{"schemaVersion":"2.0","version":"2026-10-21.workunits.2","checkedAt":"2026-10-21","nextReview":"2026-11-20","scope":"四个AI影响显著领域的具体交付任务。按接收目标组织输入、AI执行循环、逐主张证据、工具可得性与验收；同一任务供不同职业按适用主张复用，不代表职业全自动化或本站已实跑。","outlookFraming":"当前建议回答今天怎样交付；未来3–5年指2029–2031。编程的完整交付与数学的开放问题成果，支持认真准备更广泛自主工作的强情景。本站把AGI视作跨领域、能适应新任务且可靠达成复杂目标的讨论对象；这些进展是方向证据，尚不能确认某个AGI定义已满足或必在某年发生。三种情景均给出可观察信号、职责变化与现在行动，不把今天的局限或人的现有分工写成永久边界。","domains":[{"id":"computing","title":"编程与软件","summary":"按能接收的修改组织工作：功能、修复和审查可以整包委派，验收要回到实际行为。","current":"AI已能参与完整代码交付循环；明确规则与可运行环境是最实用起点。当前仓库的表现仍需自己的补丁和回归证据。","modules":["缺陷修复","约束内功能交付","变更审查"]},{"id":"math","title":"数学与建模","summary":"区分形式证明、数值模型和开放推导；三种“正确”有不同的接收方法。","current":"开放问题的证明搜索和研究产物已出现。固定形式目标可交内核检查，数值解可独立复算；题意、模型和研究意义分别核对。","modules":["固定命题形式证明","数值模型求解","推导与反例检查"]},{"id":"finance","title":"金融与财务","summary":"以来源可追溯、数字可重算的完整交付为中心，合并取数、计算和工作簿更新。","current":"可先委派明确口径的模型更新、对账与资料综合；接收检查来源、公式与异常，能力判断不等于收益承诺。","modules":["联动工作簿更新","交易对账","来源化研究备忘录"]},{"id":"analysis","title":"数据与业务分析","summary":"从业务问题到可重放分析包，区分指标、组成诊断和因果效果。","current":"数据发现、查询、检查和报告可以合为Agent循环；正确的数据语义和识别条件决定结论能走多远。","modules":["指标报表","组成变化诊断","实验与可识别性"]}],"units":[{"id":"W-C01","domainId":"computing","title":"缺陷修复：交付补丁与回归证据","goal":"让原本可复现的错误消失，同时保持已有行为。","trigger":"维护者有一个仍未解决的缺陷，并能提供仓库与故障条件。","scope":"包含定位、修改和回归循环，止于维护者接收；生产发布是另一项授权与环境决定。","inputs":[{"name":"仓库快照","detail":"提交SHA、安装与测试入口、允许改动范围。"},{"name":"失败契约","detail":"复现输入、实际与期望行为、影响对象；先核对是否已被其他补丁解决。"}],"deliverables":["可审查补丁及修改理由","同一复现步骤的修复前失败、修复后通过记录","受影响回归、未测区域和兼容性说明"],"acceptance":[{"criterion":"原错误被真实捕获","method":"在基线和候选上重放同一失败输入，观察行为或状态","failure":"只新增一个永远通过的测试，原问题仍在。"},{"criterion":"修复未破坏现有约束","method":"运行受影响调用方与边界回归；审阅测试是否被删改绕过","failure":"为通过测试移除权限检查或跳过失败用例。"},{"criterion":"交付对应可复现环境","method":"核对仓库SHA、依赖、命令与日志；环境失败单列","failure":"把依赖装不上写成缺陷已经修复。"}],"assessment":{"recommendation":"仓库和复现条件明确时，可委派Agent完成补丁与回归循环，再按行为证据接收。","conditions":["能够运行相关测试或提供隔离复现","补丁范围和关键兼容行为已明确","接收者能检查高影响变更和验收器本身"],"evidenceBasis":"SWT-Bench提供复现测试及补丁前后独立运行方法；当前修复采用建议为据此作出的编辑判断，不冒称本站真实仓库试验。","localStatus":"未对这一完整修复单元发布本站AI试验结果。","gaps":["尚缺目标仓库原失败→根因补丁→受影响回归的独立接收记录，尤其并发交错与真实依赖。"],"mode":"bounded-delivery","rationale":"失败契约将搜索空间限定为可观察行为；补丁前后运行与受影响回归能否捕获真实错误，比生成代码是否流畅更有判别力。对无法复现的投诉先接收诊断。","frontier":"公开系统已能构造仓库级复现测试和修复候选；材料不覆盖目标系统全部竞争条件及长期行为。","nextValidation":"选一个未解决的授权缺陷；冻结SHA，先让独立复现捕获基线，再接收补丁并重放用户路径和关键回归，保留失败尝试。","futureSignals":["新缺陷的独立复现、受影响回归和真实维护接收都持续成立，可考虑扩大相近输入范围。","原错误复发、测试被改弱或确认成本抵消收益，应收窄或暂停相应采用。"]},"workflow":{"before":"维护者逐步定位、手改、补测，再交由同事审阅。","now":"维护者交付故障契约；Agent自行搜索和修改，测试失败返回修复循环；接收以行为与差异为依据。","retained":["真实故障复现","关键兼容性与发布判断"],"added":["环境和上下文准备","核验测试是否覆盖根因","收集Agent反复失败模式"]},"humanResponsibility":"确认期望行为与业务影响，裁决需求冲突，决定接收和发布条件；无需逐行亲手完成补丁。","failureHandling":"不能复现时返回环境、尝试输入和缺失条件；补丁退化时保留失败证据并继续修正，不能靠缩小原需求宣布成功。","example":{"input":"缓存刷新后返回旧值；固定仓库版本和可重复的刷新/读取序列。","output":"修复缓存失效的补丁、基线失败与候选通过记录、并发读取回归结果。","failure":"只清空单元测试中的缓存，让测试变绿，却未改真实刷新路径。"},"evidenceLinks":[{"sourceId":"WU-C-SWT","supports":"提供补丁前后双运行的验收入口。","notCovered":"只生成复现测试，没有保证补丁正确或覆盖未知回归。"}],"resourceRefs":[{"module":"hierarchy","type":"task","id":"DEV-21670","version":"2026-10-02.2","relation":"partial","reason":"来源职责涵盖软件错误修正；未把整个职业任务等同本交付。"},{"module":"boundary","type":"task","id":"C03","version":"2026-10-19.boundary.1","relation":"partial","reason":"熟悉项目维护的历史证据与净交付成本可作背景。"}],"career":{"direction":"维护工程、可靠性工程","evidence":"用一份真实脱敏修复展示故障复现、影响分析、回归和本人做出的取舍。","nextStep":"对照目标岗位的仓库规模、语言及值班要求；缺生产经验时先补受限修复证据，再决定申请或选择初级维护岗位。"},"team":{"bottleneck":"缺陷积压是否来自定位慢、环境难复现，还是接收排队？","allocation":"先让Agent处理可复现且范围明确的修复；由熟悉系统者集中处理行为冲突与高影响接收。","nextStep":"选择一周真实积压，比较交付周期、返工与复核用时，再决定增工具、培养复核者或补维护人手。"},"sourceIds":["WU-C-SWT"],"reviewedAt":"2026-10-20","claims":[{"id":"C01-REPRO","statement":"缺陷测试应在基线失败、正确修复后通过；可用独立失败输入排除补丁和自写测试互相自证。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-SWT","locator":"Overview；unit test/reproduction script modes","relation":"method","supports":"提供补丁前后双运行的验收入口。","limit":"只生成复现测试，没有保证补丁正确或覆盖未知回归。"}],"progressIds":[],"configuration":null,"limitation":"只生成复现测试，没有保证补丁正确或覆盖未知回归。","evidenceStatus":"active"},{"id":"C01-DELIVER","statement":"可复现缺陷、可启动依赖和有限改动范围齐备时，可委派Agent整包搜索、修复、重测，按行为结果接收。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-SWT","locator":"Overview与gold validation","relation":"adjacent","supports":"从可运行缺陷控制迁移到修复反馈。","limit":"本站没有对新仓库运行；并发、数据损失和环境差异须按变体增加检查。"}],"progressIds":[],"configuration":null,"limitation":"本站没有对新仓库运行；并发、数据损失和环境差异须按变体增加检查。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"SWT提供容器验收器；修复Agent需接入自己的模型，语料和工具不是此仓库已修复证明。","tools":[{"name":"SWT-Bench","url":"https://github.com/logic-star-ai/swt-bench","version":"2026-10-20公开仓库HEAD 330a649a764fab2fadaea632776eeae87272f74b；实际运行另固定依赖/镜像/模型","access":"源码公开；需容器资源、目标仓库和模型账户"}]},"variants":[{"title":"并发与一致性","inputDifference":"竞争条件、异步消息或跨服务重试而非单线程错误。","assessmentChange":"分阶段接收可控制交错的复现后再修；单次通过不能证明竞争消失。","acceptanceChange":"加入屏障/故障注入、多种执行交错、幂等及事务不变量检查。"},{"title":"浏览器或设备缺陷","inputDifference":"原错误依赖渲染、权限、生命周期或具体设备。","assessmentChange":"编译和服务测试只接收代码层；须采用W-C04/W-C05的真实环境交付。","acceptanceChange":"重放对应浏览器/真机路径并核对持久状态，不能用Mock替代。"},{"title":"依赖与兼容迁移","inputDifference":"故障来自依赖升级或多个调用方版本。","assessmentChange":"允许Agent修依赖与适配，但需兼容矩阵而非只运行最新版本。","acceptanceChange":"旧/新客户端、锁文件、配置与失败回退各有结果。"}],"evidenceStatus":"active"},{"id":"W-C02","domainId":"computing","title":"约束内功能交付：从行为契约到真实系统变更","goal":"在既有系统内交付一段符合用户行为、数据、权限与资源约束的可运行功能，连同可独立验收的实现和集成证据。","trigger":"产品或开发者已明确需要增加的用户能力、不能破坏的旧行为及接收条件；代表例为给既有订单服务增加可完成的大批量CSV导出。","scope":"涵盖需求约束内的实现、真实工具/依赖操作与端到端行为接收，不规定具体代码路径。本例聚焦后端大批量导出，前端、移动或跨服务变体须按各自环境补证；不含生产发布决策。旧有限串行队列仍是exampleId=code的历史限定样例，只证明其去重/冲突/重试控制，不代表本次数据库、HTTP或并发后端已验证。","inputs":[{"name":"目标与行为契约","detail":"功能的用户操作、输入输出、禁止状态和失败恢复；固定仓库commit及允许改动范围。本例为GET /exports/orders.csv，明确租户/角色权限、筛选/排序/错误语义，以及是否要求请求开始时快照一致性。"},{"name":"复现环境","detail":"可启动服务、测试数据库版本与种子、脱敏记录集、现有接口行为及大批能力缺口的基线命令、回归入口；禁止把内存模拟数据库替代真实目标驱动而不说明。"},{"name":"资源与停止条件","detail":"目标行数N、最大响应时间T、峰值内存M由使用方按系统约束冻结；本例未设假装测得的数值。迁移、云资源或权限升级超出授权时给出具体阻塞。"}],"deliverables":["可应用补丁、依赖变更和接口/存储影响说明；如无须改表则说明判断依据","基线复现与补丁后原始记录：命令、数据快照、环境、HTTP状态、总行数、耗时与峰值内存","独立集成/回归检查及差异；涵盖跨租户隔离、全量记录、取消/断连清理与重试","可逆回退入口、已知未覆盖场景及运行预算/失败摘要；不把代码提交当部署成功"],"acceptance":[{"criterion":"新增功能在真实系统中满足契约","method":"验收者只拿冻结契约、原始夹具与补丁，在干净服务/数据库重跑；本例确认基线不具备合格大批量能力、补丁在N记录下满足T和M，保存原始测量。","failure":"界面或接口存在但无法完成真实操作；本例只跑小数据、暗加导出上限或更换硬件后宣布新增能力完成。"},{"criterion":"导出集合和权限正确","method":"独立SQL按契约生成主键与字段期望；逐项比较CSV解析后的记录，检查重复/遗漏、筛选、稳定排序及租户隔离；接受流式或分批等合法替代。","failure":"多表连接重复订单；分页边界漏行；普通用户下载其它租户记录。"},{"criterion":"边界和失败可解释","method":"验证空集、引号/逗号/换行/Unicode与契约规定的编码；模拟中途断连及数据变化，检查游标/连接释放和约定的一致性。","failure":"进程返回200但文件半截；断连后连接池耗尽；运行中变更造成不符合约定的混合快照。"},{"criterion":"旧路径与验收器可信","method":"独立运行受影响接口回归；用已知错误补丁确认验收能拒绝跨租户泄露或漏行，正确替代实现能通过。","failure":"仅作者测试全绿；验收硬编码特定函数或输出次序，拒收语义等价实现。"}],"assessment":{"recommendation":"行为契约、真实依赖和独立验收齐备时，可整包委派到“可接收的功能实现”；本例的后端导出需满足实际数据、权限和资源验收，再按团队发布权限采纳。","conditions":["目标服务及真实数据库驱动可运行","权限/记录一致性/性能阈值先冻结","验收可直接检查最终文件与服务状态"],"evidenceBasis":"DeepSWE提供相邻仓库级交付与行为verifier范式；SWE-Bench Pro审计说明测试也需校准。两者均非此CSV案例的实测。","localStatus":"此后端代表例未运行：模型试次、通过率、实际耗时/费用和人工介入量均未知（null）。旧有限队列控制不计入本例。","gaps":["具体功能的接口/数据语义与资源阈值尚需真实输入；本例未绑定目标服务。","真实数据库、断连与租户边界需重放"],"mode":"bounded-delivery","rationale":"仓库级行为任务支持把实现和验证作为整包目标；迁移到本例后端需补真实驱动、权限和资源证据。因此推荐在可执行验收下委派功能交付，不从基准数量推一次可靠性；环境缺失时交付诊断与可运行范围。","frontier":"外部已有跨仓库任务及可重放runner；公开结果证明其限定题集内可衡量交付，不覆盖长期值守或任意生产变更。","nextValidation":"在一个授权服务冻结大批量CSV功能契约，先检验参考正确实现与漏行/越权负控制，再让代理完成一次未见功能交付；保存全部重试、性能记录、人工接收时间。","futureSignals":["扩大：独立未见需求中，多种数据规模、真实驱动、故障与权限用例持续通过，维护成本可测。","收缩：回归只在模拟依赖通过，或schema/数据增长后反复漏行与资源泄漏；先缩到诊断或局部变更。"]},"workflow":{"before":"新增功能被拆成界面、函数、接口、数据与测试分别交接，最终才发现真实依赖或资源约束不满足。","now":"给代理目标仓库、复现和结果约束，让它选瓶颈实验、实现策略与修复顺序；反馈来自服务、数据库及独立验收。无需事先规定必须流式或先写哪一个文件。","retained":["业务契约取舍","目标环境和发布权限","回归接收"],"added":["把基线与最终状态放在相同环境比较","验收器正反控制","资源与连接泄漏的观测"]},"humanResponsibility":"确定接口语义冲突、可接受性能及上线窗口，提供授权环境并对正式发布负责；代码、测试与诊断可以由代理自行组织，不要求逐步批准。","failureHandling":"环境缺失时交出可运行范围、具体阻塞和接通条件；功能破坏权限/完整性则拒收整项并回到同一契约。仅性能未达标时报告实际曲线与瓶颈，不删去困难记录或静默扩大资源。","example":{"input":"代表例：给既有订单服务新增满足冻结N/T/M的大批量CSV导出能力；要求仅当前租户、按订单时间及ID稳定排序，筛选/编码语义与现有小批导出契约一致。基线不具备合格大批量能力，目标参数由使用方冻结。 性能T与内存M在实现前冻结；尚未给定目标仓库和参数，因此没有本例性能结论。快照一致或允许的弱一致语义也须在契约中明确。","output":"可运行功能补丁＋同环境基线/新增能力测量＋独立SQL/HTTP/CSV对账＋断连资源检查；这是拟定交付，尚无本例产物或运行成绩。","failure":"文件下载成功但漏掉批次边界订单，或用隐藏分页上限使小文件快速返回。旧有限队列产物只可在历史code入口查看，不能作此后端成功证据。"},"evidenceLinks":[{"sourceId":"WU-DEEPSWE-V11","supports":"提供仓库级任务、runner和行为检查的重放参照；支持从需求到可运行变更的任务组织。","notCovered":"没有运行本CSV后端，不提供其完成率；扩展到HTTP、数据库与性能是编辑迁移判断。"},{"sourceId":"WU-SWE-PRO-AUDIT","supports":"支持先审验收条件及覆盖再解释模型输赢。","notCovered":"厂商审计身份；不能照搬其比例到本站。"},{"sourceId":"WU-EVAL","supports":"支持最终状态和多条有效路径验收。","notCovered":"不是此后端模型成绩。"},{"sourceId":"WU-DEEPSWE-V1-YJS-TRIAL","supports":"Yjs冲突检测v1试次提供可读任务、模型补丁与发布者验收日志。","notCovered":"相邻库变更，不是本CSV后端、本站运行或一般成功率。"}],"resourceRefs":[{"module":"catalog","type":"task","id":"DEV-01","version":"2026-10-02.2","relation":"partial","reason":"复用有限串行队列旧场景与已保存代码；它只覆盖新单元的有限实现部分。"},{"module":"hierarchy","type":"task","id":"DEV-21670","version":"2026-10-02.2","relation":"partial","reason":"该来源任务下保存有限队列实例；实例不扩大到并发或生产支付。"}],"career":{"direction":"应用开发、业务自动化","evidence":"真实系统功能补丁、行为复现包、数据库/接口集成与性能证据；历史有限队列不能替代这些材料。","nextStep":"按目标岗位选真实使用场景；区分能写受限实现与能承担部署维护，补齐差异后再比较岗位。"},"team":{"bottleneck":"需求是否清楚，还是实现和联调消耗大量重复沟通？","allocation":"将确定需求整包交给Agent；产品/技术负责人处理取舍，独立操作者验收核心路径。","nextStep":"选一个真实系统中的限定功能，记录总执行/接收时间与失败，不用产出代码量替代交付。"},"exampleId":"code","sourceIds":["WU-DEEPSWE-V11","WU-SWE-PRO-AUDIT","WU-EVAL","WU-DEEPSWE-V1-YJS-TRIAL"],"reviewedAt":"2026-10-20","claims":[{"id":"C02-DELIVERY","statement":"外部已有按仓库状态和行为验收交付代码变更的公开任务入口；可为本任务设计独立验收，而非只验代码文本。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-DEEPSWE-V11","locator":"v1.1 corpus；/run Quickstart、What is Pier、Single task","relation":"adjacent","supports":"提供仓库级任务、runner和行为检查的重放参照。","limit":"没有运行本CSV后端，不提供其完成率。"}],"progressIds":[],"configuration":null,"limitation":"本站未进行该任务的模型运行，不能据此给出本地成功率。","evidenceStatus":"active"},{"id":"C02-VALIDATOR","statement":"评测失败可能来自题意或测试缺陷，接收器应同时拒绝错误补丁并接纳合法替代。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-SWE-PRO-AUDIT","locator":"SWE-Bench Pro audit：breaking issues分类","relation":"method","supports":"支持先审验收条件及覆盖再解释模型输赢。","limit":"厂商审计身份；不能照搬其比例到本站。"}],"progressIds":["P-C-VALIDATOR-AUDIT"],"configuration":null,"limitation":"本站未进行该任务的模型运行，不能据此给出本地成功率。","evidenceStatus":"active"},{"id":"C02-DELEGATE","statement":"在真实依赖、行为契约和独立验收齐备时，可整包委派功能实现；本例为后端大批量导出，缺少环境时先接收接口/集成诊断。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-DEEPSWE-V11","locator":"Run：任务/agent/模型/环境入口","relation":"adjacent","supports":"支持从需求到可运行变更的任务组织。","limit":"扩展到HTTP、数据库与性能是编辑迁移判断。"},{"sourceId":"WU-EVAL","locator":"Outcome-based evaluation与grader设计","relation":"method","supports":"支持最终状态和多条有效路径验收。","limit":"不是此后端模型成绩。"}],"progressIds":[],"configuration":null,"limitation":"本站未进行该任务的模型运行，不能据此给出本地成功率。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"公开runner可作试验入口；目标服务/数据库尚未提供，本例未实际运行。执行时记录实际agent/model及仓库提交，网页所列模型不是本任务实测配置。","tools":[{"name":"DeepSWE / Pier / mini-swe-agent","url":"https://deepswe.datacurve.ai/run","version":"DeepSWE v1.1；runner提交待运行时冻结","access":"源码/模型API；官方提供单任务运行，榜单使用Modal；本例必须另接目标仓库。"}]},"variants":[{"title":"不同用户界面或设备环境","inputDifference":"交付目标是前端交互或移动设备能力，而非本例后端导出。","assessmentChange":"沿用同一功能交付身份，但需实际UI/设备/状态环境；CSV例不能提供该环境完成性证据。","acceptanceChange":"新增操作—持久化—重新进入等端到端状态核验；布局或生成截图不能代替行为结果。"},{"title":"无既有复现","inputDifference":"只给用户投诉，无固定数据与超时记录。","assessmentChange":"先委派诊断和最小复现；不能直接承诺修复。","acceptanceChange":"接收可触发问题的输入、环境和测量；诊断本身不算原功能修复。"},{"title":"强一致大导出","inputDifference":"运行期间允许订单继续写入，要求请求开始时的快照。","assessmentChange":"需确认数据库隔离/游标成本后委派实现；可能必须先裁决资源与语义。","acceptanceChange":"并发写入下对照冻结快照，检查遗漏/重复和长事务影响。"},{"title":"跨服务改造","inputDifference":"导出需要跨服务授权或数据库迁移。","assessmentChange":"由单仓库交付变为分阶段集成，权限和可回退迁移独立接收。","acceptanceChange":"加入版本兼容、旧客户端、迁移回退与部分失败验收。"}],"externalExamples":[{"title":"DeepSWE v1：Yjs冲突检测公开试次","identity":"外部基准试次；发布者报告通过；未独立重放","inputUrl":"https://deepswe.datacurve.ai/data/tasks/yjs-map-conflict-detection","artifactUrl":"https://d3ujjcmjq6o8v6.cloudfront.net/trial-artifacts/yjs-map-conflict-detection__ESgfZ7p/artifacts/model.patch","verificationUrl":"https://d3ujjcmjq6o8v6.cloudfront.net/trial-artifacts/yjs-map-conflict-detection__ESgfZ7p/verifier/test-stdout.txt","configuration":"DeepSWE v1；Yjs base 7795050a74；claude-opus-4.7 high；试次 ESgfZ7p。任务、模型补丁和基线/新增测试日志分别可查。","limitation":"这是Yjs库的冲突检测变更，不是本页CSV服务的实测、已合并PR或生产部署；单个外部试次不推出一般成功率。","checkedAt":"2026-10-21","sourceIds":["WU-DEEPSWE-V1-YJS-TRIAL"]}],"evidenceStatus":"active"},{"id":"W-C03","domainId":"computing","title":"变更审查：交付可定位、可复现的问题","goal":"在合并前找出会影响用户或系统的真实问题，并说明审查覆盖。","trigger":"维护者有基线与候选差异，需要独立检查回归、数据和权限影响。","scope":"审查指定变更及相关上下文；未发现问题不等同系统安全证明，也不包含自动合并。","inputs":[{"name":"审查对象","detail":"基线/候选SHA、diff、变更目标与允许行为。"},{"name":"依赖上下文","detail":"调用方、数据约束、迁移与测试入口；优先列高影响路径。"}],"deliverables":["带文件位置、触发条件和影响的问题清单","关键问题的复现或静态推导依据","已查范围、未查路径及验收器疑点"],"acceptance":[{"criterion":"每条问题可重现或严格定位","method":"重放触发输入或沿真实调用链验证","failure":"只说“可能有性能风险”，没有路径和触发条件。"},{"criterion":"结论区分缺陷与偏好","method":"用需求和既有行为核对，允许合理替代实现","failure":"把个人命名偏好报成阻断缺陷。"},{"criterion":"审查没有靠被审补丁自证","method":"检查关键测试/规则是否被改弱，独立核验高影响点","failure":"Agent删掉失败测试后，审查只引用CI全绿。"}],"assessment":{"recommendation":"可让AI并行检查差异、定位疑点和构造复现；关键发现需核实，接收不能只看“无问题”。","conditions":["审查者可访问必要上下文","重点风险和验收准则明确","有复现路径或独立的静态依据"],"evidenceBasis":"ReviewBench提供人工参考发现、语料与离线校准方法；未知缺陷、漏报与误报需要按目标仓库另查。","localStatus":"本站未测本单元漏报与误报率。","gaps":["尚缺本仓库有效发现、误报、漏报与确认时间的盲测；参考PR发现集合不是缺陷全集。"],"mode":"staged-delivery","rationale":"审查搜索可以并行整包进行，但误报与漏报代价不同；每条发现有反证路径，才能从候选收敛为阻断。独立复现比同一Agent对自己补丁再次解释更可信。","frontier":"已有多语言真实PR参考资产和审查系统；参考标注覆盖有限，语义裁判有误差。","nextValidation":"用历史真实缺陷、合法替代与无缺陷改动组成盲测；发现后独立复现，记录漏报、误报和确认时间。","futureSignals":["在盲测真实PR中有效发现增加且确认/误报时间可控，可考虑扩大相近输入范围。","关键权限/数据缺陷漏报，或依赖裁判产生大量不可复现发现，应收窄或暂停相应采用。"]},"workflow":{"before":"同事阅读diff，手动追调用链和运行检查。","now":"AI按风险搜上下文、提出发现并构造验证；人集中裁决影响和优先级，程序检查明确约束。","retained":["独立接收","关键逻辑与权限审查"],"added":["识别审查幻觉与重复发现","检查被改动的测试能否继续捕获问题"]},"humanResponsibility":"确认问题对业务的严重性，裁决证据冲突和接受残余风险；不需要把AI的每条建议都当待办。","failureHandling":"缺上下文时明确未审区域和所需文件；不可复现的疑点保留为待证，不阻断为已证实缺陷。","example":{"input":"给订单查询增加商品明细JOIN的补丁。","output":"发现一单多商品会重复累加付款，给出两商品一付款复现与修改位置。","failure":"只检查SQL能运行，漏掉金额被JOIN放大的业务错误。"},"evidenceLinks":[{"sourceId":"WU-C-REVIEW","supports":"可检查参考发现、版本和裁判方法。","notCovered":"没有运行其模型或目标补丁；不能把未报问题当作安全证明。"}],"resourceRefs":[{"module":"catalog","type":"task","id":"DEV-04","version":"2026-10-02.2","relation":"adjacent","reason":"复用工具Agent评测与回归诊断的验收思路；工作对象由Agent协议扩展为具体代码变更。"}],"career":{"direction":"软件质量、代码审查与可靠性","evidence":"用两份真实差异展示一个有效发现与一个被排除误报，附复现和优先级理由。","nextStep":"核对目标岗位的系统深度、安全或测试工程要求；选择作品不足之处补证，而非只列会用AI审查。"},"team":{"bottleneck":"Agent生成速度超过现有审查容量，还是审查缺少上下文？","allocation":"程序先检查硬约束，AI收敛可证实疑点，系统负责人处理关键发现和跨模块影响。","nextStep":"先测每份变更的有效发现与接收耗时；优先减少重复告警，确有持续缺口再培养或招聘审查人员。"},"sourceIds":["WU-C-REVIEW"],"reviewedAt":"2026-10-20","claims":[{"id":"C03-CALIBRATE","statement":"公开ReviewBench提供真实PR及人工参考发现，可用来校准审查的有效发现和误报；参考集合本身仍不是缺陷全集。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-C-REVIEW","locator":"golden findings与offline review","relation":"direct","supports":"可检查参考发现、版本和裁判方法。","limit":"没有运行其模型或目标补丁；不能把未报问题当作安全证明。"}],"progressIds":[],"configuration":null,"limitation":"没有运行其模型或目标补丁；不能把未报问题当作安全证明。","evidenceStatus":"active"},{"id":"C03-DELIVER","statement":"可委派独立Agent追调用链、提出可验证发现，再用触发输入/不变量淘汰误报；按真实影响和覆盖接收，不按发现数量。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-REVIEW","locator":"golden findings与severity/category","relation":"partial","supports":"支持把审查结果变成可核查问题。","limit":"跨库状态、业务约束和未知漏洞须补上下文；无人发现不等于不存在。"}],"progressIds":[],"configuration":null,"limitation":"跨库状态、业务约束和未知漏洞须补上下文；无人发现不等于不存在。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"ReviewBench语料/参考和离线工具公开；需另配审查Agent、裁判及目标仓库，本站未执行。","tools":[{"name":"ReviewBench","url":"https://github.com/review-bench/ReviewBench","version":"2026-10-20公开仓库HEAD 2c34a68f73f6473a97cf35bb82baf07e9906404e；实际运行另固定依赖/镜像/模型","access":"公开语料；模型与LLM裁判需自备账户/预算"}]},"variants":[{"title":"权限与持久数据","inputDifference":"包含租户隔离、金额、迁移或删除数据。","assessmentChange":"先检查高影响不变量并构造反例；仅lint全绿不接收。","acceptanceChange":"沿调用方与存储核对权限边界、重试/回退和副作用。"},{"title":"测试与验收器修改","inputDifference":"补丁本身同时改测试或阈值。","assessmentChange":"先审测试是否仍拒绝错误，再解释CI结果。","acceptanceChange":"对被删测试、宽化断言和跳过路径重放独立负控制。"},{"title":"视觉与硬件上下文","inputDifference":"缺陷由浏览器交互或实物接口触发。","assessmentChange":"静态审查只交候选发现；运行证据转W-C04/W-C10。","acceptanceChange":"在真实目标环境验证触发与修复，不凭截图/编译确认。"}],"evidenceStatus":"active"},{"id":"W-M01","domainId":"math","title":"固定命题形式证明：找证明并交给内核核验","goal":"为一个不允许偷偷改写的形式命题交付可检查证明。","trigger":"研究者、学生或工程师已经锁定目标命题和允许依赖。","scope":"证明搜索可自动循环；自然语言题意到形式陈述另行核对。内核通过只接受这一命题与依赖下的证明。","inputs":[{"name":"固定命题","detail":"可信文件中的类型、量词、定义和假设；明确允许修改的proof区域。"},{"name":"核验环境","detail":"Lean、Mathlib与依赖revision；允许公理清单和资源预算。"}],"deliverables":["证明文件与使用的依赖","指定环境的检查日志和公理依赖记录","未解决子目标或搜索停止原因"],"acceptance":[{"criterion":"证明的是原命题","method":"比较可信声明与提交声明，核对自然语言含义","failure":"把任意实数改成非负数，得到较弱命题再宣布通过。"},{"criterion":"全部目标由允许的基础证明","method":"干净环境内核检查并审计传递公理，拒绝sorryAx或未批准公理","failure":"依赖文件藏有sorry，或添加axiom False。"},{"criterion":"可重复核验","method":"保留版本与命令；合理不同证明也应被接受","failure":"只给截图或模型说“编译成功”，没有可检查文件。"}],"assessment":{"recommendation":"可委派AI搜索证明；固定命题在指定环境通过内核与公理检查后，接收这一份证明。","conditions":["形式陈述已锁定且语义经核对","依赖可安装并精确固定","可独立检查公理与完整性"],"evidenceBasis":"AlphaProof Nexus公开研究展示在开放问题上搜索并交付Lean证明；Lean官方提供核验方法。它支持以证明文件为交付，不要求先相信模型解释。","localStatus":"本站固定实数命题的参考与替代证明已在Lean 4.19.0和固定Mathlib实际编译，额外命令、错证明、改命题等控制被拒；属于受限语法的验收器校准，新增AI试次为0。","gaps":["当前只校准固定命题的受限证明语法；新领域题型和更完整语法需独立验收与隔离运行。","对目标领域的未见命题记录搜索预算与失败，不用成功案例推普遍覆盖"],"mode":"bounded-delivery","rationale":"Nexus等外部研究表明搜索可以交付完整证明；内核允许对交付逐份接收而不先假定可靠性。本站受限命题仅校准验收器，不是找证明能力样本；新领域仍需固定声明和依赖。","frontier":"内部研究配置已在特定开放问题交付文件；公开证明服务有入口，但不能继承研究配置的预算和覆盖。","nextValidation":"对一组未泄漏、目标不变的领域引理运行公开工具，保存预算与失败；先确保正确替代、弱化命题及隐藏未证依赖分别被接收/拒绝。","futureSignals":["扩大：未见命题在固定声明下完整证明并通过依赖审计，重试/成本可追踪。","收缩：成功依赖改题、sorry或不允许公理；将交付降为未解目标与部分引理。"]},"workflow":{"before":"人寻找引理并写证明，再由同行逐段检查。","now":"人锁定问题和允许假设；AI搜索、调用Lean、根据反馈修订；机器检查形式正确，人核对题意与意义。","retained":["形式陈述与原问题一致","允许公理和研究贡献判断"],"added":["维护可搜索的形式库","检查证明依赖与版本","记录未成功搜索"]},"humanResponsibility":"明确问题及可接受假设，核对形式语义和结果意义；证明步骤可由AI产生，内核承担精确逻辑检查。","failureHandling":"预算结束但仍有目标时交出已证引理、未解目标和尝试范围；不能加假设或把未找到证明说成命题为假。","example":{"input":"固定实数x、y，目标2*x*y ≤ x^2+y^2；不增加x或y非负条件。","output":"基于(x-y)^2≥0的证明文件、核验日志和依赖列表。","failure":"改证0≤x^2+y^2，或用sorry占位，却把编译输出当完整证明。"},"evidenceLinks":[{"sourceId":"WU-NEXUS","supports":"完整搜索与形式反馈的作者报告。；已有完整产物支持限定委派。","notCovered":"不是本站实数不等式或任意新命题的概率。；不继承本站验收器校准为AI能力。"},{"sourceId":"WU-LEAN","supports":"独立形式检查方法。","notCovered":"文档不提供模型找证明成功率。"}],"resourceRefs":[],"career":{"direction":"形式化数学、验证工程、AI数学研究支持","evidence":"展示固定声明、证明文件、内核日志和自然语言对照，并解释一个被拒的弱化命题。","nextStep":"区分研究数学岗位与形式验证工程岗位，核对所需理论背景和编程能力；先补实际证明资产再选方向。"},"team":{"bottleneck":"瓶颈是找证明、形式化基础库，还是核对原意？","allocation":"把可锁定目标交给搜索Agent；让数学人员审题意，验证工程人员维护依赖与检查。","nextStep":"试点一组不同来源的引理，统计发现、形式化和检查各自成本，再决定投入工具、培养或招专门人员。"},"exampleId":"math","sourceIds":["WU-NEXUS","WU-LEAN"],"reviewedAt":"2026-10-20","claims":[{"id":"M01-SEARCH","statement":"外部研究已交付固定形式问题的证明文件，允许按产物而非生成解释接收。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-NEXUS","locator":"§2/§3 Table1、Failure Analysis","relation":"partial","supports":"完整搜索与形式反馈的作者报告。","limit":"不是本站实数不等式或任意新命题的概率。"}],"progressIds":["P-M-NEXUS"],"configuration":"Nexus v1；Gemini3.1Pro基础代理，完整配置含进化搜索/AlphaProof；题集/预算限定，成功后专家核对题意。","limitation":"不是本站实数不等式或任意新命题的概率。","evidenceStatus":"active"},{"id":"M01-CHECK","statement":"固定命题的接受须同时核对可信声明、内核和传递公理。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-LEAN","locator":"Printing Axioms / Gold Standard","relation":"method","supports":"独立形式检查方法。","limit":"文档不提供模型找证明成功率。"}],"progressIds":[],"configuration":null,"limitation":"文档不提供模型找证明成功率。","evidenceStatus":"active"},{"id":"M01-USE","statement":"当声明、环境和允许依赖已固定，可整包委派证明搜索并逐份独立接收；未完成只交实际进展。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-NEXUS","locator":"§3 Failure Analysis；结果仓库","relation":"adjacent","supports":"已有完整产物支持限定委派。","limit":"不继承本站验收器校准为AI能力。"}],"progressIds":[],"configuration":null,"limitation":"不继承本站验收器校准为AI能力。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"旧本站样例为Lean4.19.0固定Mathlib的受限校准；新的证明任务须重新固定项目，公开工具可调用不等于本题已跑。","tools":[{"name":"Lean核验","url":"https://lean-lang.org/doc/reference/latest/ValidatingProofs/","version":"历史样例Lean4.19.0；新任务按项目锁定","access":"开源工具；实际构建与依赖审计按任务运行。"},{"name":"Aristotle","url":"https://pypi.org/project/aristotlelib/2.1.0/","version":"aristotlelib2.1.0","access":"需账号/API key，本站未调用。"}]},"variants":[{"title":"固定研究引理","inputDifference":"换为领域特定形式声明，proof区以外不变。","assessmentChange":"可搜索；成本和适用库需重新核实。","acceptanceChange":"完整项目重建、可信声明对照及公理检查。"},{"title":"尚未形式化的题意","inputDifference":"只有自然语言目标。","assessmentChange":"先用W-M05交付语义对照，不把类型可编译视为已证明原题。","acceptanceChange":"先接收形式声明，再验对应证明。"}],"evidenceStatus":"active"},{"id":"W-M02","domainId":"math","title":"数值模型求解：交付解、残差和条件变化","goal":"得到满足已定义约束的数值结果，并知道条件变化后是否仍成立。","trigger":"研究或业务提出方程、目标、参数和可行性限制，需要可复算解。","scope":"模型表达、求解与独立核对组成一包；模型是否符合现实是单独的接收条件，数值收敛不是现实正确证明。","inputs":[{"name":"变量与约束","detail":"单位、取值域、整数性、目标函数、边界条件和数据版本。"},{"name":"精度与对照","detail":"容差、预算、可解析或小规模参考情形、需要检查的参数变化。"},{"name":"决策落地与情景","detail":"原始业务限制/模型方程、硬软约束解释、实施单位和变动规则；无可行方案时允许返回不可行诊断；实际效果不能只由目标值代替。"}],"deliverables":["可重跑的求解程序和参数表","解、目标值、约束残差及求解器退出状态","敏感性或边界变化报告与不可行情况","原需求—变量/约束—实现—独立检查的映射，基线与变体方案、执行适配和未建模限制"],"acceptance":[{"criterion":"结果满足原约束","method":"独立代入每条约束，检查单位、边界和整数性","failure":"将不可拆分工作包当连续变量，交付6.67个包。"},{"criterion":"声称的最优/精度有依据","method":"在小问题枚举对照；大问题看适用证书、gap或误差界","failure":"仅看到求解器success就声称任意非凸模型全局最优。"},{"criterion":"改变条件会重新计算","method":"对约束变体重跑并说明不可行或解变化","failure":"设备上限减少后照抄原解，不再检查容量。"},{"criterion":"原需求也被模型覆盖","method":"独立接收者从业务输入重新列硬约束，用不同检查代码检查方案；模拟容量减少/需求增大，基线与新方案比较可行性和代价。","failure":"优化器成功但漏掉班次休息、不可拆分任务或现实交接条件。"}],"assessment":{"recommendation":"可整包委派模型到候选解与情景报告；用独立原约束检查、界/残差和实施条件验收。复杂真实决策可先分阶段接收模型。","conditions":["模型类别适合所选求解器","整数性、单位和容差明确","有参考情形或独立可行性检查"],"evidenceBasis":"ORAgentBench提供从业务材料到可执行决策的相邻AI任务证据，SciPy文档仅支持数值方法；可行和质量/现实效果分开。","localStatus":"既有M01有人工构造的有限枚举参考；它不是AI模型试验，也未证明现实排班适配。","gaps":["用真实参数校准模型与容差","为非线性、病态或大规模问题补适当误差/最优性证据"],"mode":"bounded-delivery","rationale":"外部可执行OR任务说明模型能组织求解，但“有解”不保证目标质量。成熟求解器可接管确定性运算，采用范围取决于约束忠实度、证书与部署条件；小例枚举不代表大规模优化或真实收益。","frontier":"ORAgentBench已评价多文件/多步决策，仍有规则遗漏和可行-质量差距；它给出可试的交付范式，不给本单位职业百分比。","nextValidation":"保留历史有限整数例为检查校准；另用授权排程材料盲测，独立核对约束、目标、时间预算和改变容量后的再规划，记录被遗漏的现实限制。","futureSignals":["扩大：新行业/新规模实例，原需求映射、可行性及目标质量持续达标，实施反馈与模型一致。","收缩：建模反复漏规则，gap未闭或数据延迟导致无法执行；先交模型诊断和保守基线。"]},"workflow":{"before":"人手工转公式、录参数、运行工具、再逐条核对。","now":"给出完整约束后，AI可合并建模代码与验证报告；稳定刷新使用固定程序，异常返回模型层。","retained":["目标与现实约束确认","可行性和精度验收"],"added":["检查AI遗漏的取值域","用参数扰动发现脆弱结论"]},"humanResponsibility":"决定业务目标、约束与可接受误差；判断模型结果如何落到现实，而非把每次算术保留为人工劳动。","failureHandling":"不收敛、不可行或无界时返回状态、残差和相关约束；不能挑一个看似合理的数代替求解。","example":{"input":"非负整数x/y；2x+5y≤40、x+y≤12、y≤8；最大化3x+4y。","output":"小范围枚举参考为x=7,y=5，目标41；把y上限改4后重新得到x=8,y=4、目标40。","failure":"交出连续解20/3、16/3和目标约41.33，忽略整数约束。"},"evidenceLinks":[{"sourceId":"WU-M-ORAGENT","supports":"给出真实约束式决策的作者评估，不仅是公式或代码生成。；漏规则与质量失败决定独立验收重点。","notCovered":"107任务不是现场收益；不能将其分数分配给本整数示例。；未重跑，本例历史人工枚举不是AI任务试次。"},{"sourceId":"WU-SCIPY","supports":"限定模型类别的求解与检查入口。","notCovered":"不验证AI建模，也不证明现实约束齐全。"}],"resourceRefs":[{"module":"boundary","type":"task","id":"M01","version":"2026-10-19.boundary.1","relation":"partial","reason":"数值推导案例与有限资源约束示例部分对应；不混同形式证明。"}],"career":{"direction":"优化建模、运筹分析、科学计算","evidence":"展示模型假设、可复算程序、约束核对和一个参数变化后失效的旧方案。","nextStep":"按目标岗位核对理论、规模、数据接口要求；能做小例时先补更接近现场的规模与解释证据。"},"team":{"bottleneck":"计算慢，还是变量、约束和数据来源不一致？","allocation":"AI与求解器承担候选计算；业务确认约束，数值负责人检查模型与解。","nextStep":"将模型错误、数据错误和求解失败分别统计；据实际瓶颈决定补业务建模、数值或工程人员。"},"sourceIds":["WU-M-ORAGENT","WU-SCIPY"],"reviewedAt":"2026-10-20","claims":[{"id":"M02-OR-EVIDENCE","statement":"ORAgentBench的GPT-5.4/Codex配置在107任务中报告35.51%通过；部分可行方案仍达不到质量阈值，可行和高质量交付不同。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-M-ORAGENT","locator":"§3.4、§4.1、Table2","relation":"partial","supports":"给出真实约束式决策的作者评估，不仅是公式或代码生成。","limit":"107任务不是现场收益；不能将其分数分配给本整数示例。"}],"progressIds":["P-M-ORAGENT"],"configuration":"v1；107任务/14完整model-agent行；OpenAI用Codex其余Claude Code，high；45分钟交互、5分钟求解，离线Harbor/SCIP，默认四个OR skills。 此结果为GPT-5.4/Codex(high)的单次完整行；通过需可行且归一化质量q>0.4，非任意排程成功率。","limitation":"107任务不是现场收益；不能将其分数分配给本整数示例。","evidenceStatus":"active"},{"id":"M02-SOLVER","statement":"求解器退出状态、原约束残差与最优性证据应分别解释。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-SCIPY","locator":"linprog：bounds/integrality/result/status","relation":"method","supports":"限定模型类别的求解与检查入口。","limit":"不验证AI建模，也不证明现实约束齐全。"}],"progressIds":[],"configuration":null,"limitation":"不验证AI建模，也不证明现实约束齐全。","evidenceStatus":"active"},{"id":"M02-ADOPT","statement":"明确模型类别、约束和独立检查时，可把候选计算与变体重算整包委派；实际执行另核。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-M-ORAGENT","locator":"Failure modes；AppendixF multi-step cases","relation":"adjacent","supports":"漏规则与质量失败决定独立验收重点。","limit":"未重跑，本例历史人工枚举不是AI任务试次。"}],"progressIds":[],"configuration":null,"limitation":"未重跑，本例历史人工枚举不是AI任务试次。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"数值/OR计算与评估有公开实现；本地历史小例为人工枚举参考，现实业务实例、模型与预算尚需提供。","tools":[{"name":"SciPy线性规划","url":"https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.linprog.html","version":"文档2026-10-20读取；实际SciPy/HiGHS版本需锁定","access":"开源；必须按LP/MIP等模型选择合适入口。"},{"name":"ORAgentBench","url":"https://github.com/ORAgentBench/ORAgentBench/tree/c9eb952435a4352f33daa2a35efe0f8c76d31b28","version":"c9eb9524…；论文v1","access":"公开任务与Harbor环境；模型API/solver依赖另配，未运行。"}]},"variants":[{"title":"业务约束不完整","inputDifference":"只有“最低成本”，缺休息/交接或例外规则。","assessmentChange":"先交需求-模型差异和几种约束情景，分阶段接收。","acceptanceChange":"从原始排班/物流规则独立检查，而不是仅运行生成的约束。"},{"title":"不确定/动态再规划","inputDifference":"需求、交通或容量随时间变。","assessmentChange":"可交情景/鲁棒方案和重算入口；预测质量需另证。","acceptanceChange":"不同情景独立可行性、服务水平及再规划冻结期/实施成本。"},{"title":"数值方程与仿真","inputDifference":"非优化方程/PDE或病态系统。","assessmentChange":"原数值身份仍保留；使用W-M06的离散误差、收敛和物理检查，不套最优性gap。","acceptanceChange":"解析/高精度基准、网格/步长误差和稳定性分别检查。"},{"title":"已明确约束的静态决策","inputDifference":"目标、数据和硬软约束已冻结，不要求缺规则诊断或在线更新。","assessmentChange":"可整包委派建模代码、候选解和质量证据，稳定计算直接用solver。","acceptanceChange":"独立从原始规则核可行/整数性和目标质量，检查界/gap/容差；不附加动态或缺约束前提。"}],"evidenceStatus":"active"},{"id":"W-M03","domainId":"math","title":"推导与反例检查：找出可定位的论证缺口","goal":"帮助判断一段推导在给定定义与范围内哪里成立、哪里需要修正。","trigger":"作者已有猜想、证明草稿或数值推导，希望独立找错或找反例。","scope":"交付反例、具体缺口或限定检查结论；没找到反例不能当作完整证明。","inputs":[{"name":"论证材料","detail":"完整定义、量词、假设、引用和希望检查的结论。"},{"name":"检查范围","detail":"允许的数值/符号工具、领域范围、可用文献和搜索预算。"},{"name":"引用与版本","detail":"每个被复用引理的原文、条件及修订状态；论文/证明文件/解释的覆盖对照。"}],"deliverables":["逐步假设核对与可定位的问题","可复算反例或缺口说明","已检范围、无法判断项及下一步证明义务","引用命题—本步使用—前提匹配表；可读修订稿区分修复完成与仍未解"],"acceptance":[{"criterion":"反例真的满足原前提","method":"代入原定义独立复算，并解释违反哪条结论","failure":"反例不在定义域，或改变了原条件。"},{"criterion":"每步没有偷换条件","method":"对照量词、极限/交换条件和引用前提","failure":"把数值抽样通过解释成对全部实数成立。"},{"criterion":"报告强度匹配检查覆盖","method":"区分找到缺陷、有限检查未发现、正式证明","failure":"没找到错误便写“定理已证明”。"}],"assessment":{"recommendation":"适合让AI并行尝试证明、反驳和检索依据；可验证反例可直接推进修订，开放正确性结论仍需更强证明。","conditions":["原始论证与定义完整","反例可复算，引用可访问","接收者能核对领域假设"],"evidenceBasis":"Aletheia研究案例展示生成、检查、修订和找反例的作用；公开数学产物允许进一步独立检查。","localStatus":"本站未运行这一单元的AI论证审查试验。","gaps":["用已知正确和带隐蔽缺口的论证校准误报与漏报","重要结论可转为W-M01形式化或专项专家审阅"],"mode":"candidate-exploration","rationale":"外部研究和可修订公开产物说明候选检查有用；可复算反例可独立接收，未发现缺陷只描述搜索覆盖。审查任务不同于完成研究，也不能用同模型赞同自己当独立正确性。","frontier":"研究系统能交可读数学与部分形式产物，仍有撤回/语义争议；应把可定位的缺口与完整证明分开。","nextValidation":"在未泄漏的正确论证与含隐蔽量词/引用缺口论证上校准，保存有效反例、误报和漏报；需要全称保证时转入相应证明任务。","futureSignals":["扩大：新领域论证中能持续给出原前提内的有效缺口，引用核对和独立接收一致。","收缩：以浮点异常造假反例、引用不存在，或有限检查被包装成完整证明。"]},"workflow":{"before":"作者手工推导，同行按顺序读完并寻找疑点。","now":"AI可同时找证明路径与反例，调用符号或数值工具；人按可复算发现修订，关键命题继续形式化。","retained":["定义与假设一致","研究新颖性和意义判断"],"added":["区分探索失败和反证","核对AI引用确实支持所用步骤"]},"humanResponsibility":"确定问题与结论需要达到的证明强度，判断修订是否改变研究目标；不把同一模型自审当作独立保证。","failureHandling":"发现缺口返回位置和所需附加条件；未解决则保留搜索范围，不补造引理或文献。","example":{"input":"声称对所有实数x都有sqrt(x^2)=x。","output":"指出x=-1满足定义域而左侧1右侧-1；应改为abs(x)，或明确增加x≥0。","failure":"仅测试x=1、2、3就宣布全称命题成立。"},"evidenceLinks":[{"sourceId":"WU-ALETHEIA","supports":"原作者研究工作流报告。；把搜索和正确性证明分离。","notCovered":"已投稿不同于已接收；自然语言checker不同内核。；此为任务接收设计，本站未做AI审稿试验。"},{"sourceId":"WU-MATH-RELEASE","supports":"支持追踪实际论证与更正影响。","notCovered":"公开不等于全部有效，也不将新颖性等同正确性。"}],"resourceRefs":[],"career":{"direction":"研究助理、数学研究、算法论证审查","evidence":"给出一段原论证、一个有效反例、修订后的条件和仍未解决部分。","nextStep":"对照目标研究方向的理论基础与发表要求；把AI协作过程与本人识别的关键条件分开陈述。"},"team":{"bottleneck":"推导尝试太慢，还是专家时间被无效候选占用？","allocation":"AI扩大候选和反例搜索；工具先淘汰可复算错误，专家集中审阅剩余论证。","nextStep":"按每份有效发现耗费的专家时间比较流程，决定增加搜索预算、形式化基础或专业复核。"},"sourceIds":["WU-ALETHEIA","WU-MATH-RELEASE"],"reviewedAt":"2026-10-20","claims":[{"id":"M03-REVIEW","statement":"研究系统已展示把候选生成、检查和修订结合；这提供审查线索，不提供所有自然论证的正确性证书。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-ALETHEIA","locator":"研究案例与方法/限制；Feng26状态","relation":"partial","supports":"原作者研究工作流报告。","limit":"已投稿不同于已接收；自然语言checker不同内核。"}],"progressIds":["P-M-ALETHEIA"],"configuration":"Aletheia/DeepThink研究系统，专家协作及自然语言检查；任务、干预和预算随个案，不据此估本站概率。","limitation":"已投稿不同于已接收；自然语言checker不同内核。","evidenceStatus":"active"},{"id":"M03-SOURCE-CHECK","statement":"公开产物和修订入口让论证、引用条件与更正可逐项检查。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-MATH-RELEASE","locator":"公开结果及correction/revision入口","relation":"method","supports":"支持追踪实际论证与更正影响。","limit":"公开不等于全部有效，也不将新颖性等同正确性。"}],"progressIds":["P-M-RELEASE"],"configuration":null,"limitation":"公开不等于全部有效，也不将新颖性等同正确性。","evidenceStatus":"active"},{"id":"M03-ADOPT","statement":"在原前提内可复算的反例或确切缺口可以接收；未找到缺陷必须保留未判定范围。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-ALETHEIA","locator":"研究协作与核查范围","relation":"adjacent","supports":"把搜索和正确性证明分离。","limit":"此为任务接收设计，本站未做AI审稿试验。"}],"progressIds":[],"configuration":null,"limitation":"此为任务接收设计，本站未做AI审稿试验。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"Aletheia是研究配置；使用公开通用Agent及精确计算/检索可组织候选审查，但不继承内部表现。","tools":[{"name":"Aletheia研究说明","url":"https://deepmind.google/blog/accelerating-mathematical-and-scientific-discovery-with-gemini-deep-think/","version":"2026-02-11说明","access":"研究方法/产物，不是完整Aletheia自部署包。"},{"name":"Lean核验（形式交付适用）","url":"https://lean-lang.org/doc/reference/latest/ValidatingProofs/","version":"按目标项目锁定版本","access":"只用于适用的形式子结论，不强制自然论证走Lean。"}]},"variants":[{"title":"跨文献引理迁移","inputDifference":"核心不是算错而是原引理的紧性/可测性/独立性前提。","assessmentChange":"委派引用核对和条件映射，未满足条件则拒绝该步。","acceptanceChange":"回到原文核查前提与反例，不能仅比对标题。"},{"title":"数值发现","inputDifference":"反例来自浮点/有限网格。","assessmentChange":"先转精确证书或限定探索范围。","acceptanceChange":"精确重算及误差界；未见反例不算证明。"}],"evidenceStatus":"active"},{"id":"W-F01","domainId":"finance","title":"财报到联动工作簿：取数、对齐口径并更新模型","goal":"把锁定的新财报资料更新到既有联动模型，使关键数字可追溯、公式可重算、说明与图表随模型一致变化。","trigger":"新财报或修订资料发布，需要更新已有模型、图表及变化解释。","scope":"同一任务含取数、期间/单位/重述对齐、跨sheet依赖修复及变化说明；本样稿不包含会计政策裁决、投资建议或交易执行。输入已有公式模型，不把导出一张CSV当工作簿更新。","inputs":[{"name":"锁定报告及映射","detail":"报告版本/页面、期间、币种、单位、已批准指标定义及重述处理；每个模型字段对应的报告表和备注。"},{"name":"工作簿与改动边界","detail":"旧版xlsx及hash，Raw/Mapping/Model/Scenario/Chart等实际sheet、命名区域与公式依赖；允许修改区、历史锁定区、外链/宏/受保护表清单。"},{"name":"重算与接收环境","detail":"使用方实际Excel/WPS/其他兼容引擎及版本；材料访问授权、口径争议接收者和重要性阈值；不支持的外链或函数必须先披露。"}],"deliverables":["更新后的原生工作簿及版本，保留必要公式、命名区域、图表和审计信息","来源—页/表—指标—期间—单位转换—单元格的追溯表；重述和冲突单列","真实计算引擎的重算记录、控制总数、跨表依赖与非目标区域差异","一页变化说明，分别标历史事实、推导和假设；说明中数字可回溯到模型，附未解决项"],"acceptance":[{"criterion":"取数与口径真实相符","method":"独立从锁定报告抽查所有关键指标及重述；核对单季/累计、币种与单位。代表例1200千元必须映射1200000元。","failure":"季度累计当单季；同比基期仍用被重述值；单位转换仅在文字修正，模型未改。"},{"criterion":"最终工作簿可计算且联动","method":"在干净环境打开最终工作簿并完整重算，检查公式/命名引用/外链及合计。对收入驱动做受控扰动，验证应有公式依赖的Model、Scenario和Chart联动；承诺动态链接的说明另验联动。恢复原输入并再次重算后，逐项核对静态一页说明与最终模型一致。","failure":"只读取缓存；公式变静态数；图表仍指向旧区域；改收入后毛利率/情景不同步。"},{"criterion":"另一条路径得到相同关键数","method":"验收者从源数用独立代码或手算重算关键合计、比率与差异桥接，定义数值容差；不要求公式文本或布局逐字相同。","failure":"自动judge认为文稿好看，但关键比率算错；因合法等价公式而误拒。"},{"criterion":"范围和解释完整","method":"比较前后输入、公式、隐藏sheet、格式和图表对象，逐项解释授权改动；从最终说明追到模型再追到报告。","failure":"为修一处错误覆盖历史假设或隐藏依赖；说明引用旧缓存数字。"}],"assessment":{"recommendation":"可把来源数表到联动工作簿的候选交付整包委派；关键口径、引擎重算和非目标改动独立通过后接收。","conditions":["来源/映射/修改区锁定","真实表格引擎支持文件特性","重大数字有独立计算路径"],"evidenceBasis":"Finch v5是跨资料工作流历史评估；SpreadsheetBench 2显示局部修改分与整表正确分差距。两者支持完整产物验收，不能作为本例成绩。","localStatus":"本例AI运行数、成功率、费用与实际人工复核时间均为null；未打开/重算代表例。","gaps":["目标工作簿及外链/宏兼容尚未核实","代表例端到端与扰动验收待做"],"mode":"bounded-delivery","rationale":"公开任务族展示了跨文件财务产物可由代理执行，同时局部修改正确不足以保证整表正确。因此可委派形成完整候选并迭代，但只有独立来源核对和实际重算支持接收；未报告现代模型对本工作簿可靠性。","frontier":"论文当时模型已处理复合金融workflow与多sheet任务；现有可下载评价资料不是可直接部署的通用财务Agent。","nextValidation":"取一份授权报告/模型，冻结原始版和关键数控制；先验证公式静态化、单位错误与断链负控制，再运行完整更新并在另一引擎或独立计算路径核对。","futureSignals":["扩大：未见报告、重述及复杂依赖下，完整簿和扰动测试反复通过且复核人时下降。","收缩：关键数正确但公式/图表频繁断链，或真实引擎无法重算；缩到来源数表/诊断并保留未完成范围。"]},"workflow":{"before":"抄数、模型维护、图表刷新和说明撰写分开交接，错误常在最终联动时暴露。","now":"代理围绕同一交付图组织读取、计算与更新；对已有复杂模型先利用真实依赖，按重算反馈修复，说明从最终模型提取。稳定确定性计算可直接运行，不要求多轮模型对话。","retained":["口径与模型假设裁决","正式采纳版本"],"added":["源数据到单元格追溯","跨sheet扰动检查","重算引擎与文件特性的兼容记录"]},"humanResponsibility":"选择会计与业务口径、决定重要性和模型假设，接收对外使用版本；取数和算术可以自动完成并校验。","failureHandling":"冲突值隔离并列出各自来源；无法重算的公式明确标出依赖，不静默择数或用缓存冒充新计算。","example":{"input":"设计例：新报告收入1200千元，旧簿单位元；Model以成本/收入计算成本率，Scenario引用收入，Chart引用Model；提供期间和原成本值，不凭空补数。","output":"Raw保存原值与定位、Mapping转换单位，Model收入变1200000元并保留公式；Scenario/Chart真实重算、说明回指模型；尚无该例实际文件。","failure":"总收入正确但毛利率公式被覆盖成静态值，图表和情景继续使用旧值；文件可打开仍拒收。"},"evidenceLinks":[{"sourceId":"WU-FINCH-V5","supports":"提供完整财务产物的人评与自动judge不一致证据；支持跨文件完整性及合理替代验收。","notCovered":"历史模型、分母为workflow；不是本例或今天模型成功率；并未测本工作簿。"},{"sourceId":"WU-SPREADSHEETBENCH2","supports":"提供同类工作簿中局部修改与整表分离的证据；支持实际重算与整表检查。","notCovered":"不是当前模型能力；整表与图表评价不同，不以overall冒充统一成功率；扰动链路为本任务编辑设计，需要本地验证。"}],"resourceRefs":[{"module":"boundary","type":"task","id":"F01","version":"2026-10-19.boundary.1","relation":"partial","reason":"沿用财报取数、公式计算和期间口径的局部验收，交付层合并为工作簿包；历史事件保持原范围。"}],"career":{"direction":"财务分析、模型维护、研究运营","evidence":"提供脱敏来源表、单元格追溯、公式重算和重述处理记录。","nextStep":"按目标岗位核查会计基础、行业和建模深度；不要用能抽数替代估值判断能力。"},"team":{"bottleneck":"重复取数和模型刷新慢，还是口径争议无法接收？","allocation":"AI处理跨文件更新；分析师裁决口径与重要假设，复核者抽查关键链路并重算。","nextStep":"用一轮报告更新测错数、公式损坏和复核耗时，比较工具投入、内部培养与补建模岗位。"},"sourceIds":["WU-FINCH-V5","WU-SPREADSHEETBENCH2"],"reviewedAt":"2026-10-20","claims":[{"id":"F01-WORKFLOW","statement":"Finch v5在172个复合workflow中分别报告GPT-5.1 Pro产品代理人评66/172与自动评价72/172；完整请求与单一数值正确有区别。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FINCH-V5","locator":"Main results；Human evaluation；product agent configuration","relation":"partial","supports":"提供完整财务产物的人评与自动judge不一致证据。","limit":"历史模型、分母为workflow；不是本例或今天模型成功率。"}],"progressIds":["P-F-FINCH-V5"],"configuration":"arXiv v5 2026-04-15；172 workflows/384 tasks；GPT-5.1 Pro产品代理；外部浏览开、历史聊天关；具体每例人时/重试与费用未用于本站估计。","limitation":"本站未进行该任务的模型运行，不能据此给出本地成功率。","evidenceStatus":"active"},{"id":"F01-WHOLEBOOK","statement":"SpreadsheetBench 2金融建模100项中，Opus4.6的每任务目标单元格匹配比例平均89.69%，整表通过34/100；两者衡量不同。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-SPREADSHEETBENCH2","locator":"Evaluation / Main results：financial modeling modification vs accuracy","relation":"partial","supports":"提供同类工作簿中局部修改与整表分离的证据。","limit":"不是当前模型能力；整表与图表评价不同，不以overall冒充统一成功率。"}],"progressIds":["P-F-SPREADSHEET2"],"configuration":"arXiv v1；总集321任务，金融建模子类100任务；统一SWE-agent scaffold，最多50轮。89.69%为100项中每任务目标单元格计算值匹配比例的平均，34/100为整张输出工作簿所有单元格均匹配参考的任务比例；图表rubric另计。","limitation":"本站未进行该任务的模型运行，不能据此给出本地成功率。","evidenceStatus":"active"},{"id":"F01-ADOPT","statement":"本例应委派到可重算候选交付，并以源数、整表联动和范围差异独立接收。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FINCH-V5","locator":"Evaluation：完整工作流评价","relation":"method","supports":"支持跨文件完整性及合理替代验收。","limit":"并未测本工作簿。"},{"sourceId":"WU-SPREADSHEETBENCH2","locator":"Evaluation：recalculation与workbook-level correctness","relation":"method","supports":"支持实际重算与整表检查。","limit":"扰动链路为本任务编辑设计，需要本地验证。"}],"progressIds":[],"configuration":null,"limitation":"本站未进行该任务的模型运行，不能据此给出本地成功率。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开数据和评估代码可获得；需另选可调用模型与表格运行环境。发布论文、可安装评价管道、目标簿能重算是不同状态。","tools":[{"name":"Finch官方评估管道","url":"https://github.com/FinWorkBench/Finch","version":"论文v5；运行前固定代码commit，旧论文代码在previous分支","access":"开源评价入口；Excel预处理依赖xlwings和本机Excel，未安装运行。"},{"name":"SpreadsheetBench 2","url":"https://github.com/RUCKBReasoning/SpreadsheetBench-2","version":"论文v1；任务/代码revision运行时锁定","access":"输入/参考/输出需重算；图表评分需Windows Excel/WPS COM；不是现成财务产品。"}]},"variants":[{"title":"包含重述与并购口径","inputDifference":"新旧报告历史数字或合并范围变化。","assessmentChange":"先独立接收口径映射，代理可并行计算两套可比口径；不得自行挑选好看结果。","acceptanceChange":"重述桥接与原始来源逐项对应，说明不可比区间。"},{"title":"宏/外链驱动的估值簿","inputDifference":"原文件依赖VBA、外部数据源或插件。","assessmentChange":"环境不可满足时只委派诊断和受限更新；不能用普通xlsx库保证整簿交付。","acceptanceChange":"在实际运行环境执行宏/链接并核对关键输出；外链失效明确拒收完整任务。"},{"title":"仅来源数表","inputDifference":"不要求模型、情景或图表更新。","assessmentChange":"可缩到直接取数与单位转换，确定性程序足够。","acceptanceChange":"只验来源、期间、单位与重复；此通过不能升级为联动工作簿通过。"}],"evidenceStatus":"active"},{"id":"W-F02","domainId":"finance","title":"交易对账：算清净额并解释异常","goal":"按既定截止日与状态规则，将交易和账面逐项对齐。","trigger":"财务获得交易及总账快照，需要查明差异，而不是为了平账改原数据。","scope":"计算、匹配、异常定位与建议组成一包；过账、冲销和修改口径另由授权责任人决定。","inputs":[{"name":"冻结数据","detail":"订单、付款、退款和账面快照；稳定实体主键、币种与单位。"},{"name":"业务规则","detail":"截止时间及时区、completed/captured等状态、跨期和取消处理。"}],"deliverables":["可重放SQL或程序","逐单对账表、汇总和来源行定位","异常清单及待裁决事项"],"acceptance":[{"criterion":"防止一对多连接放大金额","method":"付款和退款分别按主键汇总后再连接，独立核对行数与总额","failure":"多商品行直接JOIN付款，重复累加同笔付款。"},{"criterion":"状态与截止日符合规则","method":"边界用例检查pending、cancelled、跨期退款","failure":"扣除尚未完成或截止日之后的退款。"},{"criterion":"差异与汇总守恒","method":"独立整数分/Decimal计算，核对逐项和合计；保留原输入","failure":"为消除差额改账面或把缺记录补成0。"}],"assessment":{"recommendation":"规则明确时，可委派AI写计算、定位差异和异常说明；结果用独立重算接收。","conditions":["币种、主键和截止口径已固定","快照可复核且不直接覆盖账本","重大异常有处理负责人"],"evidenceBasis":"官方对账模板、Beancount校验和Intuit会计产品分别提供工作组织、确定性检查与厂商采用身份；当前净额样例仍是本站参考答案。","localStatus":"代表例的校准/执行记录单列；没有据此宣称真实财务流程完成率。","gaps":["确认真实数据的多币种、手续费和迟到记录规则","用实际产物与失败记录衡量新模型，而非只检查参考算术"],"mode":"staged-delivery","rationale":"在口径、主键和快照固定时可尝试完整候选对账，独立重算及负例接收。","frontier":"已有产品将分类和对账串联；本任务同币种净额与异常可程序复算。真实多币种、迟到流水和账务调整自动执行尚无本输入记录。","nextValidation":"在冻结快照植入重复连接、跨期和跨币种边界，独立重算金额与异常。","futureSignals":["在带重复连接、跨期退款和手续费的新快照保持金额守恒且异常不漏，才扩大自动匹配范围","汇率价值日或状态规则变化造成漏项时暂停对应自动平账，重做负例"]},"workflow":{"before":"人工导表、写查询、查差额、回头定位流水。","now":"AI按合同生成查询与异常解释，固定程序重算与守恒检查；人处理口径冲突和账务动作。","retained":["截止日与状态口径","账务调整决定"],"added":["主键/连接基数检查","异常证据可追溯","未决口径单独列示"]},"humanResponsibility":"批准业务口径，判断异常应如何处理并授权账务动作；人的责任不意味着AI不能承担计算。","failureHandling":"缺失流水、重复主键或跨币种规则不明时保留异常，交出可核查差额与所缺资料；不得默默补零或强制平账。","example":{"input":"A001付款10000分、当期已完成退款2000分，账面10000；A002付款合计20000，pending5000和跨期3000不扣。","output":"A001净8000、差额-2000；A002净20000、差额0；合计净28000、差额-2000。","failure":"把pending或跨期退款扣进本期，或多商品JOIN后把付款翻倍。"},"evidenceLinks":[{"sourceId":"WU-FINANCE-AGENTS","supports":"总账对账被组织为可委派的参考工作流。","notCovered":"产品发布不证明本输入已由AI成功处理。"},{"sourceId":"WU-EVAL","supports":"用最终产物、关键负例与独立参考校准验收。","notCovered":"不能以评分器存在替代真实AI运行。"},{"sourceId":"WU-FIN-BEAN-VALIDATION","supports":"可检查确定性账务校验实现。","notCovered":"账务规则工具，不证明原始交易真实、会计政策正确或AI入账成功率。"},{"sourceId":"WU-FIN-INTUIT-AGENT","supports":"官方会计产品声明支持对账及异常处理，但本页仍以冻结快照独立重算接收。","notCovered":"产品自报非独立结果；没有本站账套、错误分母或人工介入完整日志，不使用客户节时百分比。"}],"resourceRefs":[{"module":"catalog","type":"task","id":"DATA-01","version":"2026-10-02.2","relation":"partial","reason":"复用同币种逐单净额及跨期退款旧练习；新交付不扩大成全部财务工作。"}],"career":{"direction":"财务运营、对账分析、财务数据工程","evidence":"提供查询、差额复算、源流水定位及一项异常的处理取舍。","nextStep":"按岗位核对会计口径、系统和周期要求，先补真实脱敏异常材料再选择申请或培训。"},"team":{"bottleneck":"对账时间花在计算匹配，还是寻找缺失资料与裁决？","allocation":"AI与固定程序处理计算和分流；财务负责人按异常类型裁决，工程人员维护数据接口。","nextStep":"按异常类别测复核与恢复成本，优先补最常见资料缺口，再决定自动化扩围或人员安排。"},"exampleId":"finance","sourceIds":["WU-FINANCE-AGENTS","WU-EVAL","WU-FIN-BEAN-VALIDATION","WU-FIN-INTUIT-AGENT"],"reviewedAt":"2026-10-20","claims":[{"id":"W-F02-method","statement":"固定公开实现提供账务校验方法。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-BEAN-VALIDATION","locator":"validate_active_accounts、validate_currency_constraints、validate_check_transaction_balances","relation":"method","supports":"可检查确定性账务校验实现。","limit":"账务规则工具，不证明原始交易真实、会计政策正确或AI入账成功率。"}],"progressIds":[],"configuration":null,"limitation":"账务规则工具，不证明原始交易真实、会计政策正确或AI入账成功率。","evidenceStatus":"active"},{"id":"W-F02-adoption","statement":"在口径、主键和快照固定时可尝试完整候选对账，独立重算及负例接收。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-BEAN-VALIDATION","locator":"validate_active_accounts、validate_currency_constraints、validate_check_transaction_balances","relation":"adjacent","supports":"局部工具/产物配合本任务明确可检查结果，构成试用推理依据。","limit":"无同输入整包AI运行记录，不推一般完成率。"},{"sourceId":"WU-FIN-INTUIT-AGENT","locator":"Accounting Agent、Finance Agent与产品可用性段落","relation":"partial","supports":"厂商介绍分类、对账及财务洞察的产品交付。","limit":"产品自报非独立结果；没有本站账套、错误分母或人工介入完整日志，不使用客户节时百分比。"}],"progressIds":[],"configuration":null,"limitation":"本站旧参考答案与校准控制不是新增模型试次；人工耗时、API成本和成功率未知。","evidenceStatus":"active"},{"id":"W-F02-intuit-agent","statement":"官方会计产品声明支持对账及异常处理，但本页仍以冻结快照独立重算接收。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-INTUIT-AGENT","locator":"Accounting Agent、Finance Agent与产品可用性段落","relation":"partial","supports":"官方会计产品声明支持对账及异常处理，但本页仍以冻结快照独立重算接收。","limit":"产品自报非独立结果；没有本站账套、错误分母或人工介入完整日志，不使用客户节时百分比。"}],"progressIds":["P-FIN-ACCOUNTING-PRODUCT"],"configuration":"Intuit Enterprise Suite发布声明；无本输入试次，未取得正确率/重试/总成本。","limitation":"产品自报非独立结果；没有本站账套、错误分母或人工介入完整日志，不使用客户节时百分比。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"公开代码或示例产物可读；需要另配模型、数据和重算引擎，未本站运行。","tools":[{"name":"Beancount固定代码：账户、币种与交易平衡校验","url":"https://github.com/beancount/beancount/blob/9747213802acb733ce53acc6e963f5961f47d6cd/beancount/ops/validation.py#L194-L410","version":"固定commit见URL","access":"公开读取，执行需自行配置依赖与合法数据/模型接口；成本未知"}]},"variants":[{"title":"日常净额对账","inputDifference":"同币种稳定主键，状态和期间固定","assessmentChange":"可交完整候选并独立重算","acceptanceChange":"核查连接基数、净额守恒和异常"},{"title":"多币种及迟到记录","inputDifference":"汇率价值日、费用和退款跨期增加","assessmentChange":"先固定新增规则再扩围，不继承同币种成绩","acceptanceChange":"按币种及价值日分账，对照银行回单"}],"evidenceStatus":"active"},{"id":"W-F03","domainId":"finance","title":"来源化研究备忘录：事实、假设与反证可追溯","goal":"让研究负责人收到可核查、可用于讨论的一份分析，而不只有流畅观点。","trigger":"需要回答公司、行业或财务事件的具体问题，并给定资料范围和截止时间。","scope":"交付事实、计算、情景和未决问题；不自动买卖资产，不把备忘录质量等同投资收益。","inputs":[{"name":"研究问题","detail":"决策用途、读者、截止时间、覆盖对象与重要性标准。"},{"name":"可用资料","detail":"报告、公告、电话会记录及授权数据；标识原始与转述来源。"}],"deliverables":["有来源定位的事实与数字表","区分事实、假设和推断的结论","关键反证、情景敏感性与待补信息"],"acceptance":[{"criterion":"关键主张有直接支持","method":"打开原文定位，核对日期、口径和实际表达","failure":"引用只提到主题，却不支持增长或估值结论。"},{"criterion":"计算和情景可复算","method":"保存公式、输入、单位和假设，独立重算关键项","failure":"把管理层目标当已实现业绩，或把情景当预测事实。"},{"criterion":"相反证据影响结论","method":"核查主要反证与资料缺口，说明何种事实会改判","failure":"只收集支持最初观点的资料，隐藏重要重述。"}],"assessment":{"recommendation":"可让AI整理证据、起草论证和对照情景；事实与计算可逐条接收，开放投资判断仍需按用途复核。","conditions":["问题和资料截止日具体","关键原文可访问且来源可追溯","结论需说明假设和改判条件"],"evidenceBasis":"直接研究评测和示例报告支持同类来源化研究的候选交付；来源有效、计算正确和投资价值不是同一证据。","localStatus":"本站未测该研究备忘录在真实决策中的质量或收益。","gaps":["选择同题资料包校准遗漏和错误引用","记录信息更新后的改判，不用篇幅或引用数替代质量"],"mode":"staged-delivery","rationale":"已有同类财务研究Agent任务结果和可检查报告，支持将取数、来源核对、计算与反证整合为候选；研究评分、证据口径和投资收益仍分开。","frontier":"FinFIRST已把可追踪证据和答案同时计分；Finance Agent有SEC工具研究。前者例子仍须审查地域口径，后者题目结果不覆盖整篇投资备忘录。","nextValidation":"同一问题资料包加入重述、相反证据和过期事实，检查引用是否支持每项结论及如何改判。","futureSignals":["新披露、重述和反证到来时，关键主张能持续正确更新并保留出处，才扩研究范围","引用支持主题却不支持数值，或实体期间/地域错配出现时，拒收相应研究结论"]},"workflow":{"before":"研究员检索、摘录、抄数、写稿并逐项校核。","now":"AI在同一问题下检索、计算、反证与修订；人集中判断重要性和决策使用范围。","retained":["资料适用性","关键假设和采纳决定"],"added":["主张到原文的可审计关联","防止过期事实沿用","维护反证与更新条件"]},"humanResponsibility":"确定研究用途、重要假设和采纳范围；实质判断可以参考AI建议，但须有可追溯依据。","failureHandling":"关键资料缺失就缩窄结论、列补数路径；来源互相冲突时并列冲突，不能编造一致结论。","example":{"input":"解释季度毛利变化：报告显示产品收入构成变动，管理层另给全年目标。","output":"历史变化与来源表、按已知构成的计算、目标情景及需要补的成本信息。","failure":"把全年目标写成已经实现，并声称所有毛利变化都来自AI降本。"},"evidenceLinks":[{"sourceId":"WU-FINANCE-AGENTS","supports":"研究资料综合、模型与备忘录可形成连贯交付。","notCovered":"没有证明某项投资判断正确或回报。"},{"sourceId":"WU-EVAL","supports":"开放产物可按明确标准与专家校准验收。","notCovered":"模型裁判不自动成为金融判断真值。"},{"sourceId":"WU-FIN-FINROBOT-REPORT","supports":"公开可取的报告产物入口。","notCovered":"只核对产物存在及来源身份，没有复算全报告，不作为质量或收益独立验证。"},{"sourceId":"WU-FIN-FINFIRST","supports":"在123题统一工具的研究配置中，作者报告可检查答案与证据的完成结果；分项高分不等于整题通过。","notCovered":"作者单次指定运行，自动裁判；Table2分子含美国加拿大而分母是美国，地域口径需独立审查，不能把高分当研究绝对正确。"},{"sourceId":"WU-FIN-FINANCEAGENT","supports":"早期SEC资料研究Agent已有有工具、题目、答案和评分的评估，而不只有报告演示。","notCovered":"旧配置；537题与公测50/私有150/测试337分母不能互换，不支持整套投资工作或收益。"}],"resourceRefs":[{"module":"boundary","type":"task","id":"F01","version":"2026-10-19.boundary.1","relation":"adjacent","reason":"财报数值核对可支持备忘录中的财务事实；不覆盖全部研究结论。"}],"career":{"direction":"行业研究、信用研究、投研支持","evidence":"提交一份可追溯备忘录，以及新证据到来后如何修订结论的记录。","nextStep":"按目标岗位核对行业、财务、地域和资质要求；将资料整理能力与独立研究判断分别补证。"},"team":{"bottleneck":"研究员重复收集信息，还是重要假设没有人负责？","allocation":"AI维护证据与初稿；负责人聚焦反证、关键口径与决策用途，复核者抽查关键来源。","nextStep":"用同一资料包比较遗漏、引用错误和复核时间；判断该投工具、培养研究判断还是补行业人员。"},"sourceIds":["WU-FINANCE-AGENTS","WU-EVAL","WU-FIN-FINROBOT-REPORT","WU-FIN-FINFIRST","WU-FIN-FINANCEAGENT"],"reviewedAt":"2026-10-21","claims":[{"id":"W-F03-method","statement":"已有公开研究报告产物可取得。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-FIN-FINROBOT-REPORT","locator":"report/NVDA_report.pdf；同commit README的V0安装入口","relation":"partial","supports":"公开可取的报告产物入口。","limit":"只核对产物存在及来源身份，没有复算全报告，不作为质量或收益独立验证。"}],"progressIds":[],"configuration":null,"limitation":"只核对产物存在及来源身份，没有复算全报告，不作为质量或收益独立验证。","evidenceStatus":"active"},{"id":"W-F03-adoption","statement":"已有同类财务研究Agent任务结果和可检查报告，支持将取数、来源核对、计算与反证整合为候选；研究评分、证据口径和投资收益仍分开。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-FINROBOT-REPORT","locator":"report/NVDA_report.pdf；同commit README的V0安装入口","relation":"adjacent","supports":"局部工具/产物配合本任务明确可检查结果，构成试用推理依据。","limit":"无同输入整包AI运行记录，不推一般完成率。"},{"sourceId":"WU-FIN-FINFIRST","locator":"§3.1—3.4、§4.1 Table3、Table2例子","relation":"partial","supports":"123题统一搜索/网页/Python配置的外部Agent研究。","limit":"作者单次指定运行，自动裁判；Table2分子含美国加拿大而分母是美国，地域口径需独立审查，不能把高分当研究绝对正确。"},{"sourceId":"WU-FIN-FINANCEAGENT","locator":"§2.2 Dataset split、§3、§4与AppendixC","relation":"partial","supports":"专家问题、原始披露与工具配置的研究任务证据。","limit":"旧配置；537题与公测50/私有150/测试337分母不能互换，不支持整套投资工作或收益。"}],"progressIds":[],"configuration":null,"limitation":"本站旧参考答案与校准控制不是新增模型试次；人工耗时、API成本和成功率未知。","evidenceStatus":"active"},{"id":"W-F03-finfirst","statement":"在123题统一工具的研究配置中，作者报告可检查答案与证据的完成结果；分项高分不等于整题通过。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-FINFIRST","locator":"§3.1—3.4、§4.1 Table3、Table2例子","relation":"partial","supports":"在123题统一工具的研究配置中，作者报告可检查答案与证据的完成结果；分项高分不等于整题通过。","limit":"作者单次指定运行，自动裁判；Table2分子含美国加拿大而分母是美国，地域口径需独立审查，不能把高分当研究绝对正确。"}],"progressIds":["P-FIN-FIRST"],"configuration":"15配置，搜索/网页/Python；单次指定运行；GLM-5.1裁判，缺失结果留在123分母；本文不转写职业成功率。","limitation":"作者单次指定运行，自动裁判；Table2分子含美国加拿大而分母是美国，地域口径需独立审查，不能把高分当研究绝对正确。","evidenceStatus":"active"},{"id":"W-F03-financeagent","statement":"早期SEC资料研究Agent已有有工具、题目、答案和评分的评估，而不只有报告演示。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-FINANCEAGENT","locator":"§2.2 Dataset split、§3、§4与AppendixC","relation":"partial","supports":"早期SEC资料研究Agent已有有工具、题目、答案和评分的评估，而不只有报告演示。","limit":"旧配置；537题与公测50/私有150/测试337分母不能互换，不支持整套投资工作或收益。"}],"progressIds":[],"configuration":"原文537题；Google/EDGAR等工具；50公开/150申请/337私有，全部题结果与测试榜分母不同。","limitation":"旧配置；537题与公测50/私有150/测试337分母不能互换，不支持整套投资工作或收益。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"可读固定版本代码说明及报告产物；报告PDF不是运行工具，本专业整包任务尚未安装或重放，数据接口、API及人时成本未知。","tools":[{"name":"FinRobot固定项目与安装说明","url":"https://github.com/AI4Finance-Foundation/FinRobot/tree/2717499b8e30f242640af08c4ad9afd1113c2d45","version":"commit 2717499b8e30f242640af08c4ad9afd1113c2d45","access":"开放项目说明；模型、数据、依赖需自行准备，未验证本专业任务可一键执行。"},{"name":"FinRobot已有NVDA示例研究报告","url":"https://github.com/AI4Finance-Foundation/FinRobot/blob/2717499b8e30f242640af08c4ad9afd1113c2d45/finrobot_autogen/report/NVDA_report.pdf","version":"固定commit见URL","access":"公开读取，执行需自行配置依赖与合法数据/模型接口；成本未知"}]},"variants":[{"title":"公司事件研究","inputDifference":"截止日公开披露，关键原文可访问","assessmentChange":"事实与计算可逐项接收，投资推断单独标注","acceptanceChange":"重述、反证及假设改变应实际影响结论"},{"title":"信用或交易尽调研究","inputDifference":"内部证据、契约和交易用途不同","assessmentChange":"只迁移资料组织；法律效力和独立取证另验","acceptanceChange":"补借款还款或交割条件专属反证"}],"evidenceStatus":"active"},{"id":"W-A01","domainId":"analysis","title":"指标报表：交付可重放的查询、图表与解释","goal":"将业务问题变成口径明确、数据可追溯、可重放的查询和分析包，能区分真实变化、数据缺失与口径误差。","trigger":"运营问“本周付款用户是否下滑、变化集中在哪些渠道”，已授权查询交易/账号/渠道数据并提供可冻结快照。","scope":"业务问题澄清、数据发现、查询、质量诊断和可复算解释属于一次交付；描述性分解不自动提供因果归因。稳定指标在口径接收后固化查询刷新。","inputs":[{"name":"业务问题与可比窗口","detail":"要回答付款人数变化及渠道贡献；上海时区、完整周/同比窗、付款状态、测试账号排除、未知渠道和退款处理；有争议时呈现选项和影响。"},{"name":"数据语义与访问","detail":"冻结数据快照/分区、schema及版本、主外键与粒度、数据生成代码/血缘、延迟/补发规则、脱敏只读查询入口与扫描预算。"},{"name":"独立控制","detail":"经认可的参考切片/总账控制或独立查询；指定周日16:00 UTC跨周边界、重复支付事件、多渠道映射和null账号记录。"}],"deliverables":["问题与指标定义，列出原问题到查询定义的取舍及未决语义","参数化SQL/计算代码、引擎与快照版本、运行命令/查询ID、日志和底层结果表","可复算周环比与渠道分解、图表和一页解释；未知渠道单列，渠道人数不满足可加性时明确标注","对账/连接基数/新鲜度/缺失检查与未解决异常；固定查询刷新入口及schema变化失败信号"],"acceptance":[{"criterion":"问题定义与数据口径一致","method":"接收者核对用户/事件粒度、时区与完整比较窗；独立参考切片覆盖跨周、重复、测试账号、退款与未知账户；SQL语法可不同但结果应等价。","failure":"按事件行数代替人数；不完整本周对整周；缺账号信息默认真实账号。"},{"criterion":"连接与控制数不产生虚假变化","method":"独立检查每次join前后主键数和行数；按接收规则处理渠道一对多；总人数对账，若渠道非互斥不把渠道人数和当总人数。","failure":"维表多版本使用户翻倍；把未知渠道删去后解释为渠道下降。"},{"criterion":"交付可重放","method":"验收者在只读环境用冻结快照、参数、引擎重跑SQL/代码，并从结果重建图表；稳定报表刷新使用已接收程序。","failure":"只交图片或查询文本而缺表版本；刷新时模型另写SQL导致口径漂移。"},{"criterion":"解释不过度声称且缺数据可见","method":"逐条把说明数字回指结果，核对完整性/延迟；制造缺分区或空结果，要求返回数据异常而非业务零值；描述性分解明确不证明原因。","failure":"null变零；因营销上线时间重合直接声称其导致下降；文字使用不同期间数字。"}],"assessment":{"recommendation":"有数据语义、只读查询和独立控制时，可整包委派业务问题到可重放分析；关键口径与结果通过核对后固化刷新。","conditions":["数据血缘/生成逻辑和权限可访问","比较窗口与指标语义可接收","关键结果有独立参考切片"],"evidenceBasis":"OpenAI内部数据Agent是厂商自报工程案例；公开data analyst应用提供实现参考。它们展示组织与入口，不提供本例成功率。","localStatus":"本例尚未查询数据或运行Agent；运行数、通过率、费用和人工介入量均为null。","gaps":["真实业务schema/生成逻辑未绑定","未知问题、schema变化与数据延迟条件尚未试验"],"mode":"bounded-delivery","rationale":"内部案例展示从查表到分析的整包路径，公开应用支持可见查询；迁移成立的关键是能提供业务语义与真实只读反馈。因此建议委派完整候选分析，用独立控制验收，不能把内部案例直接写成企业即插即用。","frontier":"外部内部工程案例已把动态数据发现、查询修正、报告与上下文组合；公开参考实现不等于原公司专用Agent可下载。","nextValidation":"构造授权数据的小快照，先用独立控制检验跨周/重复/缺分区等验收；再给代理原业务问题且隐藏参考SQL，核对结果并保存澄清、重试、查询扫描量和人工时间。","futureSignals":["扩大：新问题与schema/埋点变化下，代理能识别粒度差异、修正连接并在未知数据时正确返回未解决。","收缩：依靠历史固定SQL才对，或语义层过期导致稳定错数；缩到已认可报表刷新并修复语义资产。"]},"workflow":{"before":"找表、SQL、导出、作图和口径解释反复交接；执行者凭列名猜测语义。","now":"代理依据问题、语义和只读反馈自行选择数据与探索顺序；发现空结果或异常基数时诊断修订；已接收的稳定报表直接执行程序。无需规定每个问题统一查几张表。","retained":["指标用途和取舍确认","实际决策的接收责任"],"added":["查询到数据生成代码的语义核对","查询与底层结果同时交付","保留口径修正版本且不自动把猜测变为长期事实"]},"humanResponsibility":"明确业务决策及可接受口径，处理语义冲突并接收结论；发现表、写SQL、运行诊断可由代理承担。若要因果结论，另确定识别设计，不由报表验收冒充。","failureHandling":"表语义不明时列出字段、候选解释和对结果的影响；连接后空集先检查授权、时窗和粒度；缺分区时标“不完整”，不造业务数值。预算用完保留查询与已知范围，未解决不写成下滑。","example":{"input":"设计例：上海时区上一完整周相对前一周付款去重用户变化及渠道贡献；只读payment_events/users/channel_map；测试账号排除，多付款去重，未知渠道保留；快照和真实schema待提供。","output":"参数化查询与快照记录、周人数/分解表、控制切片和连接诊断、对应图表/说明；如用户跨渠道，给非可加人数或按批准归属规则重算。没有本例真实结果。 跨渠道用户存在时，可展示非互斥人数，但不能称可相加贡献；要合计贡献应先约定互斥归属或可审计分解。","failure":"周日16:00 UTC后的付款落错周；维表join重复用户；缺失分区写成0后宣布业务下滑。"},"evidenceLinks":[{"sourceId":"WU-DATA-AGENT","supports":"支持在语义上下文与结果反馈下组织完整分析任务；支持给目标、让代理自选路径并接入生成语义。","notCovered":"无本企业题目通过率；不是可下载同配置产品；周付款/渠道规则与控制为本站任务设计，未实测。"},{"sourceId":"WU-DATA-APP","supports":"提供将查询与结果交给使用者核对的实现方式；支持可追溯交付。","notCovered":"未在本站安装，版本需固定，目标数据需另适配；不提供未知业务问题的稳定性证据。"}],"resourceRefs":[{"module":"catalog","type":"task","id":"DATA-02","version":"2026-10-02.2","relation":"partial","reason":"沿用跨工具指标契约的口径与验收；新层承接完整报表交付。"}],"career":{"direction":"数据分析、分析工程、运营分析","evidence":"提交指标契约、实际查询、参考切片及口径变化后的修正记录。","nextStep":"对照目标岗位的数据工具、业务和统计深度；区分能做报表、能解释变化与能设计实验。"},"team":{"bottleneck":"重复查询多，还是表定义和指标口径分散？","allocation":"AI处理探索性查询，固定程序刷新稳定指标；分析人员维护语义与决策解释。","nextStep":"选一个每周报表，测准备、复核和刷新成本；优先修口径和接口，再选择工具、培养或招聘。"},"sourceIds":["WU-DATA-AGENT","WU-DATA-APP"],"reviewedAt":"2026-10-20","claims":[{"id":"A01-INTERNAL","statement":"OpenAI公开的内部数据Agent案例描述了查表、SQL、结果修正到报告的闭环，但这是专用内部系统的厂商自报。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-DATA-AGENT","locator":"How it works；Context is everything；Moving fast without breaking trust","relation":"partial","supports":"支持在语义上下文与结果反馈下组织完整分析任务。","limit":"无本企业题目通过率；不是可下载同配置产品。"}],"progressIds":["P-A-INTERNAL"],"configuration":"2026-01-29工程案例；文中GPT-5.2，专用数据/权限/语义与记忆；评价使用人工golden SQL及结果比较，未公开本例样本/重试/成本分母。","limitation":"本站未进行该任务的模型运行，不能据此给出本地成功率。","evidenceStatus":"active"},{"id":"A01-APP","statement":"官方公开data analyst应用可作为查询和结果可见的工程入口，不能把应用参考当能力试验。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-DATA-APP","locator":"readme：data analyst app实现及SQL展示/记忆修正","relation":"method","supports":"提供将查询与结果交给使用者核对的实现方式。","limit":"未在本站安装，版本需固定，目标数据需另适配。"}],"progressIds":["P-A-APP"],"configuration":null,"limitation":"本站未进行该任务的模型运行，不能据此给出本地成功率。","evidenceStatus":"active"},{"id":"A01-DELEGATE","statement":"本例可在独立口径/结果控制下整包委派；稳定接收后转为参数化刷新，异常才重新探索。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-DATA-AGENT","locator":"Lessons #2 Guide the Goal, Not the Path；#3 Meaning Lives in Code","relation":"adjacent","supports":"支持给目标、让代理自选路径并接入生成语义。","limit":"周付款/渠道规则与控制为本站任务设计，未实测。"},{"sourceId":"WU-DATA-APP","locator":"SQL query/result visibility","relation":"method","supports":"支持可追溯交付。","limit":"不提供未知业务问题的稳定性证据。"}],"progressIds":[],"configuration":null,"limitation":"本站未进行该任务的模型运行，不能据此给出本地成功率。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"内部Agent不可直接获取；公开示例提供自建入口。只读仓库/引擎、API账户和语义资产需使用方提供，本例未连库。","tools":[{"name":"官方data analyst应用参考","url":"https://developers.openai.com/cookbook/examples/agents_api/apps/data_analyst/readme","version":"2026-09-10发布页，2026-10-20读取；代码commit需执行时锁定","access":"公开文档/示例，需模型服务及目标数据库适配；不可等同内部Agent。"}]},"variants":[{"title":"指标定义含混","inputDifference":"只问“用户有没有变少”，未说明活跃/付款/账号粒度。","assessmentChange":"先委派问题澄清与几种可比定义的影响分析；避免静默固定最方便的定义。","acceptanceChange":"接收者确认用途与口径后，才按该定义验收数字。"},{"title":"新鲜度或schema变化","inputDifference":"最新分区迟到、事件字段改名或生成逻辑已改。","assessmentChange":"先委派数据诊断，报表可交不完整范围，不能完整接收。","acceptanceChange":"以版本/新鲜度控制识别故障；修复后重跑原比较窗。"},{"title":"要解释活动的因果效果","inputDifference":"要求回答渠道活动是否导致付款下滑。","assessmentChange":"原报表只接收描述；另需实验/准实验识别任务及假设。","acceptanceChange":"增加对照、混杂与识别假设检查，不能由相关性图替代。"}],"evidenceStatus":"active"},{"id":"W-A02","domainId":"analysis","title":"组成变化诊断：拆清总体变化来自哪里","goal":"区分组内表现变化与人群/渠道构成变化，避免被总体指标误导。","trigger":"前后总体指标明显改变，决策者需要知道哪些可观察部分造成差异。","scope":"交付描述性分解和替代解释；仅凭前后数据不判断产品或策略的因果效果。","inputs":[{"name":"分组快照","detail":"前后相同粒度的分子分母、渠道/地区等分组、缺失和重复规则。"},{"name":"分解定义","detail":"固定参考权重与计算公式；明确百分点、相对变化及新增组处理。"}],"deliverables":["可复跑分解程序和结果表","组内与组成贡献、总变化核对","可支持解释与不能归因部分"],"acceptance":[{"criterion":"分组与总量对得上","method":"核对每组分母、权重与总分子分母，先处理零分母和重复行","failure":"平均各组比率而不按分母加权。"},{"criterion":"分解完整且定义明确","method":"验证within+mix=总变化；若改公式，解释交互项归属","failure":"混用不同权重的公式后把残差藏起来。"},{"criterion":"解释没有越过识别范围","method":"复算指标并检查原因措辞，区分百分点和相对百分比","failure":"把11%到19%写成“增长8%”，或称产品造成8个百分点提升。"}],"assessment":{"recommendation":"可委派AI计算和解释组成分解；结果可独立复算，但观察分解不能自动变成因果结论。","conditions":["前后数据粒度与口径一致","分解公式及缺失组处理已规定","解释保留未观察到的混杂与机制"],"evidenceBasis":"数据Agent方法支持多轮数据检查；本单元的有限算例有明确恒等式可验收。相邻数据处理案例不等于本题AI已通过。","localStatus":"代表例的参考/反例校准另有记录；未提供真实业务因果或新AI可靠率。","gaps":["真实业务新增/重叠分组、口径变化和解释措辞尚未以独立对账校准；因果效果不在分解内。"],"mode":"bounded-delivery","rationale":"固定公式后可直接交给确定性程序复算；AI负责口径异常、分组探索及解释，不必增加Agent循环。组合恒等式通过只证明计算，不能识别干预效果。","frontier":"既有数据Agent案例是邻近分析工作法；本页8个百分点是确定性算例，不是模型试验。","nextValidation":"盲测一个构成变化和一个组内变化的数据包，加入缺组、零分母、重复与新增组，对照独立求和并核对解释。","futureSignals":["口径变化和新分组下仍满足总量核对，解释不越过描述范围，可考虑扩大相近输入范围。","缺失/重叠分组导致贡献不可比，或把分解写成因果收益，应收窄或暂停相应采用。"]},"workflow":{"before":"人分别取数、算比率、看图并猜原因。","now":"AI把数据核对、精确分解和解释组成一包，按异常回查；计算固定后可自动刷新。","retained":["分组有何业务意义","因果归因所需条件"],"added":["显式分解定义","检查总体与组内方向不一致","记录替代解释"]},"humanResponsibility":"选择有业务意义的分组，说明要做的决策；根据需要补实验或机制证据，而非把描述性分解当决策结论。","failureHandling":"缺组、零分母或口径改变时标记不可直接比较项，给补数或重定义方案；不凭空平滑出完整故事。","example":{"input":"前：A900访问/90转化、B100/20；后：A100/10、B900/180。","output":"总体11%→19%，增加8个百分点；两组转化率不变，按固定公式组内贡献0、组成贡献8个百分点。","failure":"声称两个渠道都改善或产品提升了8个百分点；忽略高转化渠道占比增加。"},"evidenceLinks":[{"sourceId":"WU-EVAL","supports":"支持独立验收与允许等价实现。","notCovered":"恒等式不检验业务口径真实，也不提供干预因果解释。"},{"sourceId":"WU-DATA-AGENT","supports":"上下文核对是迁移依据。","notCovered":"内部系统不等于公开配置；来源未测试本站的组成分解。"}],"resourceRefs":[{"module":"catalog","type":"task","id":"DATA-03","version":"2026-10-02.2","relation":"partial","reason":"保留组成分解旧场景；新交付明确公式、异常处理与职业/团队接续。"}],"career":{"direction":"产品分析、增长分析、商业分析","evidence":"提供总量到分组的复算、一个反直觉案例，以及没有作因果归因的理由。","nextStep":"对照目标岗位是否要求实验设计或预测建模；缺少这些材料时不要用报表作品冒充全部分析能力。"},"team":{"bottleneck":"团队对同一总体变化反复给出互相矛盾的解释。","allocation":"AI和程序完成一致口径的分解；业务分析者检查分组意义，实验负责人决定是否继续识别效果。","nextStep":"选一个争议指标统一分解，再比较补数据、设计实验和直接行动的成本，不以报表数量决定编制。"},"exampleId":"analysis","sourceIds":["WU-EVAL","WU-DATA-AGENT"],"reviewedAt":"2026-10-20","claims":[{"id":"A02-FORMULA","statement":"明确分组分母、参考权重和新增组约定后，组内与组成贡献可由程序精确分解并以总量恒等式验收。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-EVAL","locator":"结果验收与关键反例","relation":"method","supports":"支持独立验收与允许等价实现。","limit":"恒等式不检验业务口径真实，也不提供干预因果解释。"}],"progressIds":[],"configuration":null,"limitation":"恒等式不检验业务口径真实，也不提供干预因果解释。","evidenceStatus":"active"},{"id":"A02-USE","statement":"AI可交付分组诊断、分解与解释，数据定义不一致时应先返回不同比较口径的影响；不承诺相邻分析系统在本题的成功率。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-DATA-AGENT","locator":"Context is everything","relation":"adjacent","supports":"上下文核对是迁移依据。","limit":"内部系统不等于公开配置；来源未测试本站的组成分解。"}],"progressIds":[],"configuration":null,"limitation":"内部系统不等于公开配置；来源未测试本站的组成分解。","evidenceStatus":"active"}],"variants":[{"title":"新增退出分组","inputDifference":"两期类别集合不同或前期零分母。","assessmentChange":"先明确不可比较部分；不能静默补零造出比率。","acceptanceChange":"输出可比分组、新增/退出贡献与总量桥接，约定交互项归属。"},{"title":"多维交叉与重叠归因","inputDifference":"渠道×地区与一个对象多标签。","assessmentChange":"可以探索，但固定基准后只按一个无重复归属定义合计。","acceptanceChange":"反向核对对象去重、层级总量和权重，避免贡献被算多次。"},{"title":"要求策略因果效果","inputDifference":"希望判断活动是否造成变化。","assessmentChange":"只接收描述性机制线索；改用W-A03识别设计。","acceptanceChange":"解释需保留混杂、选择与未观测变量，不能直接推荐因果行动。"}],"availability":{"status":"mixed","detail":"计算可用SQL/Python直接执行，Agent为可选组织方式；内部数据Agent不是可下载同配置服务。","tools":[{"name":"SQL/Python分析环境","url":"https://www.python.org/","version":"执行时固定Python、统计包和数据快照；本站未新运行","access":"公开运行时；业务数据/口径与查询权限由任务方提供"}]},"evidenceStatus":"active"},{"id":"W-A03","domainId":"analysis","title":"实验分析：估计效果，或说明为什么还不能判断","goal":"把“看起来变好了”转成有识别依据的效果估计与行动条件。","trigger":"实验负责人收到A/B数据，或业务希望从观察数据判断某项改变的效果。","scope":"先检查设计与数据，再交付估计和不确定性；条件不足时，“不能识别及补救方案”也是合格产物。","inputs":[{"name":"设计与问题","detail":"随机化单位、预设指标、分配概率、触发/暴露定义、观察窗和决策门槛。"},{"name":"完整记录","detail":"分配、暴露、结果、流失和排除记录；重复单位与并行实验信息。"}],"deliverables":["设计与数据有效性检查","有条件的效应估计、区间和稳健性分析，或不可识别说明","缺口的补数/重跑方案与上线决策所需材料"],"acceptance":[{"criterion":"样本进入机制可信","method":"核对分配比例、流失、重复单位、污染及排除规则","failure":"分组比例明显异常却直接报告显著提升。"},{"criterion":"估计与设计和问题一致","method":"确认分析单位、意向处理/暴露口径、相关性与预先约定的检验","failure":"按会话随机化却按独立用户事件计算过窄区间。"},{"criterion":"结论与不确定性相称","method":"报告效应与区间、实践门槛和探索性分析；说明未满足假设","failure":"挑一个显著分组或停止时点，宣称所有用户有效。"}],"assessment":{"recommendation":"可让AI编写分析、检查设计缺口和解释区间；因果结论只在数据与识别条件满足时接收。","conditions":["分配和暴露机制可追溯","指标、单位及排除规则明确","关键假设有设计或业务证据"],"evidenceBasis":"分析Agent提供计算与复核工作法；SRM研究说明分配异常会误导决策。此处主要是方法可行性判断，不宣称AI能凭数据补出缺失的随机化。","localStatus":"本站未测本单元的AI分析准确性；旧实验场景可作资料入口。","gaps":["目标实验的分配/暴露/流失机制与停止规则尚无完整数据包验证；观察研究识别假设需领域证据。"],"mode":"staged-delivery","rationale":"程序可整包计算并核对随机分配、区间与切片，但估计对象、暴露机制和停止规则约束因果解释；异常时交付不可识别诊断有实际价值。","frontier":"来源支持实验有效性检查与Agent查询方法；未提供自动识别任意因果关系的直接成绩。","nextValidation":"冻结随机化/排除/分析计划，用保留的真值模拟数据和异常分配控制校准后，再在真实完整日志上估计效应与区间。","futureSignals":["新实验下设计诊断、估计和区间经独立复算保持一致，可考虑扩大相近输入范围。","分配/暴露链缺失、事后挑选指标或停止点导致识别失败，应收窄或暂停相应采用。"]},"workflow":{"before":"分析师写检验、查看显著性，再补充设计解释。","now":"AI可把设计诊断、计算和敏感性检验合为循环；问题不可识别时返回需要的设计或数据，不硬给答案。","retained":["目标指标和决策门槛","识别假设与上线决定"],"added":["自动检查分配/暴露不一致","防止探索结果伪装成预设检验","记录何时应停止解释"]},"humanResponsibility":"确认干预目标、随机化与业务机制，决定在不确定性下采取什么行动；AI能提出诊断但不能创造不存在的对照。","failureHandling":"分配异常或严重流失时定位链路、标记受影响结果并提出补数/重跑；观察数据只输出相应条件下的解释。","example":{"input":"计划各半分组，实际有900名A、100名B；缺少分配日志，结果指标B更高。","output":"先报告样本比例异常与缺失机制，列需要的分配/曝光日志；暂不把差异解释为处理效果。","failure":"仅按两组均值做检验，把B的高值称为可直接上线的因果提升。"},"evidenceLinks":[{"sourceId":"WU-SRM","supports":"支持以设计与异常控制保护实验解释。","notCovered":"这是统计方法研究，不是AI模型成绩。"},{"sourceId":"WU-DATA-AGENT","supports":"邻近分析工作法支持自动组织计算。","notCovered":"Agent不能从缺失记录补出真实随机化；本站无该任务运行。"}],"resourceRefs":[{"module":"catalog","type":"task","id":"DATA-04","version":"2026-10-02.2","relation":"partial","reason":"沿用实验有效性与区间审核的旧场景；新单元把不可识别及补救作为明确交付。"}],"career":{"direction":"实验分析、产品数据科学、因果推断","evidence":"展示设计检查、估计区间、关键假设和一个应暂停结论的例子。","nextStep":"根据目标岗位补统计设计、数据工程或业务实验材料；将AI计算与本人设计判断分开呈现。"},"team":{"bottleneck":"实验跑得多，但分配、记录或解释质量不稳定。","allocation":"AI做可重复诊断和计算，实验/业务负责人核对识别与行动门槛；工程修数据链路。","nextStep":"按失败来源比较改实验平台、培养设计能力和招聘数据科学人员，先修最影响可信结论的链路。"},"sourceIds":["WU-SRM","WU-DATA-AGENT"],"reviewedAt":"2026-10-20","claims":[{"id":"A03-SRM","statement":"样本比例异常是实验数据有效性警报，须诊断分配/暴露/流失链路后再解释效果；比例正常也不能证明其他假设成立。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-SRM","locator":"Sample Ratio Mismatch诊断","relation":"method","supports":"支持以设计与异常控制保护实验解释。","limit":"这是统计方法研究，不是AI模型成绩。"}],"progressIds":[],"configuration":null,"limitation":"这是统计方法研究，不是AI模型成绩。","evidenceStatus":"active"},{"id":"A03-USE","statement":"可让Agent把设计核对、程序估计、稳健性与解释合为交付；若只有观察数据则先给识别假设和不可识别范围。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-DATA-AGENT","locator":"查询与上下文核对","relation":"adjacent","supports":"邻近分析工作法支持自动组织计算。","limit":"Agent不能从缺失记录补出真实随机化；本站无该任务运行。"}],"progressIds":[],"configuration":null,"limitation":"Agent不能从缺失记录补出真实随机化；本站无该任务运行。","evidenceStatus":"active"}],"variants":[{"title":"随机对照与聚类","inputDifference":"用户、地区或门店随机且结果有重复观测。","assessmentChange":"可分阶段接收，分析单位和相关结构先冻结。","acceptanceChange":"按随机化单位重算区间、SRM、流失与多重比较；不能把事件当独立用户。"},{"title":"只有观察数据","inputDifference":"上线前后或自选参与，没有随机分配。","assessmentChange":"优先交付识别方案/敏感性而非默认处理效应；可用工具估计条件效应。","acceptanceChange":"列因果图/对照构造、平行趋势或其他识别证据与反驳测试。"},{"title":"连续查看与多指标","inputDifference":"多个时点停止、切片挑选显著项。","assessmentChange":"预设停止/纠错方法后再给行动区间；探索需显式标注。","acceptanceChange":"记录所有查看与候选，区分探索和确认，复算合适不确定性。"}],"availability":{"status":"mixed","detail":"计算可用SQL/Python直接执行，Agent为可选组织方式；内部数据Agent不是可下载同配置服务。","tools":[{"name":"SQL/Python分析环境","url":"https://www.python.org/","version":"执行时固定Python、统计包和数据快照；本站未新运行","access":"公开运行时；业务数据/口径与查询权限由任务方提供"}]},"evidenceStatus":"active"},{"id":"W-M04","domainId":"math","title":"数学研究产物：联合探索命题、构造与证明或反例","goal":"围绕一个研究问题交付语义一致的结论及可核验论证/构造/反例，或明确已解决部分与未解障碍。","trigger":"研究者有问题方向或参数化对象，尚未锁定最终构造与证明；需要探索能成立的结论及失败边界。","scope":"允许在约定空间联合搜索对象、算法/构造参数、辅助命题与证明；不可偷偷改变原问题的对象域、量词和目标。自然语言论证、精确反例、形式证明有不同接收条件；新颖性另核对，不强制全部产物先转Lean。","inputs":[{"name":"问题与题意对照","detail":"原始问题和来源、对象域、量词、边界定义及已知结果；哪些是本题/变体、哪些目标不可改变。已有成果复现与新研究分开标记。"},{"name":"可调对象及目标","detail":"允许改动的构造、参数域、待优化界/常数与约束；若只可搜索辅助引理，明确禁止把反例不存在的假设加进原题。"},{"name":"交付类型和研究环境","detail":"约定自然论证/反例/形式证明中的可接受结果，精确计算工具及版本；形式任务固定Lean/Mathlib和公理；检索范围、预算与停止条件，成本未知不假定可重现。"}],"deliverables":["问题定义与自然语言/符号声明对照，列出原题、变体、已知成果、新猜想和未决歧义","最终对象/参数/构造及性质保证成对交付；包含精确证书或证明、引理依赖、计算程序与复现命令（按产物适用）","研究记录：关键路线、被反例排除的范围、尚未解的具体目标；不是固定逐步操作日志","可读研究说明、公开来源定位及贡献/修订记录；形式成果另附项目锁定文件、构建/公理核查输出"],"acceptance":[{"criterion":"先确认研究的就是原问题或明示变体","method":"独立核对对象域、量词、下/上/自然密度或其它关键定义与原文；列题意对照，不把弱化结论当原题答案。","failure":"把“下密度>0”换成“上密度>0”；允许空集造成平凡构造仍宣称解决要求非平凡对象的题。"},{"criterion":"构造和保证相互对应","method":"独立重算候选对象/参数，核对每项约束与最终证明覆盖；改参数后旧证明必须同步重验。","failure":"数值搜索找到一个好参数但保证只证旧参数；浮点近似当精确反例。"},{"criterion":"自然论证/反例按各自对象接收","method":"自然证明由具备领域能力的独立接收者检查全部关键引理、边界与引用，可用工具补检但不宣称内核认证；有限反例精确重算并验证全部前提；渐近反例需无限构造和相应论证。","failure":"有限搜索未发现反例被当证明；自然语言关键步骤循环引用；有限n上的异常被当下密度的反例。"},{"criterion":"形式证明额外满足内核与语义","method":"只对约定形式交付：固定项目干净构建、比较可信目标声明、检查传递公理拒绝sorryAx/未批准公理，并核对可读说明与实际定理覆盖。","failure":"编译通过但依赖含sorry；形式化只覆盖支撑引理却写成主结论已证；随意增加假设。"},{"criterion":"区分完成、未解决与新颖性","method":"交付按实际覆盖标识完整结论/条件结论/反例/未解决；独立查文献与修订记录，已知成果复现不计新发现。","failure":"预算耗尽宣布命题为假；重放公开证明标“首次解决”；撤回后仍引用原结论。"}],"assessment":{"recommendation":"可把问题族探索与候选研究产物整包交给AI；按约定产物分别接收。已有公开个案证明研究循环可产生完整成果，任意新问题的完成概率与公开工具复现能力仍待核。","conditions":["原题和可调对象边界可核对","产物有独立验收途径","研究系统与实际可调用工具身份分开"],"evidenceBasis":"Nexus作者报告联合搜索参数与证明，并有Erdős125固定证明产物；不是把W-M01固定形式命题的受限本地校准外推到研究。","localStatus":"本站M04模型运行数、成功率、成本及人工时间均为null；只读过Erdős125公开证明文本，未独立构建，不继承W-M01本地控制通过。","gaps":["公开服务在新问题及限定预算下的复现尚未知","自然论证需领域独立审阅；参数联合搜索个案不等于一般数学研究可靠性"],"mode":"candidate-exploration","rationale":"外部研究及固定产物使“仅可局部辅助”的一概判断不成立；但研究系统非公开服务，个案和选择过的问题集也不支持任意新题概率。因此可委派完整探索和候选产物，是否完成由语义、正确性、覆盖与新颖性分别接收；已知复现可直接接收可重放成果，新研究保留未解决出口。 这一发布进一步支持以具体研究产物扩大观察范围，但其内部配置、人工整理和局部形式化不能压成统一自主成功率。","frontier":"Nexus研究配置包含进化搜索/AlphaProof并联合探索算法参数与证明；其公开结果可逐项检查。Erdős125是已公开成果复现，不是仍待解决的新问题，也不证明全套系统可调用。 OpenAI十月另公开内部模型研究手稿及部分形式产物，存在不同验证阶段和撤回/修订；这些公开成果属于M04直接证据，不只作为审稿背景。","nextValidation":"先在固定Nexus提交重建Erdős125并做题意/传递公理检查，校准复现验收；另选未泄漏问题或允许参数变化的独立小问题，用公开可用工具记录搜索、失败与实际人时。两类结果分开，未运行不估成功率。","futureSignals":["扩大：新问题上，对象/参数与保证共同改进且独立复核成立；题意、文献新颖性和资源记录完备，公开工具可重现。","收缩：表面成功依靠弱化目标、无证引理或挑选成功个案；改参数后保证失效；先退到已证引理/候选构造。"]},"workflow":{"before":"先凭经验挑构造或参数，再尝试证明；数值试验、证明草图和形式化相互交接，反馈较晚。","now":"代理可在允许的对象/参数空间尝试构造，同时寻找证明或反例；精确计算、文献或形式错误反馈可以改变辅助命题与搜索方向。路径依题而定，确定性穷举可直接运行；不规定所有数学问题都要并行或Lean。","retained":["原题及研究范围核对","结果意义/新颖性核查","允许依赖与资源投入"],"added":["对象与保证成对版本化","自然论证和形式产物覆盖对照","失败范围与修订传播"]},"humanResponsibility":"研究接收者确认问题含义、允许变体、领域论证与贡献定位并承担发布责任；这些核对也可由AI协助，不写成永久能力分工。形式内核只承担指定声明的逻辑检查。","failureHandling":"给出已证引理、候选构造、排除参数范围和最小未解目标，区分预算耗尽、环境不可用、语义冲突和真实反例；不靠sorry或新增假设伪装完成。已有材料撤回则检查其依赖的候选与说明。","example":{"input":"已有成果复现：A是三进制只含0、1的非负整数，B是四进制只含0、1的非负整数，判断A+B下渐近密度是否为正（计数比例liminf）。原命题不可改；有限计数只作探索。新研究变体可另约定改变进制/数字集，但须另立目标，不能继承原结论。","output":"固定公开文件在0647711a…下的target_theorem_0把该正下密度命题答案填False并给证明；交付需题意对照、可读稿、固定依赖与独立检查。本轮仅文本核对，未重放。成果入口见WU-NEXUS-125。 来源仓库将可读稿说明为沿Lean结构整理的人工稿，不当作未经整理的AI运行轨迹。","failure":"观测若干有限上界下比例较小就宣布下密度为零；把正下密度换成别的密度概念；把已公开Erdős125重新标为新发现。"},"evidenceLinks":[{"sourceId":"WU-NEXUS","supports":"支持对象/参数与证明联合探索的研究组织，及未证引理等失败模式；支持研究循环保留语义核对和未完成出口。","notCovered":"内部研究配置，不等于公开API；不把Erdős125说成参数优化实验；不同产物的具体接收条目是编辑设计。"},{"sourceId":"WU-NEXUS-125","supports":"给出本代表例可检查的对象和目标定理入口；区分研究结果集合与可运行代理。","notCovered":"只查文本；尚未本站构建/独立语义裁定，不代表新研究成功率；可下载证明不等于生成系统可获得。"},{"sourceId":"WU-LEAN","supports":"支持形式交付的公理与可信目标核对。","notCovered":"Lean文档不认证自然语言论证或数学新颖性。"},{"sourceId":"WU-ARISTOTLE-SDK210","supports":"证明工具的可调用候选入口。","notCovered":"未调用；不承诺能完成本例或复现Nexus配置。"},{"sourceId":"WU-MATH-RELEASE","supports":"内部模型研究产物的直接发布及状态。","notCovered":"不是全部开放题独立解决或公开服务稳定性。"},{"sourceId":"WU-M-MATH-OCT-FIXED","supports":"固定手稿、部分形式化、人工整理和修订入口。","notCovered":"尚未逐稿独立重放/学术裁定。"}],"resourceRefs":[],"career":{"direction":"数学研究、计算/形式化数学、研究工程","evidence":"展示原题、构造与保证、题意/文献差异、失败边界及适用的独立核验；已知复现清楚标识。","nextStep":"按目标领域选择真实问题与接收者；区分构造发现、论证正确、形式化和新颖贡献，补齐最弱的一层。"},"team":{"bottleneck":"瓶颈在候选对象、证明路线、形式库还是题意与文献定位？","allocation":"以研究产物委派代理，自行组织检索/计算/证明；领域接收者独立核对语义及贡献，工具核查适用的精确证书。","nextStep":"分别做已知成果复现和未见问题探索，记录完整成本与未解决路线；不能只展示成功题。"},"sourceIds":["WU-NEXUS","WU-NEXUS-125","WU-LEAN","WU-ARISTOTLE-SDK210","WU-MATH-RELEASE","WU-M-MATH-OCT-FIXED"],"reviewedAt":"2026-10-20","claims":[{"id":"M04-RESEARCH","statement":"Nexus论文报告联合搜索算法参数与证明保证；对象设计和证明搜索可以互相反馈。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-NEXUS","locator":"§4 Optimization Theory；§2 Agent Architecture；§3 Failure Analysis","relation":"partial","supports":"支持对象/参数与证明联合探索的研究组织，及未证引理等失败模式。","limit":"内部研究配置，不等于公开API；不把Erdős125说成参数优化实验。"}],"progressIds":["P-M-NEXUS"],"configuration":"arXiv2605.22763v1；完整研究代理含Gemini3.1Pro基础代理、进化搜索及AlphaProof；论文不同实验预算不同；成功个案成本不含完整选题搜索。本站未重放。","limitation":"本站未进行该任务的模型运行，不能据此给出本地成功率。","evidenceStatus":"active"},{"id":"M04-ARTIFACT","statement":"Erdős125正下密度变体有固定公开Lean文件及可读稿，可将“研究成果”落到具体声明而非榜单。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-NEXUS-125","locator":"L42-L54定义；L342-L370 target_theorem_0；同提交自然语言PDF","relation":"direct","supports":"给出本代表例可检查的对象和目标定理入口。","limit":"只查文本；尚未本站构建/独立语义裁定，不代表新研究成功率。"}],"progressIds":[],"configuration":null,"limitation":"固定公开产物存在不等于本站验收通过，已知复现不计新发现。","evidenceStatus":"active"},{"id":"M04-VALIDATION","statement":"自然证明、精确反例与形式证明分别接收；形式证明还要查原声明、传递公理和语义对应。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-LEAN","locator":"Validating a Lean Proof：Printing Axioms / Gold Standard","relation":"method","supports":"支持形式交付的公理与可信目标核对。","limit":"Lean文档不认证自然语言论证或数学新颖性。"},{"sourceId":"WU-NEXUS","locator":"§3 Failure Analysis；§6 Discussion：semantic verification","relation":"method","supports":"支持研究循环保留语义核对和未完成出口。","limit":"不同产物的具体接收条目是编辑设计。"}],"progressIds":[],"configuration":null,"limitation":"本站未进行该任务的模型运行，不能据此给出本地成功率。","evidenceStatus":"active"},{"id":"M04-ACCESS","statement":"Aristotle有SDK/CLI/API的官方分发入口，可作为公开证明工具候选；Nexus结果仓库不能当作Nexus服务入口。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-ARISTOTLE-SDK210","locator":"2.1.0 Quick Start / Authentication","relation":"method","supports":"证明工具的可调用候选入口。","limit":"未调用；不承诺能完成本例或复现Nexus配置。"},{"sourceId":"WU-NEXUS-125","locator":"同提交仓库README：results collection / not official Google product","relation":"direct","supports":"区分研究结果集合与可运行代理。","limit":"可下载证明不等于生成系统可获得。"}],"progressIds":[],"configuration":null,"limitation":"本站未进行该任务的模型运行，不能据此给出本地成功率。","evidenceStatus":"active"},{"id":"M04-OCT-RELEASE","statement":"OpenAI十月公开了内部模型产生的数学研究手稿与部分形式产物，并保留撤回/修订；这直接支持研究产物可逐项审查，不能改写为稳定独立解决大量开放问题。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-MATH-RELEASE","locator":"2026-10-06发布；10-15更新内容；对应公开仓库入口","relation":"direct","supports":"内部研究模型已公开具体研究产物及验证阶段。","limit":"发布者自报；10-15资助/合作不表示能力新提升；未展示所有公开产品同能力。"},{"sourceId":"WU-M-MATH-OCT-FIXED","locator":"README L12-L20/L40-L44；history.md L5-L31；ComparatorChallenges README L12-L28","relation":"direct","supports":"固定稿件、模型/人工整理、修订与形式化覆盖可追溯。","limit":"局部形式化和未独立核验需逐稿标注；文本读取不等于全部成果接收。"}],"progressIds":["P-M-RELEASE"],"configuration":"2026-10-06发布，固定1581d4cb…（10-12）；未发布内部模型；不同成果程序有例外及人工可读性编辑；README算力等值不是实际墙钟/报价，本文不估成功率或完整选题成本。","limitation":"研究产物、正确性、文献新颖性、公开工具可复现四件事分开；撤回/更正须传播到相关主张和引用。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"Nexus为研究系统，开放的是结果集合；Lean项目可重建，Aristotle有公开鉴权入口。三者不合并为“现在已能复现Nexus”。本站未调用付费服务。","tools":[{"name":"Nexus固定成果仓库","url":"https://github.com/google-deepmind/alphaproof-nexus-results/tree/0647711a71183c1ea492ad60860776617ce1ea88","version":"0647711a…；Lean4.27.0；lake-manifest锁定依赖","access":"公开结果文本与项目；生成代理不作为公开产品提供。"},{"name":"Lean proof validation","url":"https://lean-lang.org/doc/reference/latest/ValidatingProofs/","version":"实际复现使用仓库Lean4.27.0；参考文档2026-10-20读取","access":"开源内核/工具链；需安装固定依赖；文档latest不替代项目版本。"},{"name":"Aristotle CLI/API","url":"https://pypi.org/project/aristotlelib/2.1.0/","version":"aristotlelib2.1.0","access":"需账号/API key与服务预算；本站未调用；SDK链接的GitHub源码入口此前404，不声称已核实源码。"}]},"variants":[{"title":"自然语言研究命题","inputDifference":"尚无可靠形式化，可能存在含混定义。","assessmentChange":"可整包探索候选论证/反例，同时产出题意核对；不以缺Lean认定不能研究。","acceptanceChange":"领域独立审阅关键推导、引用和边界；不能标内核认证。"},{"title":"构造/参数优化","inputDifference":"允许改变算法步长、常数或数学对象，目标约束固定。","assessmentChange":"可联合探索对象和保证；采用建议仍取决于最终证书，不只看最优数值。","acceptanceChange":"构造版本、参数精确值与性质证明成对；换对象后保证重新检查。"},{"title":"已锁定形式目标","inputDifference":"只允许修改proof，不允许改变对象或命题。","assessmentChange":"转用W-M01的固定命题证明搜索，毋须重复研究立项。","acceptanceChange":"可信声明比对、内核和传递公理强制；不冒充新发现。"},{"title":"精确反例","inputDifference":"目标是反驳全称命题或提出无限反例族。","assessmentChange":"有限命题可直接用精确程序寻找证书；渐近目标需要结构论证。","acceptanceChange":"检查全部前提和精确违例；有限计算不能反驳渐近性质。"}],"evidenceStatus":"active"},{"id":"W-F04","domainId":"finance","title":"核保申请包：按固定准则交付有出处的承保建议","goal":"把一次商业保险申请及其资料，变成可追溯的适配判断、候选条件与关键缺项；保留准则未允许的例外。","trigger":"核保人员收到企业申请、经营信息和损失记录，需要按本机构版本化准则作出本次候选承保判断。","scope":"覆盖资料一致性、准则适配、限额/免赔候选和例外说明；保费模型校准、事实调查、组合风险及正式承保授权分别处理。","inputs":[{"name":"申请与附件快照","detail":"经营活动、地点、规模、既有损失等原始文件和来源位置；缺失项保留，不凭经验补值。"},{"name":"本次适用准则","detail":"产品、业务准入矩阵、限额与免赔规则、有效期间、地域和准则版本；只使用实际获准条款。"},{"name":"决策范围","detail":"可自动得出的条件与必须返回例外的规则，允许读取的数据库和工具；不把全行业通用知识代替机构准则。"}],"deliverables":["申请事实—原始页码/记录—适用准则的对应表","有理由的准入建议及限额/免赔候选；明确未决条件","仅针对会改变判断的缺项与矛盾清单，附验证路径"],"acceptance":[{"criterion":"建议能由指定准则推出","method":"独立按冻结的准入矩阵与阈值重算，逐项检查条款版本、经营分类及例外；用阈值两侧案例验证。","failure":"引用机构并未提供的产品，或把旧版本准则当现行准则。"},{"criterion":"事实与缺项被忠实保留","method":"从重要结论反查申请页/数据库记录；对空白、模糊扫描和跨文件冲突运行缺失值及矛盾控制。","failure":"给未记载的保费或被遮挡名称填入合理数字，虽有页码却没有原文依据。"},{"criterion":"输出范围可接收且不越过例外","method":"用具有明确参考判断的完整申请、缺项申请和规则冲突申请校准；分开记录建议正确、必须补问和无法判断。","failure":"为了给结论跳过关键经营风险信息，或把无足够信息一律当拒保。"}],"assessment":{"recommendation":"准则和申请资料可读取时，可把适配分析、候选条件与缺项识别整包交给Agent；按准则与出处接收建议，例外和未决事实明确返回。","conditions":["准则、产品与期间已固定且无未处理冲突","工具可读取原文和相应版本的结构化数据","候选条件与正式承保授权边界清楚"],"evidenceBasis":"UNDERWRITE已研究同类准则与工具驱动判断；IAB补充真实文件中的缺失和条款读取风险。前者是模拟企业环境，后者是文档能力，不能相加为真实核保通过率。","localStatus":"本单元无本站模型试次与业务承保结果；公开论文和厂商文档已核对，尚未重放。","gaps":["机构私有准则、例外处理及真实损失反馈尚未覆盖","缺少当前目标系统在本机构申请分布下的重复试验"],"mode":"staged-delivery","rationale":"专门评测支持AI可组织准则读取与多轮工具推理；同时暴露凭预训练知识捏造产品、重复运行不一致。因而可委派完整候选交付，但对机构特有例外逐项验收。","frontier":"研究环境已把准入、资格、限额/免赔等判断接入工具交互；现有公开证据没有证明完整承保组合的长期盈利或无干预可靠性。","nextValidation":"选择获准准则下有专家参考判断的申请包，含规则版本冲突、阈值两侧、缺项与陌生经营类别；固定模型/工具并保存全部尝试，用规则重算与专业审阅共同判定。","futureSignals":["在未见机构准则和复杂例外上保持可重复的正确建议","缺失/冲突材料能稳定拒答或只补问关键事实","外部验证显示总复核与返工成本降低，而非只增加建议产量","收缩：准则更新后反复沿用旧版、只在已知模板通过，或复核/补问成本上升时，缩至事实—准则对应包并修复知识更新。"]},"workflow":{"before":"核保人员在申请附件、分类表、系统与准则之间逐项搬运，手动整理建议。","now":"以承保建议为交付边界，Agent自主读取准则、比对材料、查询分类并修订候选；确定性限额计算交给规则程序，缺项只在影响判断时请求补充。","retained":["机构准则与适用事实必须对应","例外取舍、风险偏好和正式承保责任"],"added":["维护可读取的准则版本及变化影响","对缺失、冲突与规则外推进行针对性验收"]},"humanResponsibility":"确认本次风险偏好、例外授权和接收标准，决定正式承保；这些是当前组织责任，不预设技术永久只能辅助。","failureHandling":"保留申请和准则原件；输出哪些事实/规则尚不能对应以及需要的最小补充；工具故障、信息缺失和真实不符合准则分别返回。","example":{"input":"一家有两处经营地点的企业申请商业财产险：申请表与损失记录的经营分类不一致，准入表有新旧两版。","output":"固定有效版本，逐地点解释准入及候选免赔；列出分类冲突对应原文及其会改变的条件，未解决前保留条件建议。","failure":"采用较宽松的旧版准则，把另一地点的损失记录当当前地点，仍给出无条件可承保。"},"sourceIds":["WU-UNDERWRITE","WU-IAB"],"evidenceLinks":[{"sourceId":"WU-UNDERWRITE","supports":"模拟机构准则与工具环境中的核保判断已经接受专门评估。","notCovered":"不覆盖本机构准则、实地事实和实际承保组合结果。"},{"sourceId":"WU-IAB","supports":"真实保险文件中，缺失值、跨文档及长保单处理会影响输入可靠性。","notCovered":"厂商文档评分不能证明最终承保决策正确。"}],"resourceRefs":[],"career":{"direction":"核保方向的作品可以是一份版本明确、出处可追溯的申请判断包。","evidence":"同时展示规则适配、关键缺项和反例；不以保险问答数量表示核保能力。","nextStep":"在目标险种/机构语境核对准则与例外处理。"},"team":{"bottleneck":"资料与准则的对应及例外是候选交付的主要接收瓶颈。","allocation":"让Agent交付分析包，规则程序重算，核保接收者集中处理重要例外。","nextStep":"先量复核、补问与返工，再判断是否调整工作组合。"},"reviewedAt":"2026-10-20","claims":[{"id":"F04-result","statement":"已有专门评测在企业准则、数据库工具和模拟用户中测量核保Agent。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-UNDERWRITE","locator":"The Underwrite Environment；Experimental Details","relation":"direct","supports":"与机构准则结合的核保推理被作为完整交互任务研究。","limit":"合成企业环境、旧模型配置；不等于实际承保结果。"}],"progressIds":["P-F-UNDERWRITE"],"configuration":"论文13模型、300任务、最多50轮；ReAct/LangGraph与模拟用户；结果按文中配置解释，不录入有矛盾的榜单点值。","limitation":"本站未重放；任务环境与真实机构差异需要单独验证。","evidenceStatus":"active"},{"id":"F04-adoption","statement":"可委派固定准则下的候选建议包，以规则重算和事实出处验收。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-UNDERWRITE","locator":"Results and Analysis；企业专有知识失败","relation":"partial","supports":"支持工具与规则驱动分析，并暴露需核查的失败模式。","limit":"从模拟环境迁移到目标机构仍有输入与规则差异。"},{"sourceId":"WU-IAB","locator":"§4、§6","relation":"adjacent","supports":"文档处理与缺失值案例支持设置来源核对和拒判条件。","limit":"没有验证本任务的核保结论。"}],"progressIds":[],"configuration":null,"limitation":"这是从公开证据推导的采用建议，尚无本机构成功率。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"通用Agent与规则/SQL工具可组成执行路径；UNDERWRITE论文环境和机构生产系统的可得性不同，本站未验证其可下载运行入口。","tools":[{"name":"UNDERWRITE环境说明","url":"https://arxiv.org/html/2602.00456v1","version":"v1，2026-01-31","access":"公开论文；不标为已取得完整运行系统"},{"name":"Cooper文档系统","url":"https://www.askcooper.ai/labs/insurance-agent-benchmark","version":"IAB读取于2026-10-20","access":"厂商服务与自报评估，接入及费用另核对"}]},"variants":[{"title":"完整且规则内的标准申请","inputDifference":"关键字段与原文一致、准则无冲突。","assessmentChange":"可以在同一交付内完成适配与候选条件，不需人为拆开每次查询。","acceptanceChange":"核对准则重算与出处，抽查边界条件。"},{"title":"冲突、陌生经营或例外申请","inputDifference":"分类或资料矛盾，或准则没有覆盖。","assessmentChange":"交付缩为条件建议和最小澄清问题；不能填造机构规则。","acceptanceChange":"检查冲突是否完整暴露，例外是否由获准标准处理。"}],"evidenceStatus":"active"},{"id":"W-F05","domainId":"finance","title":"理赔证据包：交付责任判断候选与可重算理算","goal":"把固定保单与一次损失的证据，整理为有出处的责任候选、可重算金额和未决调查问题。","trigger":"理赔人员收到保单、批单、事故材料和费用清单，需要明确本次损失适用条款与候选理算。","scope":"覆盖已提供事实的条款适用和金额计算；不由文件推断证据必真，不把欺诈风险线索当欺诈事实，调查与支付执行分别授权。","inputs":[{"name":"合同版本包","detail":"保单、批单、有效期间及适用顺序，免赔、分项/累计限额和已赔付记录。"},{"name":"损失证据快照","detail":"事故时间、费用原件与调查材料；标示事实冲突、重复票据和未确认事项。"},{"name":"理算规则","detail":"本案认可的计算范围、币种/税费、分摊、舍入和缺项处理；不给AI自由改合同规则。"}],"deliverables":["事实—证据位置—控制条款对照，含批单覆盖关系","可执行理算明细、关键中间值与独立重算方式","责任候选、反面证据、未决调查及可导致改判的条件"],"acceptance":[{"criterion":"适用条款与时间线相符","method":"从候选责任逐条反查保单/批单和事故事实；检查生效、失效、除外与恢复条款的适用顺序。","failure":"只读主险条款忽略批单，或将事故后生效条款套到事故时点。"},{"criterion":"理算能够独立重算","method":"用独立脚本重算免赔、分项与累计限额、已赔金额及费用重复；测试限额与免赔的边界输入。","failure":"总额看似合理，但先后扣减顺序错误或重复票据被计入。"},{"criterion":"不把材料空缺或风险线索写成事实","method":"使用相互矛盾、缺失和无法判定的证据控制，检查引用确有该内容，并列出需要调查的最小问题。","failure":"凭照片相似认定欺诈，或为获得完整结论补造事故时间与费用。"}],"assessment":{"recommendation":"可将条款查找、证据对照与规则理算组织为一个候选交付；金额用程序重算，责任依据逐项核对。缺少关键事实时应交付未决问题和条件结论。","conditions":["完整合同版本与事实快照可读取","理算公式与重要边界能够固定","事实真实性、争议与支付授权另有处理机制"],"evidenceBasis":"官方理赔示例提供可检查的集成路径；IAB有文档读取、跨文件及缺失值评估。前者是演示、后者是子能力，当前整案委派建议属于有范围的迁移推断。 寿险身故演示没有实现本例财产险的费用、免赔与累计限额模型，迁移差异必须另外验证。","localStatus":"本站未运行理赔Agent或真实案件；无本任务成功率、成本与赔付效果。","gaps":["从文档子任务到真实责任认定的迁移尚待直接试验","预置保单和纯文本演示不覆盖真实证据、复杂批单或长期争议"],"mode":"staged-delivery","rationale":"规则理算可机械核验，文档与参考实现支持建立候选交付；涉及证据真伪、合同冲突与事实调查的结论缺少直接可靠性证据，采用范围因此限定为材料内可追溯判断。 此处AWS来源仅是寿险身故的资料—保单—候选结论集成示例；本例财产损失中的发票去重、免赔顺序、累计限额和已赔余额是编辑设计，尚无该机制的直接实现或整案能力证据。","frontier":"公开实现已把读取资料、条款核对、候选结论连接为流程；文件处理研究仍观察到不存在字段被编造，不能以接入成功代表整案正确。","nextValidation":"先用独立规则计算校准财产险发票去重、免赔顺序、分项/累计限额及已赔余额；再选同险种、固定合同和专家参考案件，含批单优先与缺失事实，逐条检查责任、金额、应拒判和整案交付，保留全部尝试。","futureSignals":["在复杂批单与冲突证据案件上有独立整案评估","文档缺失与条款定位失败能可靠暴露而非伪造出处","真实接收、争议和返工记录支持扩大自动处理范围","收缩：复杂批单或重复费用下持续算错，先缩至事实—条款定位包；规则计算与合同结构修复后再扩大。"]},"workflow":{"before":"理赔人员分别读合同、核票、查条款、手工计算和撰写结论，材料交接容易丢失上下文。","now":"围绕同一案件候选结论，Agent自行关联证据与条款、调用确定性理算并反查来源；冲突触发补证或条件结论，不固定为六个Agent或某种调用顺序。","retained":["事实真实性与合同适用需要各自依据","正式责任处理、争议与支付责任"],"added":["合同版本与条款依赖维护","引用真实性、拒判和理算边界的自动控制"]},"humanResponsibility":"确定案件接收标准及调查/支付权限，处理有争议事实和责任取舍；不把当前机构安排当作技术永久上限。","failureHandling":"保留原件与计算快照；按条款无法定位、资料冲突、规则计算错误、工具故障分别返回。未知事实可形成附依赖的条件金额或区间，不能冒充已确认的无条件金额与责任。","example":{"input":"同一财产损失有主保单与后续批单，两张费用发票内容重叠，另有一笔此前已赔金额。","output":"逐项列批单影响、去重依据、免赔和剩余额度计算，以及责任判断依赖的事故时间；附可重算脚本和未决证据。","failure":"总额扣一次免赔后即报可支付，忽略累计限额与已赔金额，并把不清晰票据当已确认支出。"},"sourceIds":["WU-IAB","WU-CLAIMS-DEMO"],"evidenceLinks":[{"sourceId":"WU-IAB","supports":"文档理解评估提供条款、跨材料、缺失与引用错误的直接子任务证据。","notCovered":"没有验证责任认定、真实调查或整案处理成功率。"},{"sourceId":"WU-CLAIMS-DEMO","supports":"官方参考实现展示将资料、保单和候选结论连接的技术路径。","notCovered":"预置场景和纯文本输入不证明生产可用或真实证据核查。"}],"resourceRefs":[],"career":{"direction":"理赔作品应展示事实—条款—金额—未决事项的可追溯关系。","evidence":"完整材料内判断、独立重算和合理拒判比流畅结论更有证明力。","nextStep":"按目标险种核对合同结构、调查和争议处理。"},"team":{"bottleneck":"复杂事实、合同依赖与实际争议决定接收成本。","allocation":"Agent整合候选包，程序核算，接收者按重要未决事项安排调查。","nextStep":"先在固定历史案件影子运行，记录错误与总复核时间。"},"reviewedAt":"2026-10-20","claims":[{"id":"F05-method","statement":"寿险身故理赔已有官方参考实现，但其演示依赖预置保单和纯文本资料。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-CLAIMS-DEMO","locator":"README：Demo Limitations；What It Does","relation":"method","supports":"可检查一种资料到候选结论的集成路径及其限制。","limit":"未运行；不是实际理赔可靠性测量。"}],"progressIds":[],"configuration":null,"limitation":"可获得实现入口不等于本案件能成功，也不能代替事实调查。","evidenceStatus":"active"},{"id":"F05-adoption","statement":"可形成材料内的责任与理算候选包，按条款和独立计算接收，未决事实单列。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-IAB","locator":"§4文档任务；§6失败实例","relation":"partial","supports":"支持设置条款定位、跨文件和缺失值检查。","limit":"文档任务到责任认定存在迁移缺口。"},{"sourceId":"WU-CLAIMS-DEMO","locator":"README API Reference与Demo Limitations","relation":"method","supports":"说明可连接的系统接口和演示边界。","limit":"没有提供真实业务整案通过率。"}],"progressIds":[],"configuration":null,"limitation":"明确为采用设计，后续需同险种案件级验证；不从子任务分数推出整案概率。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"AWS样例源码公开，需自行配置云环境、模型与业务系统；本轮只核对README和固定提交，未部署或执行。 实际可下载的是寿险身故演示；财产险合同/费用/免赔/累计限额模型及相应接口需要另行接入，不能直接复用为本例实现。","tools":[{"name":"AWS Claims Processing Demo","url":"https://github.com/aws-samples/sample-agentic-insurance-claims-processing-demo/tree/0bab9ac348e830b016c3faaa2a31ab6618dbf385","version":"0bab9ac348e830b016c3faaa2a31ab6618dbf385","access":"开源示例；需AWS账号、模型访问与有成本的基础设施"}]},"variants":[{"title":"合同与事实完整的规则内案件","inputDifference":"保单/批单和费用可读取，没有关键事实争议。","assessmentChange":"候选责任、金额与出处可作为同一交付。","acceptanceChange":"检查条款优先、限额/免赔和独立金额重算。"},{"title":"事实冲突或需要现场调查","inputDifference":"事故原因、时间或票据真实性不能由现有材料确定。","assessmentChange":"交付范围转为条件结论和调查问题，不能由语言推理补齐事实。","acceptanceChange":"检查冲突和依赖是否明确，未确认项不得被写成确定结论。"}],"evidenceStatus":"active"},{"id":"W-C04","domainId":"computing","title":"浏览器功能交付：操作、渲染与可访问性一起验收","goal":"交付可在目标浏览器完成真实用户任务的界面变更，附持久状态、视觉差异和可访问性证据。","trigger":"给定页面、用户意图和设计参照，定义提交后/刷新后应保留的状态以及加载、无数据、失败和无权限行为。","scope":"范围为已给设计/行为目标的页面或组件，不把视觉截图等同交互成功；原创品牌设计另定验收。","inputs":[{"name":"目标状态","detail":"给定页面、用户意图和设计参照，定义提交后/刷新后应保留的状态以及加载、无数据、失败和无权限行为。"},{"name":"环境与差异","detail":"固定仓库、API数据种子、浏览器/视口、语言、网络条件与访问身份；性能预算按真实业务冻结。"},{"name":"接收基线","detail":"保留原页面trace和响应数据，允许Agent选DOM/截图/代码工具探索，关键结果从服务状态反查。"}],"deliverables":["可应用界面补丁及组件/状态接口变化","真实浏览器操作trace、状态对照和视口截图","键盘/焦点/语义检查、性能前后测量及未覆盖辅助技术范围"],"acceptance":[{"criterion":"任务确实完成","method":"执行输入—提交—刷新—重进，核对服务端或本地持久状态与错误恢复。","failure":"只截图漂亮页面，点击按钮无效果。"},{"criterion":"不同使用条件可接收","method":"对目标视口、加载/失败态、键盘导航及焦点恢复逐项检查；axe只负责可自动发现的规则。","failure":"把零axe告警写成所有辅助技术用户都能完成任务。"},{"criterion":"改善可重复","method":"在同一数据/网络/设备条件比较性能与布局，保留失败trace。","failure":"只更改阈值或关闭慢路径来声称性能改善。"}],"assessment":{"recommendation":"可整包委派已定义交互的实现与浏览器迭代；以真实状态变化接收，键盘和辅助技术另留明确检查。","conditions":["固定仓库、API数据种子、浏览器/视口、语言、网络条件与访问身份；性能预算按真实业务冻结。","保留原页面trace和响应数据，允许Agent选DOM/截图/代码工具探索，关键结果从服务状态反查。"],"evidenceBasis":"视觉JavaScript软件缺陷已有专属评估；原研究观察到通用修复系统跨到视觉与另一语言时明显退化；可用Playwright与axe检查常见自动化可访问性缺陷，但需要交互和辅助技术条件验证补足覆盖。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["目标页面的持久状态、跨浏览器、键盘/读屏与性能矩阵尚未实际重放。"],"mode":"bounded-delivery","rationale":"视觉任务已有独立研究对象，通用Python修复成绩不能迁移成界面完成率；浏览器反馈能把画面、动作和状态放进同一修正循环。","frontier":"SWE-bench Multimodal给出视觉JavaScript任务及历史实验，尚不能证明目标产品的跨浏览器与可访问性。","nextValidation":"选择一个有异步提交和失败恢复的真实页面，先冻结端点/状态控制；让Agent交付并重放键盘、窄屏和慢网路径。","futureSignals":["新页面和浏览器下真实任务、持久状态及键盘流程持续通过，可考虑扩大相近输入范围。","视觉相似但交互/焦点/错误恢复退化，需收窄支持矩阵，应收窄或暂停相应采用。"]},"workflow":{"before":"设计、开发、手工点验和缺陷来回交接。","now":"Agent可自主修改页面、启动浏览器、观察trace并迭代，程序核对数据状态；使用者按目标接收体验。","retained":["确定用户任务、视觉取舍和需要支持的使用条件；辅助技术体验缺口应由相应使用场景验证。"],"added":["在同一数据/网络/设备条件比较性能与布局，保留失败trace。"]},"humanResponsibility":"确定用户任务、视觉取舍和需要支持的使用条件；辅助技术体验缺口应由相应使用场景验证。","failureHandling":"缺API、浏览器或登录身份时交付可重现阻塞与已完成范围；不能用静态图宣告功能交付。","example":{"input":"筛选订单后导出，移动宽度下错误提示不挡住重试按钮。","output":"补丁、真实请求/文件结果、键盘焦点与三类状态trace。","failure":"按钮只是下载伪CSV，刷新后过滤条件丢失。"},"sourceIds":["WU-C-VISUAL","WU-C-PW","WU-C-PWMCP"],"evidenceLinks":[{"sourceId":"WU-C-VISUAL","supports":"证明视觉修复需要独立任务分布。","notCovered":"2024历史系统，不能当成2026当前前端能力上限。"},{"sourceId":"WU-C-PW","supports":"给出自动规则和边界。","notCovered":"工具检查不是Agent表现或全面WCAG认证。"},{"sourceId":"WU-C-PWMCP","supports":"提供操作与观察的可用接口。","notCovered":"迁移到具体产品仍需独立数据与体验检查。"}],"resourceRefs":[],"career":{"direction":"前端开发、质量工程","evidence":"可应用界面补丁及组件/状态接口变化","nextStep":"选择一个有异步提交和失败恢复的真实页面，先冻结端点/状态控制；让Agent交付并重放键盘、窄屏和慢网路径。"},"team":{"bottleneck":"视觉任务已有独立研究对象，通用Python修复成绩不能迁移成界面完成率；浏览器反馈能把画面、动作和状态放进同一修正循环。","allocation":"Agent可自主修改页面、启动浏览器、观察trace并迭代，程序核对数据状态；使用者按目标接收体验。","nextStep":"用一项真实授权交付核对可接收产物、失败、返工和接收耗时；据此调整工具/环境与团队分工。"},"claims":[{"id":"C04-VISUAL","statement":"视觉JavaScript软件缺陷已有专属评估；原研究观察到通用修复系统跨到视觉与另一语言时明显退化。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-C-VISUAL","locator":"Abstract：任务定义与视觉/跨语言实验摘要","relation":"partial","supports":"证明视觉修复需要独立任务分布。","limit":"2024历史系统，不能当成2026当前前端能力上限。"}],"configuration":"SWE-bench M：617任务、17个JS库；论文SWE-agent等原配置，非本站运行。","progressIds":["P-C-VISUAL-REPAIR"],"limitation":"2024历史系统，不能当成2026当前前端能力上限。","evidenceStatus":"active"},{"id":"C04-ACCESS","statement":"可用Playwright与axe检查常见自动化可访问性缺陷，但需要交互和辅助技术条件验证补足覆盖。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-PW","locator":"Introduction/Example accessibility tests","relation":"method","supports":"给出自动规则和边界。","limit":"工具检查不是Agent表现或全面WCAG认证。"}],"configuration":null,"progressIds":[],"limitation":"工具检查不是Agent表现或全面WCAG认证。","evidenceStatus":"active"},{"id":"C04-DELIVER","statement":"行为契约和目标浏览器齐备时，AI可把实现、操作反馈和修复合成一个交付循环；新状态由独立控制验收。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-PWMCP","locator":"README：browser tools/configuration","relation":"method","supports":"提供操作与观察的可用接口。","limit":"迁移到具体产品仍需独立数据与体验检查。"}],"configuration":null,"progressIds":[],"limitation":"迁移到具体产品仍需独立数据与体验检查。","evidenceStatus":"active"},{"id":"C04-PERFORMANCE","statement":"可在固定浏览器/设备/数据/网络下委派性能诊断与改动，按可重复测量和真实功能回归接收。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-PWMCP","locator":"browser configuration/tools","relation":"method","supports":"公开操作/观察接口支持工程测量组织。","limit":"视觉JS修复研究不评性能；本页面无改善试验。"}],"configuration":null,"progressIds":[],"limitation":"视觉JS修复研究不评性能；本页面无改善试验。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"浏览器工具公开；模型、应用和浏览器版本需执行时固定，本站未安装。","tools":[{"name":"Playwright MCP","url":"https://github.com/microsoft/playwright-mcp","version":"2026-10-20公开仓库HEAD ab6d67cbf8b619338f096980806fdbe29432cc4d；实际运行另固定依赖/镜像/模型","access":"开源；Node与浏览器环境，目标应用账户另备"}]},"variants":[{"title":"响应式与视觉还原","inputDifference":"既有设计有桌面/窄屏差异。","assessmentChange":"可整包交付，先冻结内容和交互，避免用藏内容制造相似。","acceptanceChange":"检查断点、长文本、溢出与行为；截图须与真实DOM状态对应。"},{"title":"可访问性与输入设备","inputDifference":"使用键盘、读屏、缩放或不同输入法。","assessmentChange":"自动规则可批量修复；复杂读屏流程需单独核验。","acceptanceChange":"逐步检查名称/角色、顺序、焦点返回和错误朗读。"},{"title":"性能与兼容迁移","inputDifference":"改变构建、组件、SSR/客户端状态或大列表。","assessmentChange":"按浏览器矩阵分阶段接收。","acceptanceChange":"同数据负载记录性能、hydration和旧组件/接口回归。"},{"title":"业务仪表盘交互","inputDifference":"目标BI/浏览器环境、业务口径/结果内核、角色与发布入口固定。","assessmentChange":"组合A01数据内核和真实界面实现接收；查询Agent案例不是仪表盘完成成绩。","acceptanceChange":"交组件/BI文件和运行入口，独立操作筛选/联动/下钻并对账分母/结果；验证角色权限、刷新/缓存/空态/错误态和重新进入。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C05","domainId":"computing","title":"移动应用交付：离线、生命周期与设备状态","goal":"交付在指定设备和系统版本上可运行的移动功能或修复，并证明进后台、权限变化与升级后的关键状态。","trigger":"提供现有app、仓库、签名构建方式、目标用户动作及应保存/同步的数据。","scope":"原生/跨平台均按目标OS接收；Android操作基准仅是状态验收参照，不能提供iOS代码或真机传感器成绩。","inputs":[{"name":"设备任务","detail":"提供现有app、仓库、签名构建方式、目标用户动作及应保存/同步的数据。"},{"name":"平台条件","detail":"冻结OS/设备、原生或跨平台框架、账号、网络、权限初态与升级前本地库快照。"},{"name":"状态约束","detail":"明确断网、重连、冲突、后台/进程死亡、通知/外设和升级失败时的可接受状态；敏感操作限定授权。"}],"deliverables":["可安装构建、补丁与版本/依赖清单","设备操作记录、生命周期/离线同步/权限检查结果","旧版升级数据核对、崩溃/内存迹象与回退或阻塞说明"],"acceptance":[{"criterion":"功能在设备上成立","method":"安装并执行核心任务，重启app后从存储/服务确认结果。","failure":"模拟器截图通过就声称所有真机可用。"},{"criterion":"环境变化不丢状态","method":"切断网络、旋转、进后台/销毁活动、撤销权限，再重进核对冲突策略与持久数据。","failure":"重连覆盖本地编辑或后台返回产生重复提交。"},{"criterion":"升级可恢复","method":"从旧版本真实数据升级并验证登录、schema和失败路径；设备特性以实物或对应硬件证据接收。","failure":"全新安装通过，却升级丢数据或通知不达。"}],"assessment":{"recommendation":"可委派实现到设备状态核验；先接收可安装构建，再以生命周期、离线和升级矩阵接收完整交付。","conditions":["冻结OS/设备、原生或跨平台框架、账号、网络、权限初态与升级前本地库快照。","明确断网、重连、冲突、后台/进程死亡、通知/外设和升级失败时的可接受状态；敏感操作限定授权。"],"evidenceBasis":"AndroidWorld在真实Android模拟器上以可变参数和持久状态检查任务，说明手机操作结果可以独立于截图评价；ActivityScenario可控制活动生命周期，便于验证后台、销毁与重建引起的状态错误。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["目标OS/真机、离线冲突、后台回收及旧数据升级尚无完整设备矩阵，Android操作任务不能替代。"],"mode":"staged-delivery","rationale":"AndroidWorld说明移动状态变化可设定独立奖励；但操作app与开发app不同。借助平台测试接口可形成可验收闭环，传感器/后台和平台差异决定接收边界。","frontier":"AndroidWorld公开了API33模拟器上可参数化的既有app操作与持久状态检查；原论文做了任务变体稳健性分析。Android官方ActivityScenario可操纵生命周期供检查；二者不构成新app编程、iOS或真机耗电/外设的模型成绩。","nextValidation":"给一个可脱敏的离线编辑app，冻结冲突策略；在真实目标设备执行编辑—断网—后台杀进程—重连—升级并检查两端记录。","futureSignals":["新OS/旧数据与真机矩阵下同步、权限、后台和升级均可复验，可考虑扩大相近输入范围。","模拟器与真机分歧、生命周期丢状态或平台限制使路径失效，应收窄或暂停相应采用。"]},"workflow":{"before":"开发后由测试人员在几台设备手工试用，再补平台问题。","now":"Agent可构建、控制模拟器/设备、查日志并反复修复；实际状态检查决定何时交付，不固定工具调用顺序。","retained":["确定冲突、数据保留、分发与设备支持政策；平台签名/商店与高影响操作须有实际权限。"],"added":["从旧版本真实数据升级并验证登录、schema和失败路径；设备特性以实物或对应硬件证据接收。"]},"humanResponsibility":"确定冲突、数据保留、分发与设备支持政策；平台签名/商店与高影响操作须有实际权限。","failureHandling":"没有真机/账号/签名时列出可构建部分与阻塞，不把无法观察的生命周期结果写成通过。","example":{"input":"离线修改地址，进程被系统终止后重新联网，服务端同期也有更新。","output":"安装包、两端变更记录和明确冲突结果；保留升级前后数据。","failure":"仅重跑UI测试，却丢失用户离线编辑。"},"sourceIds":["WU-C-ANDROID-REPO","WU-C-ANDROID-LIFE","WU-C-ANDROID"],"evidenceLinks":[{"sourceId":"WU-C-ANDROID-REPO","supports":"提供状态化任务入口与API33参考环境。","notCovered":"是既有app操作，不是移动代码交付评测。"},{"sourceId":"WU-C-ANDROID-LIFE","supports":"平台测试支持生命周期验收。","notCovered":"Android方法不覆盖iOS及全部真机硬件。"},{"sourceId":"WU-C-ANDROID","supports":"动态状态任务提供迁移参照。","notCovered":"没有本app和当前模型试验，采用建议须由设备矩阵校准。"}],"resourceRefs":[],"career":{"direction":"移动应用开发","evidence":"可安装构建、补丁与版本/依赖清单","nextStep":"给一个可脱敏的离线编辑app，冻结冲突策略；在真实目标设备执行编辑—断网—后台杀进程—重连—升级并检查两端记录。"},"team":{"bottleneck":"AndroidWorld说明移动状态变化可设定独立奖励；但操作app与开发app不同。借助平台测试接口可形成可验收闭环，传感器/后台和平台差异决定接收边界。","allocation":"Agent可构建、控制模拟器/设备、查日志并反复修复；实际状态检查决定何时交付，不固定工具调用顺序。","nextStep":"用一项真实授权交付核对可接收产物、失败、返工和接收耗时；据此调整工具/环境与团队分工。"},"claims":[{"id":"C05-STATE","statement":"AndroidWorld在真实Android模拟器上以可变参数和持久状态检查任务，说明手机操作结果可以独立于截图评价。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-C-ANDROID-REPO","locator":"README：environment/tasks；Installation","relation":"partial","supports":"提供状态化任务入口与API33参考环境。","limit":"是既有app操作，不是移动代码交付评测。"}],"configuration":null,"progressIds":["P-C-MOBILE-STATE"],"limitation":"是既有app操作，不是移动代码交付评测。","evidenceStatus":"active"},{"id":"C05-LIFECYCLE","statement":"ActivityScenario可控制活动生命周期，便于验证后台、销毁与重建引起的状态错误。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-ANDROID-LIFE","locator":"ActivityScenario lifecycle testing","relation":"method","supports":"平台测试支持生命周期验收。","limit":"Android方法不覆盖iOS及全部真机硬件。"}],"configuration":null,"progressIds":[],"limitation":"Android方法不覆盖iOS及全部真机硬件。","evidenceStatus":"active"},{"id":"C05-DELIVER","statement":"在目标设备/权限可用并冻结同步规则时，可将移动代码实现与状态反馈委派为一包；未经设备核验不接收平台特性交付。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-ANDROID","locator":"Abstract；dynamic tasks and robustness","relation":"partial","supports":"动态状态任务提供迁移参照。","limit":"没有本app和当前模型试验，采用建议须由设备矩阵校准。"}],"configuration":null,"progressIds":[],"limitation":"没有本app和当前模型试验，采用建议须由设备矩阵校准。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"Android公开环境可用；本任务iOS与真机能力需实际工具链，本站没有运行。","tools":[{"name":"AndroidWorld","url":"https://github.com/google-research/android_world","version":"2026-10-20公开仓库HEAD f8e78c0ab0bb8b5daa681afd0296ed0193a98f1f；实际运行另固定依赖/镜像/模型","access":"开源模拟器环境；模型API自备，默认legacy模型须另确认"},{"name":"Android ActivityScenario","url":"https://developer.android.com/guide/components/activities/testing","version":"按目标AndroidX/OS固定","access":"Android工具链；真机/模拟器及应用测试权限"}]},"variants":[{"title":"离线同步与冲突","inputDifference":"本地和服务端可并行改同一对象。","assessmentChange":"先冻结冲突语义再整包实现，不能默认最后写入覆盖。","acceptanceChange":"断网重试、重复消息、时钟变化、进程终止后对账。"},{"title":"权限与生命周期","inputDifference":"通知、相机、传感器、后台运行或进程被回收。","assessmentChange":"逐设备/OS分阶段验收；模拟器不足时保留硬件验证缺口。","acceptanceChange":"拒绝/撤销权限、后台限制、重建、外设失联及内存迹象。"},{"title":"升级与分发","inputDifference":"旧版本安装、本地schema和签名存在。","assessmentChange":"构建可委派，分发权限与真实升级单独确认。","acceptanceChange":"旧数据升级、签名/版本、迁移中断和恢复；不能只新装。"},{"title":"iOS或跨平台","inputDifference":"代码运行在另一OS或桥接插件。","assessmentChange":"只迁移任务方法，不迁移Android能力成绩。","acceptanceChange":"使用该平台生命周期、权限、真机与崩溃记录重新接收。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C06","domainId":"computing","title":"机器学习实验包：数据、训练与可解释的保留集表现","goal":"在预算和数据使用边界内交付可重现模型或实验结论，包含独立保留集、切片误差和失败记录。","trigger":"定义预测对象、可用时间点、错误代价与接收指标；研究须给可反驳假设与复现目标。","scope":"覆盖数据/标签诊断、训练或适配、基线与消融；上线服务与真实业务收益另以W-C15/W-A03接收。","inputs":[{"name":"研究或预测目标","detail":"定义预测对象、可用时间点、错误代价与接收指标；研究须给可反驳假设与复现目标。"},{"name":"数据与预算","detail":"固定数据版本、标签来源、许可、训练/验证/保留划分、硬件、时间/费用和可用预训练权重。"},{"name":"独立评分","detail":"评分器/保留集权限与开发环境分开；明确不可检索答案、训练污染和多次试验选择规则。"}],"deliverables":["数据/标签/划分清单和泄漏检查","可重跑训练或微调代码、配置、权重与基线比较","保留集/误差切片、试验账本、资源和未达到目标的解释"],"acceptance":[{"criterion":"评价不被开发污染","method":"按时间/实体/组划分并审查特征可用时点，保留集由独立入口评分。","failure":"同一用户重复进入训练和测试，或使用未来标签。"},{"criterion":"结果可重现且代价明确","method":"从固定环境重新训练/加载权重，核对种子、预处理、依赖、尝试次数与实际资源。","failure":"只交最好的截图，丢掉失败与调参预算。"},{"criterion":"泛化和结论相称","method":"对关键业务群体、漂移与简单基线比较；研究结果含消融与可反证范围。","failure":"验证分上升就声称已改善生产效果或提出新理论。"}],"assessment":{"recommendation":"可委派数据检查、训练搜索和模型产物；先校准保留集与泄漏控制，再按冻结指标接收模型，不用竞赛得分替代部署判断。","conditions":["固定数据版本、标签来源、许可、训练/验证/保留划分、硬件、时间/费用和可用预训练权重。","评分器/保留集权限与开发环境分开；明确不可检索答案、训练污染和多次试验选择规则。"],"evidenceBasis":"MLE-bench的公开研究在Kaggle工程任务上实测了Agent的数据准备、模型训练与实验能力；它支持离线实验交付存在，不能直接代表生产收益；MLflow能记录训练参数、代码版本、指标和输出文件，便于交付完整试验账本。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["目标标签/划分与保留集、资源预算和未来时间窗泛化未独立验证，不能继承竞赛分数。"],"mode":"staged-delivery","rationale":"MLE-bench证明训练与实验可以被Agent组织成可评分交付；其预算、公开数据与榜单条件特定。独立保留集与数据机制限制迁移，不能用多次搜索最好一次作稳定表现。","frontier":"公开训练任务已有实测与可运行资产；当前排行榜暂停新提交，跨方法可比性有限，尚不能证明本组织数据上的泛化。","nextValidation":"取一个未泄漏的小型真实数据问题，锁定时间/实体保留集与预算，先跑简单参考，再让Agent交完整模型和失败账本；服务化另测。","futureSignals":["在新时间窗/实体保留集持续超过简单基线且资源可追踪，可考虑扩大相近输入范围。","标签/未来特征泄漏、切片退化或只保留最好一次结果，应收窄或暂停相应采用。"]},"workflow":{"before":"数据准备、特征、调参、训练与报告由多人交接。","now":"Agent可按评分反馈自主选择模型、代码与实验，确定性训练器直接运行；保留集和预算与搜索过程分离。","retained":["确认目标、数据权利、错误代价和上线门槛；不需要手工逐步挑特征，但须保护独立评分。"],"added":["对关键业务群体、漂移与简单基线比较；研究结果含消融与可反证范围。"]},"humanResponsibility":"确认目标、数据权利、错误代价和上线门槛；不需要手工逐步挑特征，但须保护独立评分。","failureHandling":"数据质量或划分不成立时先交诊断和修复方案；预算内未达标是结果，不生成虚构成绩。","example":{"input":"按时间预测设备故障，维修后信息不能在预测时使用。","output":"固定时间划分、基线/候选权重、关键设备切片与资源记录。","failure":"随机分行让同一设备泄漏进保留集，得高分却无法新设备泛化。"},"sourceIds":["WU-C-MLE","WU-C-MLFLOW","WU-C-MLE-REPO"],"evidenceLinks":[{"sourceId":"WU-C-MLE","supports":"训练交付有直接研究而非只由编程成绩外推。","notCovered":"历史竞赛样本；本页不采用或迁移其奖牌比例。"},{"sourceId":"WU-C-MLFLOW","supports":"提供可追踪实验接口。","notCovered":"记录不证明无泄漏或指标适合业务。"},{"sourceId":"WU-C-MLE-REPO","supports":"可运行训练及评分资产是迁移基础。","notCovered":"官方暂停新提交，任何目标任务成功率仍未知。"}],"resourceRefs":[],"career":{"direction":"机器学习工程、数据科学、计算机研究","evidence":"数据/标签/划分清单和泄漏检查","nextStep":"取一个未泄漏的小型真实数据问题，锁定时间/实体保留集与预算，先跑简单参考，再让Agent交完整模型和失败账本；服务化另测。"},"team":{"bottleneck":"MLE-bench证明训练与实验可以被Agent组织成可评分交付；其预算、公开数据与榜单条件特定。独立保留集与数据机制限制迁移，不能用多次搜索最好一次作稳定表现。","allocation":"Agent可按评分反馈自主选择模型、代码与实验，确定性训练器直接运行；保留集和预算与搜索过程分离。","nextStep":"用一项真实授权交付核对可接收产物、失败、返工和接收耗时；据此调整工具/环境与团队分工。"},"claims":[{"id":"C06-TRAIN","statement":"MLE-bench的公开研究在Kaggle工程任务上实测了Agent的数据准备、模型训练与实验能力；它支持离线实验交付存在，不能直接代表生产收益。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-C-MLE","locator":"Abstract；75 competitions；resource scaling","relation":"partial","supports":"训练交付有直接研究而非只由编程成绩外推。","limit":"历史竞赛样本；本页不采用或迁移其奖牌比例。"}],"configuration":"原论文75个Kaggle竞赛，含o1-preview+AIDE等配置；模型/脚手架和预算按原实验区分；未在本站重放。","progressIds":["P-C-ML-TRAIN"],"limitation":"历史竞赛样本；本页不采用或迁移其奖牌比例。","evidenceStatus":"active"},{"id":"C06-TRACK","statement":"MLflow能记录训练参数、代码版本、指标和输出文件，便于交付完整试验账本。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-MLFLOW","locator":"Tracking Runs；Runs/Models","relation":"method","supports":"提供可追踪实验接口。","limit":"记录不证明无泄漏或指标适合业务。"}],"configuration":null,"progressIds":[],"limitation":"记录不证明无泄漏或指标适合业务。","evidenceStatus":"active"},{"id":"C06-ADOPT","statement":"在独立保留集和预算受控时，可整包委派训练搜索；采用决策以保留集切片、资源和下游条件作阶段接收。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-MLE-REPO","locator":"README：leaderboard update/agents","relation":"partial","supports":"可运行训练及评分资产是迁移基础。","limit":"官方暂停新提交，任何目标任务成功率仍未知。"}],"configuration":null,"progressIds":[],"limitation":"官方暂停新提交，任何目标任务成功率仍未知。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"训练框架和记录工具公开；竞赛/模型/计算资源有条件且本文未消耗新试验。","tools":[{"name":"MLE-bench","url":"https://github.com/openai/mle-bench","version":"2026-10-20公开仓库HEAD 507f92e1138bb6e40dac5c6ee7a6758e6424bf97；实际运行另固定依赖/镜像/模型","access":"公开代码；Kaggle条款、GPU/数据存储与模型API另备"},{"name":"MLflow Tracking","url":"https://mlflow.org/docs/latest/ml/tracking/","version":"Tracking文档；具体包版本在实际试验固定","access":"开源记录工具；存储/访问控制另配置"}]},"variants":[{"title":"标签与评测集","inputDifference":"标签噪声、样本不平衡或实体/时间依赖。","assessmentChange":"先接收数据诊断与划分，再搜索模型。","acceptanceChange":"抽核标签、实体隔离、未来特征和重复；保留所有清洗影响。"},{"title":"微调与基础模型","inputDifference":"调用预训练权重、检索或第三方模型。","assessmentChange":"增加许可、上下文与生成评测，不能沿用分类准确率。","acceptanceChange":"无检索泄漏，任务/安全控制独立，记录外部API版本。"},{"title":"研究复现与消融","inputDifference":"目标是论文结果或新方法比较。","assessmentChange":"按贡献与实验接收；理论结论与新颖性另核。","acceptanceChange":"对应W-C14的代码/结果/消融及复现差异。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C07","domainId":"computing","title":"数据管道交付：接入、转换、补数和可信数据集","goal":"交付可重放且可对账的数据管道/数据集，保持业务键、口径、时效与访问契约。","trigger":"固定源/目标schema、业务键、口径、血缘与希望形成的数据集；明确修正优先级和消费接口。","scope":"工作对象是来源到目标的数据转换；接入、质量修复、增量/历史重放是同一管道交付的条件变化。数据库物理还原与跨版本切换另见W-C16/W-C17。","inputs":[{"name":"数据目标","detail":"固定源/目标schema、业务键、口径、血缘与希望形成的数据集；明确修正优先级和消费接口。"},{"name":"运行条件","detail":"给连接器/调度器/数据库版本、数据快照、权限、重复/迟到/缺失规则和保留政策。"},{"name":"接收控制","detail":"独立准备总量/金额/键约束与下游查询，冻结时效/成本预算；允许测试中断和幂等重放。"}],"deliverables":["可重建接入/转换代码、schema/语义模型及调度配置","源到目标对账、重复迟到/质量诊断和补数结果","血缘/权限/保留说明、刷新时效/成本和中断恢复控制"],"acceptance":[{"criterion":"数据语义与完整性正确","method":"以业务键、金额、约束及下游查询核对源到目标；异常分类而非直接丢弃。","failure":"行数相等但JOIN放大金额或口径混用。"},{"criterion":"补数和重跑幂等","method":"施加重复/迟到/中断，重跑并核对人工修正与增量水位。","failure":"补数覆盖后来修正或重跑产生重复数据。"},{"criterion":"可供消费并可维护","method":"验证数据新鲜度、访问正反矩阵、血缘和schema变化下的下游回归，记录资源。","failure":"数据正确但刷新/权限失效，或消费者取到旧schema。"}],"assessment":{"recommendation":"可委派Agent整包实现与修复管道，先按源到目标控制证明正确，再在代表负载接收时效与治理；固定刷新可直接程序执行。","conditions":["给连接器/调度器/数据库版本、数据快照、权限、重复/迟到/缺失规则和保留政策。","独立准备总量/金额/键约束与下游查询，冻结时效/成本预算；允许测试中断和幂等重放。"],"evidenceBasis":"Spider2-DBT公开了DuckDB/dbt仓库级任务和运行入口，适合检验schema到转换结果的真实状态。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["连接器、迟到/修正语义、水位和真实下游消费尚需对账与中断重放；云账户访问有公告限制。"],"mode":"staged-delivery","rationale":"Spider2-DBT提供数据转换仓库任务入口；真实流式/历史补数属于有明确差异的迁移。把对账、水位和修正语义固定，Agent才能从schema搜索到可接收数据。","frontier":"Spider2-DBT有公开DuckDB/dbt仓库级转换任务和runner，能检验schema到结果；其本地路径不等同全天候流式恢复。维护公告指出Snowflake账户访问受阻，不能把云账号可用性写成已重现结果。","nextValidation":"冻结一包重复/迟到/人工修正数据，校准独立对账，再让Agent交管道并注入中断；从下游读取验证重放与新鲜度。","futureSignals":["新分区/迟到事件下幂等对账、下游口径和新鲜度同时成立，可考虑扩大相近输入范围。","补数覆盖修正、数据重复或schema变化造成静默错误，应收窄或暂停相应采用。"]},"workflow":{"before":"资料、实现/处理与验证分散交接，完成记录易脱离实际结果。","now":"Agent可查询元数据、写转换、重放并对账，异常才重新探索；程序持续刷新冻结口径。","retained":["确定业务键、数据权利、口径与保留政策，负责消费者语义变化。"],"added":["验证数据新鲜度、访问正反矩阵、血缘和schema变化下的下游回归，记录资源。"]},"humanResponsibility":"确定业务键、数据权利、口径与保留政策，负责消费者语义变化。","failureHandling":"对账失败保留异常与可重放中间态；不能删行凑总量或静默改水位。","example":{"input":"迟到的订单事件与人工修正同时进入日报。","output":"可重放管道、键/金额对账、修正优先级及新鲜度。","failure":"重复事件多记收入，补数又覆盖人工修正。"},"sourceIds":["WU-C-SPIDER"],"evidenceLinks":[{"sourceId":"WU-C-SPIDER","supports":"提供数据工程产物任务资产。","notCovered":"不是全天候流式处理成绩；Snowflake公告中断不能转写成本地DBT不可用。"}],"resourceRefs":[],"career":{"direction":"数据管道交付：接入、转换、补数和可信数据集","evidence":"可重建接入/转换代码、schema/语义模型及调度配置","nextStep":"冻结一包重复/迟到/人工修正数据，校准独立对账，再让Agent交管道并注入中断；从下游读取验证重放与新鲜度。"},"team":{"bottleneck":"Spider2-DBT提供数据转换仓库任务入口；真实流式/历史补数属于有明确差异的迁移。把对账、水位和修正语义固定，Agent才能从schema搜索到可接收数据。","allocation":"Agent可查询元数据、写转换、重放并对账，异常才重新探索；程序持续刷新冻结口径。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C07-PIPELINE","statement":"Spider2-DBT公开了DuckDB/dbt仓库级任务和运行入口，适合检验schema到转换结果的真实状态。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-C-SPIDER","locator":"Overview；Spider2-DBT quickstart","relation":"partial","supports":"提供数据工程产物任务资产。","limit":"不是全天候流式处理成绩；Snowflake公告中断不能转写成本地DBT不可用。"}],"configuration":null,"progressIds":[],"limitation":"不是全天候流式处理成绩；Snowflake公告中断不能转写成本地DBT不可用。","evidenceStatus":"active"},{"id":"C07-PIPELINE-ADOPT","statement":"可委派来源到数据集的代码与诊断，独立核对口径、重放和下游状态后接收；不继承数据库恢复实验成绩。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-SPIDER","locator":"Spider2-DBT任务入口","relation":"partial","supports":"仓库级转换任务可检查。","limit":"流式迟到/治理与本组织数据未实测。"}],"configuration":null,"progressIds":[],"limitation":"流式迟到/治理与本组织数据未实测。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"Spider2-DBT有本地DuckDB/dbt入口；Snowflake另有账号暂停公告，不混同可用性。","tools":[{"name":"Spider2-DBT","url":"https://github.com/xlang-ai/Spider2","version":"2026-10-20公开仓库HEAD cafb867313aab4e674652054198f383cf4018943；实际运行另固定依赖/镜像/模型","access":"开源本地入口；模型/资源另备"}]},"variants":[{"title":"增量与历史补数","inputDifference":"迟到、重复、缺组和人工修正交错。","assessmentChange":"先固定业务键/水位/优先级再实现。","acceptanceChange":"多次重放、异常/修正、总量桥接与下游一致。"},{"title":"语义数据集与治理","inputDifference":"多人消费、指标口径、权限/保留变化。","assessmentChange":"数据与语义/访问接口一起接收。","acceptanceChange":"口径版本、血缘、权限正反例、到期保留和消费者回归。"},{"title":"性能与流水线恢复","inputDifference":"刷新延迟、成本或任务中断。","assessmentChange":"按管道结果恢复，不扩成数据库物理恢复。","acceptanceChange":"代表吞吐/成本、水位、幂等断点及下游新鲜度。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C08","domainId":"computing","title":"事故恢复：从信号到持续业务恢复","goal":"在授权环境定位故障并恢复指定用户路径，交付数据/权限不变量与观察窗内恢复证据。","trigger":"给服务拓扑、SLI/SLO、正常流量、现象/预期配置和禁止损害的数据/网络边界；没有事件时给计划变更的目标状态。","scope":"工作对象是正在失败的服务及恢复状态；计划基础设施交付W-C19、指标/告警W-C20、复盘整改W-C21分别接收。","inputs":[{"name":"事故目标","detail":"给受影响用户路径、服务/数据库/网络拓扑、现象、最近变化、SLI/SLO与禁止损害的数据/权限边界。"},{"name":"可观察与可操作环境","detail":"固定namespace/账户、配置、最近发布、指标/日志/trace、依赖状态、费用/时间预算及可执行回退。"},{"name":"独立流量探针","detail":"事先冻结用户路径、写后读/事务校验、告警和资源阈值；操作前保留状态及时间线。"}],"deliverables":["诊断、实际获准恢复动作或待执行边界与配置/代码差异","事件时间线、诊断证据、失败尝试及恢复后流量结果","回退/观察窗、容量/成本差异和复发行动的可追踪记录"],"acceptance":[{"criterion":"用户目标恢复","method":"执行真实关键请求、错误率/延迟和数据一致性探针，在观察窗验证。","failure":"关闭告警、重启进程后只看pod Running。"},{"criterion":"操作范围与副作用可接受","method":"比较权限/网络/配置与数据前后差异，检查其他服务与回退。","failure":"为连通服务开放全部网络或删除队列数据。"},{"criterion":"恢复经得起后续变化","method":"重新施加代表负载或第二事件，检查原修复、容量和复发条件。","failure":"一次请求成功就宣告根因解决，后续发布立即复发。"}],"assessment":{"recommendation":"可委派Agent关联信号、提出/实施授权内修复并持续验证；影响面大或缺回退时先交诊断与演练，不以根因叙述替代恢复。","conditions":["固定namespace/账户、配置、最近发布、指标/日志/trace、依赖状态、费用/时间预算及可执行回退。","事先冻结用户路径、写后读/事务校验、告警和资源阈值；操作前保留状态及时间线。"],"evidenceBasis":"AIOpsLab公开提供故障注入、负载、遥测与按任务结果评估的接口，可用于验证Agent是否改变了真实服务状态。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["目标服务拓扑、权限、事务与第二事件下的恢复尚未演练；原论文/试验床不能替代。"],"mode":"staged-delivery","rationale":"AIOpsLab可将故障、负载、观察和评估接在同一环境，支持按系统状态验收；真实云拓扑、权限及业务不变量决定迁移边界。稳定配置可直接用脚本执行，Agent主要处理异常和不完整信息。","frontier":"AIOpsLab公开连接故障注入、负载、遥测、动作轨迹与检测/定位/分析/缓解评估；这是可执行状态试验床。数据库变体另有DBA-Bench的有负载运维实测，显示诊断、恢复和安全执行仍有缺口；本事故未重放。","nextValidation":"在隔离微服务环境冻结DNS/端口或资源故障，先证明探针捕获错误，再让Agent修复并重复负载/第二事件，保留每条动作与恢复时间。","futureSignals":["连续故障及后续变更下用户路径恢复，数据和权限保持完整，可考虑扩大相近输入范围。","只消除告警而业务未恢复，或恢复动作破坏约束/反复复发，应收窄或暂停相应采用。"]},"workflow":{"before":"值班工程师切换日志、图表、云控制台和同事排查。","now":"Agent自主关联遥测、查看配置、试验可逆动作与复测；确定性探针持续检查服务，越过权限或预算返回阻塞。","retained":["设定服务目标、风险容忍和操作授权，负责不可逆切换与事故沟通；接收以服务证据而非人的操作占比。"],"added":["重新施加代表负载或第二事件，检查原修复、容量和复发条件。"]},"humanResponsibility":"设定服务目标、风险容忍和操作授权，负责不可逆切换与事故沟通；接收以服务证据而非人的操作占比。","failureHandling":"信号矛盾时保留假设和反证，不把最近变更当唯一根因；恢复失败交付尝试、当前状态和升级包，维持能恢复的路径。","example":{"input":"发布后支付请求超时，服务还活着但下游连接受策略阻断。","output":"配置差异、受限修复、支付探针和回退/第二次负载结果。","failure":"直接允许全网出站，让告警清零却越过网络边界。"},"sourceIds":["WU-C-AIOPS","WU-C-DBA"],"evidenceLinks":[{"sourceId":"WU-C-AIOPS","supports":"可运行事故任务与独立状态检查方法。","notCovered":"框架存在不是目标事故已恢复，也不是长期SRE成绩。"},{"sourceId":"WU-C-DBA","supports":"数据库事故有直接状态运维评估。","notCovered":"不是本站故障或所有生产数据库的成功率。"}],"resourceRefs":[],"career":{"direction":"SRE、云网络、数据库运行与服务维护","evidence":"可重建配置/恢复补丁与实际执行或待执行边界","nextStep":"在隔离微服务环境冻结DNS/端口或资源故障，先证明探针捕获错误，再让Agent修复并重复负载/第二事件，保留每条动作与恢复时间。"},"team":{"bottleneck":"AIOpsLab可将故障、负载、观察和评估接在同一环境，支持按系统状态验收；真实云拓扑、权限及业务不变量决定迁移边界。稳定配置可直接用脚本执行，Agent主要处理异常和不完整信息。","allocation":"Agent自主关联遥测、查看配置、试验可逆动作与复测；确定性探针持续检查服务，越过权限或预算返回阻塞。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C08-ENV","statement":"AIOpsLab公开提供故障注入、负载、遥测与按任务结果评估的接口，可用于验证Agent是否改变了真实服务状态。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-C-AIOPS","locator":"README Problems：Fault/Workload/Evaluator","relation":"partial","supports":"可运行事故任务与独立状态检查方法。","limit":"框架存在不是目标事故已恢复，也不是长期SRE成绩。"}],"configuration":null,"progressIds":[],"limitation":"框架存在不是目标事故已恢复，也不是长期SRE成绩。","evidenceStatus":"active"},{"id":"C08-ADOPT","statement":"可在受限权限与独立用户探针下委派环境改造/故障恢复；先演练可逆变更，再接收持续业务恢复及副作用。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-AIOPS","locator":"Detection/Localization/Analysis/Mitigation","relation":"partial","supports":"支持按不同结果分别接收。","limit":"真实网络、数据、依赖及成本需任务现场验证。"}],"configuration":null,"progressIds":[],"limitation":"真实网络、数据、依赖及成本需任务现场验证。","evidenceStatus":"active"},{"id":"C08-DBSTATE","statement":"DBA-Bench在活跃负载下观察到诊断、业务恢复和安全执行之间的缺口；数据库事故必须同时检查三者。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-C-DBA","locator":"v2 Abstract；bounded/persistent scenarios","relation":"partial","supports":"数据库事故有直接状态运维评估。","limit":"不是本站故障或所有生产数据库的成功率。"}],"progressIds":["P-C-DBA-STATE"],"configuration":"v2：PostgreSQL/TiDB，231有界与10持续场景、8自动基线；原始模型/脚手架与约束按论文区分，本站未重放。","limitation":"仅按任务/环境解释，不能由根因分类推断实际恢复。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开AIOpsLab提供故障/负载/遥测试验床；真实事故权限、拓扑和用户探针另备。","tools":[{"name":"AIOpsLab","url":"https://github.com/microsoft/AIOpsLab","version":"2026-10-20公开仓库HEAD ccf08d0d1d5fa5b30f120e2e8549662d44411b35；实际运行另固定依赖/镜像/模型","access":"开源；独立集群/命名空间、遥测与模型API自备"}]},"variants":[{"title":"事故与渐进发布","inputDifference":"运行中异常，可能伴随发布或依赖变化。","assessmentChange":"按观察—可逆干预—恢复证据分阶段，但Agent可自行选择调查路径。","acceptanceChange":"金丝雀/回滚、用户请求、事务状态与完整时间线。"},{"title":"数据库运行故障","inputDifference":"锁/连接/配置/容量或依赖故障影响数据库业务，未必需要备份还原。","assessmentChange":"按数据库事务与负载接收；确需备份还原再交W-C16，不能为修复牺牲数据。","acceptanceChange":"事务/约束/权限、原负载与恢复后再次变化；保持数据和审计。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C10","domainId":"computing","title":"嵌入式交付：固件、接口与可重放设备行为","goal":"交付在指定板卡/外设/实时约束下可运行的固件变更，连同实物或明确标为仿真的接口、时序和故障恢复证据。","trigger":"给板卡/芯片修订、SDK/RTOS、引脚与接线、外设datasheet、输入刺激和可接受输出。","scope":"代码、原理图、驱动和跨平台适配按目标硬件联合接收；仿真只能支持其建模范围，不能认证电气安全、EMC或工业最坏时序。","inputs":[{"name":"设备目标","detail":"给板卡/芯片修订、SDK/RTOS、引脚与接线、外设datasheet、输入刺激和可接受输出。"},{"name":"约束与仪器","detail":"冻结内存/时序/功耗/并发预算及异常条件，提供可刷写设备、调试器/串口/逻辑分析或等效可信观测。"},{"name":"恢复条件","detail":"准备可回退镜像、启动方式、看门狗/掉电/升级失败策略及不会损坏设备的测试范围。"}],"deliverables":["固件/驱动/配置、可构建镜像与接线版本","协议/中断/资源/时序观测和失败输入回放","跨平台差异、实物/仿真覆盖边界与升级恢复产物"],"acceptance":[{"criterion":"接口语义成立","method":"在目标接线/外设施加输入，核对寄存器/协议/采样与实际输出；软件日志与外部观测交叉验证。","failure":"编译通过但引脚或采样单位错误。"},{"criterion":"资源与并发可接收","method":"测试中断/任务交错、DMA与内存生命周期、代表及最坏可测负载，记录时延/堆栈/丢样。","failure":"轮询演示通过，却中断下丢数据或阻塞超时。"},{"criterion":"失败不会不可恢复","method":"在明确保护范围内中断升级/断电/外设失联，核对旧镜像启动和诊断信息。","failure":"只测正常烧写，半途掉电后设备无法启动。"}],"assessment":{"recommendation":"可委派固件候选、构建和仿真修正；有目标设备与独立接口/时序观测后，再接收实际硬件交付。","conditions":["冻结内存/时序/功耗/并发预算及异常条件，提供可刷写设备、调试器/串口/逻辑分析或等效可信观测。","准备可回退镜像、启动方式、看门狗/掉电/升级失败策略及不会损坏设备的测试范围。"],"evidenceBasis":"EmbedAgent针对固件编程、原理图和跨平台迁移做了专属实验，不同平台/语言表现差异显著；这支持对设备条件分开判断；官方代码提供Wokwi模拟与任务资产，可检查固件/电路的联合结果。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["目标实物电气/接口、最坏可测时序、DMA/资源与掉电恢复证据尚缺。"],"mode":"staged-delivery","rationale":"EmbedAgent直接表明板卡、语言、原理图与迁移条件影响表现，不能照搬通用代码成绩。Agent可自己搜索datasheet和调工具，但必须让物理反馈进入接收，而不是由代码自述完成。","frontier":"EmbedAgent有三平台/九组件虚拟电路评估，不能覆盖工业板卡、电气约束、真实DMA/cache与最坏时序；本文不迁移分数。","nextValidation":"冻结一项传感器采集/驱动任务，让Agent交镜像并在受保护板卡运行；以外部信号回放、超时/失联和升级中断控制验证。","futureSignals":["换板卡/组件后实物接口、资源和失败恢复继续满足约束，可考虑扩大相近输入范围。","仿真与物理信号不符、实时边界或安全属性未被观测，应收窄或暂停相应采用。"]},"workflow":{"before":"固件编程、接线、调试与测试分别往返，代码可编译就进入实物排查。","now":"Agent可联合修改代码/配置/电路描述，按编译和设备反馈搜索；独立刺激和仪器决定真实行为是否接收。","retained":["确定硬件保护、时序/资源与失效政策；产品认证和不可逆烧录由有相应权限者决定。"],"added":["在明确保护范围内中断升级/断电/外设失联，核对旧镜像启动和诊断信息。"]},"humanResponsibility":"确定硬件保护、时序/资源与失效政策；产品认证和不可逆烧录由有相应权限者决定。","failureHandling":"设备或观测不足只交明确仿真/构建成果；保留无法解释的波形、复位与错误，不将外设返回默认值当正常读数。","example":{"input":"I2C传感器定时采样并DMA发串口，外设偶发失联，固件需可更新。","output":"镜像、引脚/SDK清单、采样波形、超时/并发与恢复结果。","failure":"仿真输出稳定，却实物地址/电压错误或中断竞争丢样。"},"sourceIds":["WU-C-EMBED","WU-C-EMBED-REPO"],"evidenceLinks":[{"sourceId":"WU-C-EMBED","supports":"嵌入式任务有直接研究。","notCovered":"虚拟电路、小板卡任务，不是工业硬件可靠率。"},{"sourceId":"WU-C-EMBED-REPO","supports":"存在可检查仿真入口。","notCovered":"API/仿真与实物条件不同，不能代替所有硬件测试。"}],"resourceRefs":[],"career":{"direction":"嵌入式开发","evidence":"固件/驱动/配置、可构建镜像与接线版本","nextStep":"冻结一项传感器采集/驱动任务，让Agent交镜像并在受保护板卡运行；以外部信号回放、超时/失联和升级中断控制验证。"},"team":{"bottleneck":"EmbedAgent直接表明板卡、语言、原理图与迁移条件影响表现，不能照搬通用代码成绩。Agent可自己搜索datasheet和调工具，但必须让物理反馈进入接收，而不是由代码自述完成。","allocation":"Agent可联合修改代码/配置/电路描述，按编译和设备反馈搜索；独立刺激和仪器决定真实行为是否接收。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C10-SIM","statement":"EmbedAgent针对固件编程、原理图和跨平台迁移做了专属实验，不同平台/语言表现差异显著；这支持对设备条件分开判断。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-C-EMBED","locator":"Abstract；roles and cross-platform migration","relation":"partial","supports":"嵌入式任务有直接研究。","limit":"虚拟电路、小板卡任务，不是工业硬件可靠率。"}],"configuration":"v3：126案例、9组件、3平台、10种LLM；编程/设计原理图/迁移与检索-编译反馈分别评估，本站未运行。","progressIds":["P-C-EMBED-SIM"],"limitation":"虚拟电路、小板卡任务，不是工业硬件可靠率。","evidenceStatus":"active"},{"id":"C10-TOOLS","statement":"官方代码提供Wokwi模拟与任务资产，可检查固件/电路的联合结果。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-C-EMBED-REPO","locator":"README evaluate/Wokwi","relation":"partial","supports":"存在可检查仿真入口。","limit":"API/仿真与实物条件不同，不能代替所有硬件测试。"}],"configuration":null,"progressIds":[],"limitation":"API/仿真与实物条件不同，不能代替所有硬件测试。","evidenceStatus":"active"},{"id":"C10-ADOPT","statement":"可把候选固件、配置和调试委派为一包，但跨到目标实物后只按其接口、时序、资源和恢复证据接收。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-EMBED","locator":"Platform differences/feedback","relation":"partial","supports":"从平台差异推导阶段接收。","limit":"本板卡、RTOS、工具版本和异常条件需要现场核验。"}],"configuration":null,"progressIds":[],"limitation":"本板卡、RTOS、工具版本和异常条件需要现场核验。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"有公开仿真任务；实物刷写、外部观测、RTOS/SDK按任务提供，没有本站设备试验。","tools":[{"name":"EmbedAgent/Embedbench","url":"https://github.com/icip-cas/EmbedAgent","version":"2026-10-20公开仓库HEAD 877bae655797d01180b562a673a4dfdaaddbadd6；实际运行另固定依赖/镜像/模型","access":"源码公开；模型、Wokwi和目标工具链条件另核"}]},"variants":[{"title":"硬件接口与跨平台","inputDifference":"更换MCU、SDK、RTOS或外设接线。","assessmentChange":"先核对地址/引脚/中断/库语义；旧板成绩不迁移。","acceptanceChange":"协议刺激、边界值、校准/单位和失联。"},{"title":"实时资源与并发","inputDifference":"中断、DMA、cache、低功耗与多个任务交互。","assessmentChange":"分阶段验证可观测不变量，未测最坏时序必须标明。","acceptanceChange":"锁/竞争、堆栈、采样丢失和时延/唤醒检查。"},{"title":"启动升级与恢复","inputDifference":"引导、分区、镜像签名或掉电升级。","assessmentChange":"候选可委派，接收需保护下的实物恢复。","acceptanceChange":"不同断点掉电、镜像校验、回退启动与原用户数据。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C11","domainId":"computing","title":"系统方案交付：需求、备选架构与验收契约","goal":"交付依据实际流程/数据/接口的系统方案，含可追溯需求、选型取舍、迁移设计和关键假设验证。","trigger":"给现状工作流、参与角色、信息需求、失败/成本与需要改善的可检验结果。","scope":"产物为系统分析与可实施契约；项目范围/资源承诺与实际推进状态由W-C22接收，系统采用由W-C24验证。","inputs":[{"name":"问题与业务目标","detail":"给现状工作流、参与角色、信息需求、失败/成本与需要改善的可检验结果。"},{"name":"系统事实","detail":"冻结已有接口/schema、用户样例、系统边界、性能/安全/法规/预算约束与证据来源。"},{"name":"方案检验","detail":"明确备选方案、无法改变的约束、原型环境和验收标准；外部产品/同事意见要有真实来源。"}],"deliverables":["现状/信息需求与目标、可追溯需求和验收条件","备选架构/选型与成本收益/约束/失败恢复取舍","接口/数据映射、集成迁移设计和关键假设原型结果"],"acceptance":[{"criterion":"需求不是凭空补齐","method":"抽查角色任务/数据到原记录；不明确处给影响对照并取得实际取舍。","failure":"由一句需求自行假定所有角色/权限。"},{"criterion":"方案能满足约束","method":"按接口/数据样例与原型检查关键兼容/容量/成本假设，保留替代方案。","failure":"画架构图却无法处理实际数据或依赖。"},{"criterion":"契约可用于交付接收","method":"需求→接口/数据→验收可追踪，迁移含切换/失败恢复与运营/采用条件。","failure":"方案只有产品功能清单，没有数据转换或验收。"}],"assessment":{"recommendation":"可委派现状核对、备选设计和原型验证；以事实与约束闭合接收方案，不以篇幅或图表数量接收。","conditions":["冻结已有接口/schema、用户样例、系统边界、性能/安全/法规/预算约束与证据来源。","明确备选方案、无法改变的约束、原型环境和验收标准；外部产品/同事意见要有真实来源。"],"evidenceBasis":"公开公司基准提供跨网站/文档/代码的工作资料环境，可作为系统资料整合与事实检查的方法参照。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["真实需求歧义、供应商接口与成本假设尚需项目样例/原型反证；没有直接方案质量实测。"],"mode":"staged-delivery","rationale":"公司环境和跨工具资料可支持整合方法，但没有本系统选型的直接能力成绩。AI价值在联合搜索方案、验证接口/数据假设并呈现可执行取舍。","frontier":"TheAgentCompany公开跨网站/文档/代码的公司任务及结果检查，证明有可操作工作资料环境；该资料未单独展示本系统需求、选型与迁移方案的AI完成性，因此关键方案须由实际样例/原型接收。","nextValidation":"用一个真实集成/替换决策冻结样例接口和数据，给两个关键冲突约束；检查Agent备选是否可执行并以原型反证错误假设。","futureSignals":["关键方案假设经实际接口/数据原型反复成立，取舍仍可追溯，可考虑扩大相近输入范围。","选型依赖未兑现、需求冲突被隐藏或成本约束失真，应收窄或暂停相应采用。"]},"workflow":{"before":"资料、实现/处理与验证分散交接，完成记录易脱离实际结果。","now":"Agent可按目标访问资料、构造备选、计算成本并原型验证；实际业务取舍确定后生成可追踪契约。","retained":["确认业务目标、利益冲突和成本风险取舍；专业/法规约束由真实依据确定。"],"added":["需求→接口/数据→验收可追踪，迁移含切换/失败恢复与运营/采用条件。"]},"humanResponsibility":"确认业务目标、利益冲突和成本风险取舍；专业/法规约束由真实依据确定。","failureHandling":"事实不齐交影响分析与最低核验，不能虚构接口能力/供应商承诺。","example":{"input":"把手工订单对账接入旧系统，预算与数据保留政策固定。","output":"备选方案、接口/身份/数据映射、样例原型和验收/迁移契约。","failure":"只推荐最流行架构，不处理旧接口和数据保留。"},"sourceIds":["WU-C-COMPANY-REPO","WU-C-COMPANY"],"evidenceLinks":[{"sourceId":"WU-C-COMPANY-REPO","supports":"有可检查工作资料和结果环境。","notCovered":"并未提供本系统方案或选型成绩。"},{"sourceId":"WU-C-COMPANY","supports":"邻近跨工具任务支持资料整合的可行性。","notCovered":"真实组织方案质量与采用效果需原型和实际用户核验。"}],"resourceRefs":[],"career":{"direction":"系统方案交付：需求、备选架构与验收契约","evidence":"现状/信息需求与目标、可追溯需求和验收条件","nextStep":"用一个真实集成/替换决策冻结样例接口和数据，给两个关键冲突约束；检查Agent备选是否可执行并以原型反证错误假设。"},"team":{"bottleneck":"公司环境和跨工具资料可支持整合方法，但没有本系统选型的直接能力成绩。AI价值在联合搜索方案、验证接口/数据假设并呈现可执行取舍。","allocation":"Agent可按目标访问资料、构造备选、计算成本并原型验证；实际业务取舍确定后生成可追踪契约。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C11-CONTEXT","statement":"公开公司基准提供跨网站/文档/代码的工作资料环境，可作为系统资料整合与事实检查的方法参照。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-C-COMPANY-REPO","locator":"README environment/task roles/result-based grading","relation":"partial","supports":"有可检查工作资料和结果环境。","limit":"并未提供本系统方案或选型成绩。"}],"configuration":null,"progressIds":[],"limitation":"并未提供本系统方案或选型成绩。","evidenceStatus":"active"},{"id":"C11-ADOPT","statement":"可委派备选系统方案与关键假设原型，接收需以真实需求、接口/数据、成本和可执行验收闭合。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-COMPANY","locator":"simulated work environment","relation":"partial","supports":"邻近跨工具任务支持资料整合的可行性。","limit":"真实组织方案质量与采用效果需原型和实际用户核验。"}],"configuration":null,"progressIds":[],"limitation":"真实组织方案质量与采用效果需原型和实际用户核验。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开可重建公司环境可研究工作法；真实工单/看板/日历访问与写权限需组织提供，未接入本站。","tools":[{"name":"TheAgentCompany","url":"https://github.com/TheAgentCompany/TheAgentCompany","version":"2026-10-20公开仓库HEAD 98b68ef82a47690c316f42fddb05baafaab56851；实际运行另固定依赖/镜像/模型","access":"开源；隔离部署和模型账户自备"}]},"variants":[{"title":"流程与需求契约","inputDifference":"信息需求、角色和当前流程不一致。","assessmentChange":"先交可检验目标及口径取舍，再形成接口/数据验收。","acceptanceChange":"需求来源、权限/数据样例和正反行为。"},{"title":"选型与迁移设计","inputDifference":"需买/建/替换系统，存在成本与旧系统约束。","assessmentChange":"备选并列，先原型挑战高风险假设。","acceptanceChange":"接口/数据转换、容量/费用、切换/恢复设计；实现另接收。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C12","domainId":"computing","title":"验收器与测试资产交付：能拒绝错误，也能接受合法替代","goal":"交付可运行的行为测试/复现与风险覆盖说明，使它捕获已知错误而不把正确实现误判为失败。","trigger":"给用户/系统应满足的结果、重要不变量、历史缺陷与受影响路径；明确无法自动判定的体验或专业条件。","scope":"单元、集成、端到端和非功能按风险选取；不把测试条数、覆盖率或CI绿当作系统质量。复杂环境检查结合对应任务包。","inputs":[{"name":"行为与风险","detail":"给用户/系统应满足的结果、重要不变量、历史缺陷与受影响路径；明确无法自动判定的体验或专业条件。"},{"name":"独立控制","detail":"固定基线、一个正确/合法替代、一个有实际错误的版本及环境/数据；避免仅复制被测实现逻辑。"},{"name":"可运行反馈","detail":"提供稳定启动、测试隔离、依赖/时钟/并发控制、日志与允许的运行成本。"}],"deliverables":["可重放测试/复现脚本、测试数据与环境说明","正确/错误/合法替代控制结果和失败诊断","风险到检查的覆盖映射、遗漏、flaky原因及发布差距"],"acceptance":[{"criterion":"确实区分错误","method":"同一检查在已知错误版本失败、修复/合法替代通过，失败原因对应需求。","failure":"为某种代码写法锁死断言，合法替代也失败。"},{"criterion":"不靠环境偶然通过","method":"重复重放、清空状态、变顺序，明确超时/依赖与并发；记录不稳定失败。","failure":"测试共享脏数据库或sleep碰巧通过。"},{"criterion":"覆盖与发布结论相称","method":"按关键风险、权限/数据/环境路径给剩余风险，使用独立缺陷控制挑战测试。","failure":"覆盖率100%就认定没有遗漏或安全问题。"}],"assessment":{"recommendation":"可委派Agent从行为到测试/复现再校准；先证明检查能区分正反控制，再拿它接收功能或发布。","conditions":["固定基线、一个正确/合法替代、一个有实际错误的版本及环境/数据；避免仅复制被测实现逻辑。","提供稳定启动、测试隔离、依赖/时钟/并发控制、日志与允许的运行成本。"],"evidenceBasis":"SWT-Bench公开给出仓库级复现测试、gold patches和执行日志，能检查错误前失败与修复后通过。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["尚缺本产品独立真实缺陷、正确与合法替代控制，以及flaky/确认时间记录。"],"mode":"bounded-delivery","rationale":"SWT-Bench直接把复现测试前后运行分开，并区分复现脚本与全套测试。把验收器当交付对象可以合并测试生成/环境诊断/回归，而不强迫先写大量映射实现的测试。","frontier":"公开有仓库级测试生成与gold控制；它的覆盖变化不等于业务查全率，真实浏览器/移动/硬件验收还需环境。","nextValidation":"用一组真实缺陷、正确修复和不同正确实现盲测，冻结需求后让Agent生成检查，再由独立运行核对拒绝/接收、稳定性与诊断质量。","futureSignals":["独立新缺陷被捕获，正确和合理替代被接收，flaky成本可控，可考虑扩大相近输入范围。","测试只迎合某种写法、环境偶然通过或关键错误漏检，应收窄或暂停相应采用。"]},"workflow":{"before":"QA按用例写脚本，开发修环境，业务最后点验。","now":"Agent可自主构造输入、启动环境、生成检查并用控制修订；程序反复运行，人按风险判断何种证据足够。","retained":["确定质量风险和发布容忍；验收器的语义与不可自动判断范围需独立核对。"],"added":["按关键风险、权限/数据/环境路径给剩余风险，使用独立缺陷控制挑战测试。"]},"humanResponsibility":"确定质量风险和发布容忍；验收器的语义与不可自动判断范围需独立核对。","failureHandling":"发现验收器错误先修正并重跑双方控制；环境失败单列，不把失败转跳过后宣布通过。","example":{"input":"付款重复回调造成双记账，正确替代可用事务或唯一键。","output":"两种正确实现通过，双记账和缺权限的实现失败；附隔离数据和事件回放。","failure":"只检查某函数被调用一次，错过真实账本副作用。"},"sourceIds":["WU-C-SWT"],"evidenceLinks":[{"sourceId":"WU-C-SWT","supports":"提供直接测试任务资产。","notCovered":"复现脚本模式不包含全套回归；语料不覆盖全部产品风险。"}],"resourceRefs":[],"career":{"direction":"软件质量、测试工程","evidence":"可重放测试/复现脚本、测试数据与环境说明","nextStep":"用一组真实缺陷、正确修复和不同正确实现盲测，冻结需求后让Agent生成检查，再由独立运行核对拒绝/接收、稳定性与诊断质量。"},"team":{"bottleneck":"SWT-Bench直接把复现测试前后运行分开，并区分复现脚本与全套测试。把验收器当交付对象可以合并测试生成/环境诊断/回归，而不强迫先写大量映射实现的测试。","allocation":"Agent可自主构造输入、启动环境、生成检查并用控制修订；程序反复运行，人按风险判断何种证据足够。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C12-REPRO","statement":"SWT-Bench公开给出仓库级复现测试、gold patches和执行日志，能检查错误前失败与修复后通过。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-C-SWT","locator":"Overview；Gold validation；evaluation logs","relation":"partial","supports":"提供直接测试任务资产。","limit":"复现脚本模式不包含全套回归；语料不覆盖全部产品风险。"}],"configuration":null,"progressIds":[],"limitation":"复现脚本模式不包含全套回归；语料不覆盖全部产品风险。","evidenceStatus":"active"},{"id":"C12-ADOPT","statement":"可让AI交付并校准验收器，先用真实正反控制和合理替代测试它，再用它判断其他Agent产物。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-SWT","locator":"Unit Test mode vs Reproduction Script mode","relation":"partial","supports":"支持按实际执行结果检查测试。","limit":"不能由已知控制表现推出未知缺陷检出率。"}],"configuration":null,"progressIds":[],"limitation":"不能由已知控制表现推出未知缺陷检出率。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"开源测试资产可检查；没有运行新的模型生成测试。","tools":[{"name":"SWT-Bench","url":"https://github.com/logic-star-ai/swt-bench","version":"2026-10-20公开仓库HEAD 330a649a764fab2fadaea632776eeae87272f74b；实际运行另固定依赖/镜像/模型","access":"开源；容器、仓库与模型调用预算另备"}]},"variants":[{"title":"端到端与真实环境","inputDifference":"浏览器、设备、DB或硬件行为重要。","assessmentChange":"脚本可委派，但必须与目标环境实际状态连接。","acceptanceChange":"采用W-C04/C05/C07/C10环境不变量，Mock边界显式记录。"},{"title":"非功能与对抗输入","inputDifference":"性能、负载、权限或并发条件主导风险。","assessmentChange":"冻结负载/故障/对抗控制后分阶段接收。","acceptanceChange":"预算与阈值不能事后修改；检查负控制和副作用。"},{"title":"回归与发布","inputDifference":"测试来自修复或持续交付。","assessmentChange":"交付发布差距与证据，不能只给总通过数。","acceptanceChange":"历史错误重放、flaky隔离、关键风险未覆盖与版本差异。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C13","domainId":"computing","title":"IT用户任务恢复与配置交付","goal":"在获准设备/账户中恢复一个具体用户任务，交付可核对的配置状态、操作/诊断包和面向用户的验证说明。","trigger":"给受影响任务、设备/OS/账号身份、错误、时间、影响范围与用户能确认的恢复结果。","scope":"目标是特定用户原来失败任务的恢复；新设备/软件配置与迁移见W-C23，用户指引/培训采用见W-C24。不能访问设备时只接收具体诊断升级包。","inputs":[{"name":"用户目标与故障","detail":"给受影响任务、设备/OS/账号身份、错误、时间、影响范围与用户能确认的恢复结果。"},{"name":"可操作与保护","detail":"冻结可用遥测、管理入口、允许远程动作、数据备份、恢复凭据和不允许删除的资料；用户私人内容不扩散。"},{"name":"验证与升级","detail":"提供原用户任务的重放方式、正常参照和升级联系人/诊断字段；不能观察设备时允许返回具体阻塞。"}],"deliverables":["原因线索与可回放操作/配置差异","原用户任务恢复或仍受阻的状态证据","用户步骤/迁移验收、诊断升级包和可维护知识条目"],"acceptance":[{"criterion":"恢复对应原用户","method":"以同设备/身份重复原工作，检查登录/权限/数据/网络，必要时重启重进。","failure":"管理员能登录就称用户已恢复。"},{"criterion":"配置不扩大风险","method":"核对权限、策略、版本、备份和数据前后；拒绝以关闭保护或清空资料来消除报错。","failure":"禁用所有安全策略解决登录，却扩大越权。"},{"criterion":"交接可继续","method":"保存错误/版本/已试步骤和时间线，指引由用户或另一支持者重放；记录不能远程解决部分。","failure":"复制通用教程，无具体设备状态且没有升级信息。"}],"assessment":{"recommendation":"可委派信息收集、诊断和获准配置操作；以原用户实际任务恢复接收，缺设备/身份/遥测时交可用升级包。","conditions":["冻结可用遥测、管理入口、允许远程动作、数据备份、恢复凭据和不允许删除的资料；用户私人内容不扩散。","提供原用户任务的重放方式、正常参照和升级联系人/诊断字段；不能观察设备时允许返回具体阻塞。"],"evidenceBasis":"微软IT管理员随机试验在登录排障、策略与设备故障情景观察到Copilot辅助的准确性/耗时改善；这不是无人维修实测；Security Copilot在Entra/Intune等工具中的接入受租户、产品权限和容量影响；不能由产品文档推定已经在本用户可用。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["尚未访问目标设备/用户身份验证恢复、保护和重启结果，原RCT不评无人维修。"],"mode":"staged-delivery","rationale":"IT管理员RCT支持人加Copilot的信息和故障工作重组；OSWorld2显示长流程仍会丢约束。将恢复目标和配置副作用同时检查，可把AI建议转成真实交付，不能把答对问卷当修好电脑。","frontier":"直接证据是Entra/Intune辅助管理员情景，通用跨应用基准也非企业端点恢复专属。物理故障、目标租户策略与当前设备效果仍未知。","nextValidation":"取一台获准测试设备和测试账户，冻结一次真实登录/策略故障；让Agent交诊断并执行可逆改动，由原用户路径、权限负例与重启验证恢复。","futureSignals":["新设备/策略条件下原用户任务恢复，数据和保护保持完整，可考虑扩大相近输入范围。","管理员视角通过而用户仍受阻，或依赖过度提权/清数据，应收窄或暂停相应采用。"]},"workflow":{"before":"支持者问用户、翻日志、查知识库，再远程操作或升级。","now":"Agent自主关联用户描述、设备/账号遥测与知识，执行授权内可逆修复并重测；恢复失败时一次性交付可继续的诊断包。","retained":["核对用户身份/设备授权、数据保护与高影响账号动作，确认用户任务是否真正恢复。"],"added":["保存错误/版本/已试步骤和时间线，指引由用户或另一支持者重放；记录不能远程解决部分。"]},"humanResponsibility":"核对用户身份/设备授权、数据保护与高影响账号动作，确认用户任务是否真正恢复。","failureHandling":"用户不可配合、设备不在线或缺管理权限时返回具体收集步骤和升级包；不声称已在不可见设备操作。","example":{"input":"新策略后用户无法访问业务应用，管理员仍可访问。","output":"该身份/设备策略证据、受限修改、重启后的用户访问与越权负例。","failure":"重置所有策略或删除本地资料，暂时不报错但原任务仍失败。"},"sourceIds":["WU-C-ITRCT","WU-C-ITACCESS","WU-C-OS2"],"evidenceLinks":[{"sourceId":"WU-C-ITRCT","supports":"IT支持工作有直接领域试验。","notCovered":"厂商相关、参与者加Copilot；本页不转用总体提升幅度。"},{"sourceId":"WU-C-ITACCESS","supports":"明确当前可用条件。","notCovered":"未在本站租户启用，不保证所有设备动作可执行。"},{"sourceId":"WU-C-OS2","supports":"长流程约束风险支持结果接收。","notCovered":"通用桌面任务只作迁移边界，不提供IT修复率。"}],"resourceRefs":[],"career":{"direction":"IT支持、用户服务与端点管理","evidence":"原因线索与可回放操作/配置差异","nextStep":"取一台获准测试设备和测试账户，冻结一次真实登录/策略故障；让Agent交诊断并执行可逆改动，由原用户路径、权限负例与重启验证恢复。"},"team":{"bottleneck":"IT管理员RCT支持人加Copilot的信息和故障工作重组；OSWorld2显示长流程仍会丢约束。将恢复目标和配置副作用同时检查，可把AI建议转成真实交付，不能把答对问卷当修好电脑。","allocation":"Agent自主关联用户描述、设备/账号遥测与知识，执行授权内可逆修复并重测；恢复失败时一次性交付可继续的诊断包。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C13-ITRCT","statement":"微软IT管理员随机试验在登录排障、策略与设备故障情景观察到Copilot辅助的准确性/耗时改善；这不是无人维修实测。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-C-ITRCT","locator":"Abstract；three scenarios","relation":"partial","supports":"IT支持工作有直接领域试验。","limit":"厂商相关、参与者加Copilot；本页不转用总体提升幅度。"}],"configuration":"v2：Entra/Intune中有/无Security Copilot的随机组；三情景，问答/自由回答口径区分；v2修正时间数字。","progressIds":["P-C-IT-RCT"],"limitation":"厂商相关、参与者加Copilot；本页不转用总体提升幅度。","evidenceStatus":"active"},{"id":"C13-ACCESS","statement":"Security Copilot在Entra/Intune等工具中的接入受租户、产品权限和容量影响；不能由产品文档推定已经在本用户可用。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-ITACCESS","locator":"Eligibility；permissions；agents deployment","relation":"method","supports":"明确当前可用条件。","limit":"未在本站租户启用，不保证所有设备动作可执行。"}],"configuration":null,"progressIds":[],"limitation":"未在本站租户启用，不保证所有设备动作可执行。","evidenceStatus":"active"},{"id":"C13-RESTORE","statement":"Agent可整合诊断与可逆配置，但完整恢复必须由原用户任务、身份和数据/保护不变量验证。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-OS2","locator":"Abstract：hidden state/constraints","relation":"partial","supports":"长流程约束风险支持结果接收。","limit":"通用桌面任务只作迁移边界，不提供IT修复率。"}],"configuration":null,"progressIds":[],"limitation":"通用桌面任务只作迁移边界，不提供IT修复率。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"有条件商业助手与开源桌面试验环境；未连接任何用户设备。","tools":[{"name":"Security Copilot","url":"https://learn.microsoft.com/en-us/copilot/security/security-copilot-inclusion","version":"2026-10-20核查；租户实际版本/启用另记","access":"需产品许可、容量、角色与组织设置"},{"name":"OSWorld","url":"https://github.com/xlang-ai/OSWorld","version":"2026-10-20公开仓库HEAD 9f4b4d34319589f3ba6947da6045b0955b08900e；实际运行另固定依赖/镜像/模型","access":"开源隔离桌面环境，模型与系统资源另备"}]},"variants":[{"title":"用户登录与设备配置","inputDifference":"身份/策略/客户端版本交互，含安装或连接。","assessmentChange":"先还原原用户故障，按最小可逆变更恢复。","acceptanceChange":"原身份成功、越权负例、重启和策略回归。"},{"title":"复杂升级与物理缺口","inputDifference":"硬件、网络管理或其他团队权限超出远程范围。","assessmentChange":"交具体可继续诊断包，不假装完成维修。","acceptanceChange":"版本/错误/时间线、复现、尝试及责任范围清楚；升级后再验证。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C14","domainId":"computing","title":"算法候选交付：代码搜索、反例与正确性证据","goal":"针对固定问题交付新算法/系统候选及对应的正确性、复杂性或经验改进证据，未解决则说明预算与卡点。","trigger":"已有精确定义的问题和基线，需要探索更好的候选算法/系统，并用独立评价器接收。","scope":"对象为明确问题上的新算法/系统候选与可检查论证，允许返回预算内未解决；已有论文贡献的复现另由W-C25接收。","inputs":[{"name":"问题与基线","detail":"输入域、输出语义、已有算法/结果、目标性质、不可放宽约束与必要文献。"},{"name":"评价与预算","detail":"独立正确性/性能评价器、保留输入、硬件与计算/时间/尝试预算；理论结论给命题和允许假设。"},{"name":"验证环境","detail":"编译/求解/反例工具和对照实现固定；错误/合法替代先校准，搜索者不能静默改评分器。"}],"deliverables":["候选算法/系统、精确定义与可运行实现","反例/正确性/复杂性论证或保留输入下的同条件实验和失败账本","相对基线改进、退化边界、预算及未解决部分"],"acceptance":[{"criterion":"问题语义保持","method":"独立核对输入域、输出和约束，反例/边界及保留输入挑战候选。","failure":"放宽约束或只缓存公开样例。"},{"criterion":"证据匹配主张","method":"一般性/复杂性给完整论证，经验改善同条件比较；数值搜索不能代替一般证明。","failure":"样例更快就声称一般最优。"},{"criterion":"结果可重查","method":"固定环境重跑基线/候选，保留选择、预算与失败，记录评价器缺口。","failure":"只给最好一次，隐藏改过评分器。"}],"assessment":{"recommendation":"可委派新候选、实现和反例循环；按独立评价接收有效成果，理论推广与新颖性另核。","conditions":["独立正确性/性能评价器、保留输入、硬件与计算/时间/尝试预算；理论结论给命题和允许假设。","编译/求解/反例工具和对照实现固定；错误/合法替代先校准，搜索者不能静默改评分器。"],"evidenceBasis":"AlphaEvolve专用系统报告算法改进，并有部分公开成果；本站未重放。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["新问题语义、可获得搜索配置、独立评价器和一般复杂性/正确性论证尚需逐项核验。"],"mode":"candidate-exploration","rationale":"问题和独立评价器可把算法候选搜索与反证接成循环；结果可以包含代码、可核查特例或论证，但不能从得分直接推广一般理论。","frontier":"专用研究系统已有可检查算法成果；搜索配置、任务选择与预算限制公开Agent迁移，目标问题表现未知。","nextValidation":"固定一个算法问题、基线、评价器和预算，先校准错误/合法替代，再搜索候选并在保留输入重跑，独立检查理论主张。","futureSignals":["新问题和独立保留输入上候选有可核验改进，论证不偷换目标，可考虑扩大相近输入范围。","钻评价器漏洞、仅公开样例变好或一般性主张无相应证明，应收窄或暂停相应采用。"]},"workflow":{"before":"候选、实现、反例与资源实验分散进行。","now":"Agent可联合搜索算法/实现/验证路径，程序或求解器给反馈；目标语义、独立评价器和预算保持固定。","retained":["研究价值、语义与可接受证明由研究负责人确定。"],"added":["核对保留输入、失败账本与评价器未被偷换。"]},"humanResponsibility":"选择研究价值、语义、预算与证据要求，核对新颖性和一般性。","failureHandling":"预算内未找到有效候选就交失败/卡点；评价器有误先修正再重评，不以更容易问题宣告成功。","example":{"input":"在保持精确输出下优化图算法，原算法和输入分布固定。","output":"候选代码、反例搜索、保留图性能/内存和语义论证；或预算内未改善说明。","failure":"只对公开小图缓存答案，称已改善一般复杂度。"},"sourceIds":["WU-C-EVOLVE","WU-C-EVOLVE-ART"],"evidenceLinks":[{"sourceId":"WU-C-EVOLVE","supports":"直接算法探索研究。","notCovered":"专用系统/选择性成果；不等于任意公开Agent或任意理论问题。"},{"sourceId":"WU-C-EVOLVE-ART","supports":"有可定位产物。","notCovered":"本站未运行产物。"}],"resourceRefs":[],"career":{"direction":"新算法与系统研究","evidence":"候选算法/系统、精确定义与可运行实现","nextStep":"固定一个算法问题、基线、评价器和预算，先校准错误/合法替代，再搜索候选并在保留输入重跑，独立检查理论主张。"},"team":{"bottleneck":"候选代码很多，但独立正确性/资源证据不足。","allocation":"Agent可联合搜索算法/实现/验证路径，程序或求解器给反馈；目标语义、独立评价器和预算保持固定。","nextStep":"固定一个算法问题、基线、评价器和预算，先校准错误/合法替代，再搜索候选并在保留输入重跑，独立检查理论主张。"},"claims":[{"id":"C14-ALGORITHM","statement":"AlphaEvolve白皮书报告在可自动评价的算法与计算基础设施问题上产生改进，并有部分公开产物；可把算法和验证共同纳入委派研究。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-C-EVOLVE","locator":"Abstract；evolution/evaluators；computational problems","relation":"partial","supports":"直接算法探索研究。","limit":"专用系统/选择性成果；不等于任意公开Agent或任意理论问题。"}],"configuration":"2025白皮书专用AlphaEvolve；LLM代码变异、演化与一个或多个任务评价器；不推算未报告任务分母/成本。","progressIds":["P-C-ALGORITHM"],"limitation":"专用系统/选择性成果；不等于任意公开Agent或任意理论问题。","evidenceStatus":"active"},{"id":"C14-ARTIFACT","statement":"公开AlphaEvolve成果库可作为独立检查和复现入口；不能把成果库当完整搜索系统已经开放。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-C-EVOLVE-ART","locator":"README与结果目录","relation":"partial","supports":"有可定位产物。","limit":"本站未运行产物。"}],"configuration":null,"progressIds":[],"limitation":"本站未运行产物。","evidenceStatus":"active"},{"id":"C14-ADOPT","statement":"在固定问题和独立评价器下可委派算法搜索，接收候选与反例/论证；一般性、复杂性和新颖性不能只靠样例得分。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-EVOLVE","locator":"evaluation and algorithm search","relation":"partial","supports":"专用研究展示搜索与验证联合组织。","limit":"公开Agent与新问题仍需独立任务检查。"}],"configuration":null,"progressIds":[],"limitation":"公开Agent与新问题仍需独立任务检查。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开研究成果可检查；研究系统可获得性、预算和目标问题依赖单独核实。","tools":[{"name":"AlphaEvolve公开结果","url":"https://github.com/google-deepmind/alphaevolve_results","version":"2026-10-20公开仓库HEAD 4226acbf237ff9ad10ba7673a2af127a2d8a5971；实际运行另固定依赖/镜像/模型","access":"可读成果公开；完整搜索服务/配置不据此假定可得"}]},"variants":[{"title":"理论与反例","inputDifference":"接收命题、复杂性或正确性证明，而非经验提升。","assessmentChange":"候选探索可委派，完整论证/反例才算相应结果。","acceptanceChange":"命题语义、假设、边界、逻辑依赖；若选形式化再验内核与公理。"},{"title":"经验模型与系统","inputDifference":"接收新模型、调度/数据库等系统原型。","assessmentChange":"按实际运行、资源和保留分布阶段接收。","acceptanceChange":"W-C06实验/消融与W-C08系统负载，避免只汇报单次最好。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C15","domainId":"computing","title":"模型推理服务交付：效果、资源与回滚","goal":"把固定权重和特征处理交付为可调用批/在线推理服务，满足效果、资源、故障与版本回滚条件。","trigger":"已有模型产物，需要在目标硬件与接口下提供可接收推理。","scope":"产物为固定模型/特征到批或在线推理服务；检索与工具Agent工作流另由W-C26接收，业务因果效果转W-A03。","inputs":[{"name":"推理目标","detail":"模型输入输出schema、效果切片/错误代价、并发/尾延迟/费用预算。"},{"name":"固定环境","detail":"权重/Tokenizer/预后处理、特征、依赖、硬件、数据快照与服务身份固定。"},{"name":"独立接收","detail":"离线基线/保留输入、缺失/新类别、依赖失败/负载、旧服务与回滚入口及漂移动作。"}],"deliverables":["推理服务代码/镜像和权重/特征/依赖版本","离线线上一致性、保留切片、异常输入及负载/实际资源结果","版本化流量、回滚/依赖失败/漂移响应和未支持输入说明"],"acceptance":[{"criterion":"版本与效果一致","method":"同保留输入核对权重/特征/Tokenizer/输出，检查缺失、新类别和关键切片。","failure":"服务载入旧权重却继承新训练成绩。"},{"criterion":"负载与失败可接收","method":"固定并发/规模，测尾延迟/吞吐/资源，施加资源不足、依赖超时和异常输入。","failure":"只看HTTP200或最快一次，隐藏资源耗尽。"},{"criterion":"版本可恢复","method":"按版本对比效果/资源，演练减流/回滚及旧模型-特征一致性，核验漂移响应。","failure":"canary存在但回滚后模型/特征不匹配。"}],"assessment":{"recommendation":"可委派推理服务实现和演练；以效果/版本一致性、负载与回滚阶段接收，再决定获准流量。","conditions":["权重/Tokenizer/预后处理、特征、依赖、硬件、数据快照与服务身份固定。","离线基线/保留输入、缺失/新类别、依赖失败/负载、旧服务与回滚入口及漂移动作。"],"evidenceBasis":"MLflow可记录模型/试验谱系，KServe0.20发布提供流量切分与canary机制；服务行为可以按版本单独核验；实验参数/代码/指标/产物应能连到服务版本，否则离线成绩无法解释线上差异。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["目标权重/特征、推理硬件、线上切片/漂移与真实回滚未验证；没有AI部署成功率。"],"mode":"staged-delivery","rationale":"模型工程需要把实验和运行状态接起来。公开KServe提供版本流量/观察机制，MLflow记录模型谱系；这些工具不提供AI部署成绩，采用建议是结合可验证状态作出的编辑判断。","frontier":"MLflow追踪谱系与KServe版本化部署机制公开；方法存在不等于本模型/硬件的AI交付可靠率，也不能以存活推断业务收益。","nextValidation":"固定一个模型API和旧版本，核对保留输入/特征，再测异常/并发/资源并回滚，检查旧服务真实输出。","futureSignals":["新模型/数据版本的线上线下效果、资源和回滚持续一致，可考虑扩大相近输入范围。","特征错位、漂移未响应或服务存活掩盖效果退化，应收窄或暂停相应采用。"]},"workflow":{"before":"训练交权重，工程接接口，运行团队另补版本/负载和回滚。","now":"Agent联合实现预处理、推理接口和部署验证；稳定推理由程序执行，效果/环境反馈驱动修正。","retained":["效果门槛、服务身份、生产流量与费用授权。"],"added":["版本化效果/资源与回滚后模型-特征一致性检查。"]},"humanResponsibility":"确定错误代价、服务身份、资源和流量授权，裁决效果/费用取舍。","failureHandling":"缺目标硬件/保留数据时交具体阻塞；效果或资源不达标保留失败，不改阈值隐藏。","example":{"input":"离线故障预测器部署API，需处理缺失特征、新类别和并发。","output":"同输入版本一致性、切片效果、资源/失败与回滚输出。","failure":"接口存活但特征错位，回滚也载入不匹配权重。"},"sourceIds":["WU-C-KSERVE","WU-C-MLFLOW"],"evidenceLinks":[{"sourceId":"WU-C-KSERVE","supports":"提供公开部署机制。","notCovered":"机制存在不是模型效果或AI上线可靠率。"},{"sourceId":"WU-C-MLFLOW","supports":"提供谱系/版本记录接口。","notCovered":"记录正确性和线上效果仍由任务验证。"}],"resourceRefs":[],"career":{"direction":"模型推理与机器学习服务","evidence":"推理服务代码/镜像和权重/特征/依赖版本","nextStep":"固定一个模型API和旧版本，核对保留输入/特征，再测异常/并发/资源并回滚，检查旧服务真实输出。"},"team":{"bottleneck":"模型离线表现与服务效果/资源/版本可能不一致。","allocation":"Agent联合实现预处理、推理接口和部署验证；稳定推理由程序执行，效果/环境反馈驱动修正。","nextStep":"固定一个模型API和旧版本，核对保留输入/特征，再测异常/并发/资源并回滚，检查旧服务真实输出。"},"claims":[{"id":"C15-TRACK","statement":"MLflow可记录模型/试验谱系，KServe0.20发布提供流量切分与canary机制；服务行为可以按版本单独核验。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-KSERVE","locator":"0.20 release：traffic splitting/canary","relation":"method","supports":"提供公开部署机制。","limit":"机制存在不是模型效果或AI上线可靠率。"}],"configuration":null,"progressIds":[],"limitation":"机制存在不是模型效果或AI上线可靠率。","evidenceStatus":"active"},{"id":"C15-LINEAGE","statement":"实验参数/代码/指标/产物应能连到服务版本，否则离线成绩无法解释线上差异。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-MLFLOW","locator":"Tracking Runs；Models","relation":"method","supports":"提供谱系/版本记录接口。","limit":"记录正确性和线上效果仍由任务验证。"}],"configuration":null,"progressIds":[],"limitation":"记录正确性和线上效果仍由任务验证。","evidenceStatus":"active"},{"id":"C15-ADOPT","statement":"可委派模型推理工程包，以独立效果、资源/故障和版本回滚接收；不继承训练成绩或服务存活作为业务收益。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-KSERVE","locator":"canary与versioned traffic","relation":"partial","supports":"分版本接收有工具基础。","limit":"未在目标环境运行，本任务不提供成功率或业务收益。"}],"configuration":null,"progressIds":[],"limitation":"未在目标环境运行，本任务不提供成功率或业务收益。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"MLflow/KServe机制公开；目标模型权利、硬件、依赖与部署版本另固定，未在本站运行。","tools":[{"name":"KServe","url":"https://github.com/kserve/website/blob/main/blog/2026-08-06-KServe-0.20-release.md","version":"0.20发布；具体CRD/Runtime镜像在执行固定","access":"开源；K8s/推理资源、模型权利与云权限另备"},{"name":"MLflow","url":"https://mlflow.org/docs/latest/ml/tracking/","version":"按实际服务固定MLflow/模型/存储版本","access":"开源记录工具；API和存储权限另配"}]},"variants":[{"title":"模型推理与漂移","inputDifference":"训练模型作为批/在线服务。","assessmentChange":"先核对特征/权重一致性，再接收负载和漂移响应。","acceptanceChange":"保留切片、未知类别、输入schema、延迟/成本和再训练/回退。"},{"title":"上线试点与回滚","inputDifference":"需流量切分、效果比较及失败恢复。","assessmentChange":"部署演练可整包交付，生产流量/业务判断另有授权与实验。","acceptanceChange":"版本化效果/资源、旧服务回滚与状态完整；收益转W-A03验证。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C16","domainId":"computing","title":"数据库还原：交付可读写的目标恢复点","goal":"从备份与日志交付一个经验证的恢复实例，报告实际恢复范围、时点与所需时间。","trigger":"给指定恢复点、关键事务/数据不变量、业务读写验收、RPO/RTO目标及待恢复范围。","scope":"工作对象为已备份数据库的物理/逻辑恢复；源库变更迁移、数据转换与查询调优分别另接收。","inputs":[{"name":"恢复目标","detail":"给指定恢复点、关键事务/数据不变量、业务读写验收、RPO/RTO目标及待恢复范围。"},{"name":"备份与环境","detail":"固定数据库版本、基础备份、WAL/归档链、配置/权限副本、隔离实例与存储资源；记录缺口。"},{"name":"核验与隔离","detail":"避免写回源库，保留恢复日志和目标时点事务，对关键用户读写及权限有独立探针。"}],"deliverables":["可复现还原脚本、备份/日志链与环境清单","目标点事务/约束/权限及读写检查","实际可恢复范围、时间/资源及未恢复或缺日志说明"],"acceptance":[{"criterion":"恢复到正确状态","method":"独立核对目标点前后事务、约束和关键业务查询，不只看数据库启动。","failure":"实例启动但恢复到错误时间或漏事务。"},{"criterion":"恢复链和配置真实可用","method":"检查基础备份/WAL连续性、实际还原记录及单独配置/权限；不能用pg_dump冒充物理WAL基础备份。","failure":"文件存在却日志缺段，或配置/权限未恢复。"},{"criterion":"恢复满足业务与代价","method":"在恢复实例做读写/负载探针，记录实际RPO/RTO、资源和不可恢复范围。","failure":"恢复操作完成但业务仍不可用且没有说明。"}],"assessment":{"recommendation":"可委派恢复准备与受限还原演练；只有真实目标实例、事务核对与业务探针通过才接收恢复，生产替换另须实际权限。","conditions":["固定数据库版本、基础备份、WAL/归档链、配置/权限副本、隔离实例与存储资源；记录缺口。","避免写回源库，保留恢复日志和目标时点事务，对关键用户读写及权限有独立探针。"],"evidenceBasis":"DBA-Bench显示数据库诊断、有效恢复和约束满足必须分开评价；恢复声明需要真实业务状态；基础备份、连续WAL与独立配置保护构成PostgreSQL时间点恢复的重要条件。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["本备份/WAL链、目标时点事务、配置/权限与实际RPO/RTO尚未隔离还原。"],"mode":"staged-delivery","rationale":"备份存在与恢复成功不同；PostgreSQL机制给出可核对恢复条件，DBA-Bench强调恢复目标与安全需要同时验证。","frontier":"PostgreSQL18公开提供基础备份+连续WAL的时间点恢复机制；DBA-Bench v2实测活跃PostgreSQL/TiDB的运维状态，观察诊断正确仍可能无法恢复、安全与恢复亦可能冲突。前者是方法，后者是有限场景模型结果，均不证明本备份链可恢复。","nextValidation":"在隔离实例还原一份授权备份，冻结目标事务，加入缺WAL负控制并校验目标点、配置、用户读写和实际恢复时间。","futureSignals":["不同恢复点/备份周期均可独立还原并满足业务与恢复预算，可考虑扩大相近输入范围。","WAL缺段、目标点不符或恢复后读写/配置失效，应收窄或暂停相应采用。"]},"workflow":{"before":"资料、实现/处理与验证分散交接，完成记录易脱离实际结果。","now":"Agent可核对备份/日志、生成/执行获准还原并根据日志修正；程序直接验证事务和业务状态。","retained":["确认恢复点、可丢失数据和生产替换权限，保护原始备份/取证材料。"],"added":["在恢复实例做读写/负载探针，记录实际RPO/RTO、资源和不可恢复范围。"]},"humanResponsibility":"确认恢复点、可丢失数据和生产替换权限，保护原始备份/取证材料。","failureHandling":"链不完整时准确报告最近可恢复点与缺口；保持源库不受写入，不伪造完成。","example":{"input":"需还原误删发生前一分钟，后续合法事务需要另行保存。","output":"隔离恢复实例、目标事务核对、差异和恢复时间。","failure":"只恢复到最后备份，误删和合法变更边界不明。"},"sourceIds":["WU-C-DBA","WU-C-PITR"],"evidenceLinks":[{"sourceId":"WU-C-DBA","supports":"直接状态运维研究。","notCovered":"论文场景不是本站备份恢复试次。"},{"sourceId":"WU-C-PITR","supports":"支持恢复方案和结果核对。","notCovered":"原生机制不是AI成绩。"}],"resourceRefs":[],"career":{"direction":"数据库还原：交付可读写的目标恢复点","evidence":"可复现还原脚本、备份/日志链与环境清单","nextStep":"在隔离实例还原一份授权备份，冻结目标事务，加入缺WAL负控制并校验目标点、配置、用户读写和实际恢复时间。"},"team":{"bottleneck":"备份存在与恢复成功不同；PostgreSQL机制给出可核对恢复条件，DBA-Bench强调恢复目标与安全需要同时验证。","allocation":"Agent可核对备份/日志、生成/执行获准还原并根据日志修正；程序直接验证事务和业务状态。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C16-STATE","statement":"DBA-Bench显示数据库诊断、有效恢复和约束满足必须分开评价；恢复声明需要真实业务状态。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-C-DBA","locator":"v2 bounded/persistent evaluation","relation":"partial","supports":"直接状态运维研究。","limit":"论文场景不是本站备份恢复试次。"}],"configuration":"v2：231有界场景、10持续场景，PostgreSQL/TiDB；8自动基线；按原环境/任务评估，本站未重放。","progressIds":["P-C-DBA-STATE"],"limitation":"论文场景不是本站备份恢复试次。","evidenceStatus":"active"},{"id":"C16-PITR","statement":"基础备份、连续WAL与独立配置保护构成PostgreSQL时间点恢复的重要条件。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-PITR","locator":"25.3.1/25.3.5/Caveats","relation":"method","supports":"支持恢复方案和结果核对。","limit":"原生机制不是AI成绩。"}],"configuration":null,"progressIds":[],"limitation":"原生机制不是AI成绩。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"PostgreSQL原生恢复工具公开；实际备份、日志和恢复资源由任务提供。","tools":[{"name":"PostgreSQL PITR","url":"https://www.postgresql.org/docs/18/continuous-archiving.html","version":"文档18；执行锁定源/目标DB与备份格式","access":"隔离恢复实例、授权备份/日志和存储"}]},"variants":[{"title":"日志链与物理恢复","inputDifference":"目标点依赖基础备份与归档。","assessmentChange":"先验证连续性再承诺可恢复范围。","acceptanceChange":"缺段、时间线和目标点事务，配置另核。"},{"title":"逻辑恢复与业务重建","inputDifference":"只恢复选定对象/逻辑数据，不能套用物理PITR。","assessmentChange":"按目标对象与依赖单独设计。","acceptanceChange":"对象/约束/权限、增量差异及完整业务查询。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C17","domainId":"computing","title":"数据与schema切换：兼容迁移及可回退状态","goal":"交付从旧系统/schema到目标版本的数据迁移、回填和切换证据，保持旧新读写兼容及失败时数据可接续。","trigger":"给源/目标版本和映射、调用方/消费者、迁移目标、数据保留与不允许丢失的事务。","scope":"工作对象是版本/系统切换，不是日常管道刷新或历史备份还原；可以按兼容/回填/切换多阶段接收同一迁移。","inputs":[{"name":"切换契约","detail":"给源/目标版本和映射、调用方/消费者、迁移目标、数据保留与不允许丢失的事务。"},{"name":"并存环境","detail":"固定旧新应用/数据库、代表数据/负载、并发写入、回填预算及实际权限。"},{"name":"核对与退出条件","detail":"准备键/金额/约束对账、兼容矩阵和回滚/向前修复路径；明确不可逆点和停机边界。"}],"deliverables":["兼容schema/数据映射、迁移/回填和切换代码","旧新版本/并发写与断点下的数据对账结果","切换判据、实际停顿/资源、回滚或向前修复演练"],"acceptance":[{"criterion":"语义迁移正确","method":"按业务键、金额、空值/编码/时区及约束独立核对映射。","failure":"行数相同但金额精度/时区或身份映射改变。"},{"criterion":"并存与中断可接续","method":"重放旧新客户端、并发写/回填中断和重复执行。","failure":"旧客户端无法写或新数据被回填覆盖。"},{"criterion":"切换和退出真可执行","method":"演练流量/写入切换、未完成检测与回退/向前修复，核对业务不变量。","failure":"文档有rollback按钮，实际新写入无法恢复。"}],"assessment":{"recommendation":"可委派兼容方案、迁移实现和演练；按旧新并存、回填及切换结果阶段接收，先揭示不可逆点再决定执行。","conditions":["固定旧新应用/数据库、代表数据/负载、并发写入、回填预算及实际权限。","准备键/金额/约束对账、兼容矩阵和回滚/向前修复路径；明确不可逆点和停机边界。"],"evidenceBasis":"Spider2-DBT提供可检查的仓库级数据转换任务，DBA-Bench强调业务状态与约束同时核验；旧新并存/切换建议为基于这些邻近任务的编辑迁移，不冒称迁移实测。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["旧新并存/持续写入、语义映射、不可逆点和退出路径尚无目标迁移实测。"],"mode":"staged-delivery","rationale":"仓库转换和数据库状态评测提供邻近依据；跨版本迁移的关键是持续写入和退出策略，需要明确标为任务适配而非假装有直接迁移成功率。","frontier":"Spider2-DBT已经公开仓库级数据转换任务；DBA-Bench进一步以活跃数据库目标/约束检查状态。这些是迁移的邻近基础，当前未核实涵盖本旧新并存/切换任务的直接AI实验，因此本单元采用建议为方法推断。","nextValidation":"冻结一份旧库/调用方快照，用重复执行、并发写与中断测试迁移；独立对账并演练退出或向前修复。","futureSignals":["旧新并存、持续写和中断回填下语义对账与退出都可复验，可考虑扩大相近输入范围。","回填覆盖新数据、旧客户端失败或不可逆点未被识别，应收窄或暂停相应采用。"]},"workflow":{"before":"资料、实现/处理与验证分散交接，完成记录易脱离实际结果。","now":"Agent可生成兼容schema、回填与切换脚本，并按对账/失败反馈修改；稳定批处理直接程序执行。","retained":["确定语义映射、停机/丢失容忍和不可逆点授权，维护调用方兼容责任。"],"added":["演练流量/写入切换、未完成检测与回退/向前修复，核对业务不变量。"]},"humanResponsibility":"确定语义映射、停机/丢失容忍和不可逆点授权，维护调用方兼容责任。","failureHandling":"对账或兼容失败就保留旧写路径；没有可行回退时明确向前修复前提，不能虚构可逆。","example":{"input":"账户ID改型，旧客户端与新服务并存且要持续写入。","output":"兼容字段、回填、旧新客户端与中断对账、切换/退出结果。","failure":"先删旧字段通过新测试，旧客户端无法访问。"},"sourceIds":["WU-C-SPIDER","WU-C-DBA"],"evidenceLinks":[{"sourceId":"WU-C-SPIDER","supports":"数据转换的工程任务作为邻近方法基础。","notCovered":"没有目标跨版本迁移直接AI结果；所有切换控制是本任务设计。"},{"sourceId":"WU-C-DBA","supports":"状态型评价支持独立迁移验收。","notCovered":"数据库运维研究不提供本迁移的成功率。"}],"resourceRefs":[],"career":{"direction":"数据与schema切换：兼容迁移及可回退状态","evidence":"兼容schema/数据映射、迁移/回填和切换代码","nextStep":"冻结一份旧库/调用方快照，用重复执行、并发写与中断测试迁移；独立对账并演练退出或向前修复。"},"team":{"bottleneck":"仓库转换和数据库状态评测提供邻近依据；跨版本迁移的关键是持续写入和退出策略，需要明确标为任务适配而非假装有直接迁移成功率。","allocation":"Agent可生成兼容schema、回填与切换脚本，并按对账/失败反馈修改；稳定批处理直接程序执行。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C17-MIGRATE","statement":"Agent可生成并校验迁移候选，但接收需独立检验语义、旧新并存及中断退出，不能从普通SQL任务推出迁移已可靠。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-SPIDER","locator":"Spider2-DBT repository workflows","relation":"partial","supports":"数据转换的工程任务作为邻近方法基础。","limit":"没有目标跨版本迁移直接AI结果；所有切换控制是本任务设计。"}],"configuration":null,"progressIds":[],"limitation":"没有目标跨版本迁移直接AI结果；所有切换控制是本任务设计。","evidenceStatus":"active"},{"id":"C17-STATE","statement":"数据库状态评价应同时检查业务目标和安全约束，不能只以脚本退出码接收切换。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-DBA","locator":"operational objectives and safety","relation":"partial","supports":"状态型评价支持独立迁移验收。","limit":"数据库运维研究不提供本迁移的成功率。"}],"configuration":null,"progressIds":[],"limitation":"数据库运维研究不提供本迁移的成功率。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"SQL/数据库原生迁移可执行；目标工具/引擎版本与退出路径需实际冻结，尚无本任务模型试验。","tools":[{"name":"PostgreSQL 原生SQL与备份/恢复","url":"https://www.postgresql.org/docs/18/continuous-archiving.html","version":"以实际源/目标DB和迁移脚本版本为准；备份恢复文档18仅作退出准备","access":"获准旧新实例、数据/备份与兼容客户端"}]},"variants":[{"title":"在线扩展与收缩","inputDifference":"旧新schema并存、不能停写。","assessmentChange":"按兼容/回填/切换接收，Agent内部方法自由。","acceptanceChange":"旧新读写、并发和水位，最后才验收旧结构清理。"},{"title":"停机或跨系统切换","inputDifference":"允许冻结写入或目标引擎不同。","assessmentChange":"先校验可转换范围与实际停顿预算。","acceptanceChange":"编码/类型/依赖、最终增量、流量切换与退出。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C18","domainId":"computing","title":"性能与容量改善：同一负载下证明收益与代价","goal":"交付可重复测得的系统性能/容量改善，保留正确性、资源成本和退化边界。","trigger":"固定目标查询/请求/数据规模、延迟/吞吐/容量目标及不能牺牲的正确性和成本约束。","scope":"工作对象是固定负载与约束下的性能变化；数据库、服务与云网络环境可分变体，但优化目标和独立前后测量不变。","inputs":[{"name":"性能问题","detail":"固定目标查询/请求/数据规模、延迟/吞吐/容量目标及不能牺牲的正确性和成本约束。"},{"name":"可比环境","detail":"冻结硬件/DB/服务/网络配置、数据和代表负载、缓存冷热、并发与测量窗。"},{"name":"独立测量","detail":"准备基线/候选、资源/成本记录和正确性回归；允许保留未改善或只特定条件改善的结果。"}],"deliverables":["瓶颈证据与可应用的代码/索引/配置候选","同条件前后性能、资源/费用与正确性结果","适用负载/退化边界、回退和容量建议"],"acceptance":[{"criterion":"对比条件可比","method":"多次重放同数据/并发/缓存条件并报告波动和热身。","failure":"只给最快一次或拿冷基线比热候选。"},{"criterion":"改善没有偷换目标","method":"核对正确性/一致性/权限及其他请求，测写放大、内存和费用。","failure":"删结果、跳校验或昂贵扩容却只报延迟下降。"},{"criterion":"边界可复用","method":"改变负载/数据规模/失败条件复查，说明瓶颈转移及回退。","failure":"单条SQL变快但整服务锁争用和成本恶化。"}],"assessment":{"recommendation":"可委派瓶颈定位、候选和基准循环；按同条件收益/代价和正确性接收，未改善也是可用诊断。","conditions":["冻结硬件/DB/服务/网络配置、数据和代表负载、缓存冷热、并发与测量窗。","准备基线/候选、资源/成本记录和正确性回归；允许保留未改善或只特定条件改善的结果。"],"evidenceBasis":"DBA-Bench在活跃负载中区分业务目标恢复与违规执行，支持把性能结果与约束一起判断；可用AIOpsLab负载和轨迹接口构造受控前后比较；实际测量与正确性应独立于优化者。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["目标负载的可比测量、混合读写/尾延迟/费用及新规模退化尚未校准。"],"mode":"staged-delivery","rationale":"数据库有负载运维研究和AIOps负载工具提供邻近依据；性能交付可由独立程序比较，Agent主要搜索原因/配置，不能用一次最好成绩证明容量。","frontier":"DBA-Bench v2把真实负载和数据库状态放入评估，AIOpsLab则提供可控制负载/轨迹接口；前者报告任务状态差距，后者是工程工具。它们未展示本目标负载下的AI优化收益，须同条件前后测量。","nextValidation":"选一个可重放慢查询/服务压力，校准负载和测量；冻结预算，让Agent交候选，多次测性能、正确性和资源，改变负载挑战结论。","futureSignals":["保留负载下多次收益超过测量波动且资源代价符合预算，可考虑扩大相近输入范围。","收益仅来自缓存/更易输入，或写入/正确性与费用退化，应收窄或暂停相应采用。"]},"workflow":{"before":"资料、实现/处理与验证分散交接，完成记录易脱离实际结果。","now":"Agent自主分析profile/执行计划、修改代码/配置并重测；程序执行冻结基准与资源记录。","retained":["确定服务目标、成本与一致性取舍，批准涉及容量/费用的实际变更。"],"added":["改变负载/数据规模/失败条件复查，说明瓶颈转移及回退。"]},"humanResponsibility":"确定服务目标、成本与一致性取舍，批准涉及容量/费用的实际变更。","failureHandling":"环境不可比时先修测量；无改善保留profile和失败，不篡改阈值/采样。","example":{"input":"查询随订单量变慢，候选增加索引。","output":"读写混合负载前后延迟、锁、存储/写放大与金额结果。","failure":"只测读请求，忽略索引让写入超时。"},"sourceIds":["WU-C-DBA","WU-C-AIOPS"],"evidenceLinks":[{"sourceId":"WU-C-DBA","supports":"直接DB状态运维研究作为邻近证据。","notCovered":"不是本站性能优化实测，也不是跨云/网络提升率。"},{"sourceId":"WU-C-AIOPS","supports":"有负载与记录接口。","notCovered":"框架不保证选对工作负载或性能改善。"}],"resourceRefs":[],"career":{"direction":"性能与容量改善：同一负载下证明收益与代价","evidence":"瓶颈证据与可应用的代码/索引/配置候选","nextStep":"选一个可重放慢查询/服务压力，校准负载和测量；冻结预算，让Agent交候选，多次测性能、正确性和资源，改变负载挑战结论。"},"team":{"bottleneck":"数据库有负载运维研究和AIOps负载工具提供邻近依据；性能交付可由独立程序比较，Agent主要搜索原因/配置，不能用一次最好成绩证明容量。","allocation":"Agent自主分析profile/执行计划、修改代码/配置并重测；程序执行冻结基准与资源记录。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C18-STATE","statement":"DBA-Bench在活跃负载中区分业务目标恢复与违规执行，支持把性能结果与约束一起判断。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-C-DBA","locator":"v2 live workloads and operational objectives","relation":"partial","supports":"直接DB状态运维研究作为邻近证据。","limit":"不是本站性能优化实测，也不是跨云/网络提升率。"}],"configuration":"DBA-Bench v2：活跃PostgreSQL/TiDB，231有界场景与10持续场景，8自动基线；按论文具体模型/脚手架和约束解释，不是本站性能优化试验。","progressIds":["P-C-DBA-STATE"],"limitation":"不是本站性能优化实测，也不是跨云/网络提升率。","evidenceStatus":"active"},{"id":"C18-MEASURE","statement":"可用AIOpsLab负载和轨迹接口构造受控前后比较；实际测量与正确性应独立于优化者。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-AIOPS","locator":"Workload/Evaluator/trace","relation":"method","supports":"有负载与记录接口。","limit":"框架不保证选对工作负载或性能改善。"}],"configuration":null,"progressIds":[],"limitation":"框架不保证选对工作负载或性能改善。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"框架公开；目标云权限、遥测与可回退环境由实际任务提供，不要求先接生产。","tools":[{"name":"AIOpsLab","url":"https://github.com/microsoft/AIOpsLab","version":"2026-10-20公开仓库HEAD ccf08d0d1d5fa5b30f120e2e8549662d44411b35；实际运行另固定依赖/镜像/模型","access":"开源；独立集群/命名空间、遥测与模型API自备"}]},"variants":[{"title":"数据库查询与争用","inputDifference":"索引、计划、锁、存储与混合读写。","assessmentChange":"先冻结混合负载，不以单条查询接收整库。","acceptanceChange":"计划/profile、锁、写放大、数据结果与故障回退。"},{"title":"服务与网络容量","inputDifference":"连接池、缓存、网络路径、扩容或资源配置。","assessmentChange":"明确用户路径和费用预算后迭代。","acceptanceChange":"吞吐/尾延迟、正确性、故障域、费用与相邻路径。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C19","domainId":"computing","title":"基础设施配置交付：可重建、可访问、可拒绝","goal":"交付可重建的云/网络/系统配置，使指定服务连通、隔离和故障域约束在实际环境成立。","trigger":"给服务/网络/账户边界、DNS/证书/路由、容量/故障域约束和必须拒绝的连接。","scope":"工作对象是期望基础设施状态与配置差异；事故抢修、性能优化和数据切换分别由W-C08/W-C18/W-C17接收。","inputs":[{"name":"目标拓扑与访问","detail":"给服务/网络/账户边界、DNS/证书/路由、容量/故障域约束和必须拒绝的连接。"},{"name":"可部署环境","detail":"固定云/集群/网络版本、账户/namespace、IaC提供器、配额、费用上限与可逆部署范围。"},{"name":"接收矩阵","detail":"预先准备可连/不可连、身份、重建/漂移和故障切换检查，保留旧配置。"}],"deliverables":["IaC/网络/系统配置与依赖、拓扑和费用假设","实际部署/重建与访问正反矩阵、故障域检查","漂移/监控/回退及未满足配额/外部依赖"],"acceptance":[{"criterion":"期望状态真实成立","method":"部署后独立访问端点、DNS/证书/路由与身份，不只检查计划文件。","failure":"IaC plan成功却服务不可访问。"},{"criterion":"隔离没有被打开","method":"从未授权身份/网络做拒绝控制，检查最小权限、密钥和暴露面。","failure":"把规则改为全允许来通过连通检查。"},{"criterion":"状态可维护可回退","method":"销毁/重建演练限于隔离资源，检测漂移和一个故障域失效，核对回退。","failure":"手工改资源未入代码，重建时丢失配置。"}],"assessment":{"recommendation":"在隔离配额/权限与正反矩阵齐备时可整包委派配置实现和重建验证；稳定配置直接程序部署。","conditions":["固定云/集群/网络版本、账户/namespace、IaC提供器、配额、费用上限与可逆部署范围。","预先准备可连/不可连、身份、重建/漂移和故障切换检查，保留旧配置。"],"evidenceBasis":"AIOpsLab公开配置故障/真实K8s试验床，可把服务状态与配置联系起来检查。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["目标云/网络的重建、可连/拒绝矩阵、故障域与实际费用尚未验证。"],"mode":"staged-delivery","rationale":"AIOpsLab展示真实微服务与配置故障可检查，提供邻近环境方法；计划基础设施的成功应由目标状态和隔离检查决定。","frontier":"AIOpsLab公开可部署微服务/K8s环境及配置故障注入、负载和结果检查，已经能把配置与服务状态连起来；这是可检查实现范围，不是目标云IaC交付、连通隔离或费用上的模型成绩。","nextValidation":"给一份隔离拓扑与访问矩阵，让Agent交IaC，独立重建并检查正反访问、漂移、配额/成本和故障域。","futureSignals":["新环境重建、访问拒绝矩阵与故障域检查保持一致，可考虑扩大相近输入范围。","只能手工修到可用，或开放权限/隐藏费用换取连通，应收窄或暂停相应采用。"]},"workflow":{"before":"资料、实现/处理与验证分散交接，完成记录易脱离实际结果。","now":"Agent可生成配置、比较实际状态并迭代；部署器直接执行，独立探针验证连通和拒绝。","retained":["确认云费用、暴露边界、账户授权和故障域需求，批准影响共享环境的配置。"],"added":["销毁/重建演练限于隔离资源，检测漂移和一个故障域失效，核对回退。"]},"humanResponsibility":"确认云费用、暴露边界、账户授权和故障域需求，批准影响共享环境的配置。","failureHandling":"缺配额/身份/外部证书时明确阻塞与可部署部分，不能以扩大权限或隐藏失败完成交付。","example":{"input":"新服务仅允许应用子网访问，需跨区故障切换。","output":"可重建配置、两类身份访问/拒绝、DNS/证书与切换结果。","failure":"从自己账户可访问就声称所有隔离规则正确。"},"sourceIds":["WU-C-AIOPS"],"evidenceLinks":[{"sourceId":"WU-C-AIOPS","supports":"提供环境和故障资产。","notCovered":"事故试验床不是目标云IaC成绩。"}],"resourceRefs":[],"career":{"direction":"基础设施配置交付：可重建、可访问、可拒绝","evidence":"IaC/网络/系统配置与依赖、拓扑和费用假设","nextStep":"给一份隔离拓扑与访问矩阵，让Agent交IaC，独立重建并检查正反访问、漂移、配额/成本和故障域。"},"team":{"bottleneck":"AIOpsLab展示真实微服务与配置故障可检查，提供邻近环境方法；计划基础设施的成功应由目标状态和隔离检查决定。","allocation":"Agent可生成配置、比较实际状态并迭代；部署器直接执行，独立探针验证连通和拒绝。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C19-CONFIG","statement":"AIOpsLab公开配置故障/真实K8s试验床，可把服务状态与配置联系起来检查。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-C-AIOPS","locator":"Fault injectors/Problems","relation":"partial","supports":"提供环境和故障资产。","limit":"事故试验床不是目标云IaC成绩。"}],"configuration":null,"progressIds":[],"limitation":"事故试验床不是目标云IaC成绩。","evidenceStatus":"active"},{"id":"C19-ADOPT","statement":"对基础设施可以委派实现和重建反馈，接收需同时验证连通、拒绝、故障域和实际资源。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-AIOPS","locator":"Fault/Workload/Evaluator","relation":"partial","supports":"受控状态验证迁移为配置接收。","limit":"云服务/网络差异需本地正反矩阵验证。"}],"configuration":null,"progressIds":[],"limitation":"云服务/网络差异需本地正反矩阵验证。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"框架公开；目标云权限、遥测与可回退环境由实际任务提供，不要求先接生产。","tools":[{"name":"AIOpsLab","url":"https://github.com/microsoft/AIOpsLab","version":"2026-10-20公开仓库HEAD ccf08d0d1d5fa5b30f120e2e8549662d44411b35；实际运行另固定依赖/镜像/模型","access":"开源；独立集群/命名空间、遥测与模型API自备"}]},"variants":[{"title":"网络与访问配置","inputDifference":"路由、DNS、证书、策略或身份。","assessmentChange":"先冻结连通/拒绝矩阵再实施。","acceptanceChange":"从不同身份/子网测试解析、端口、证书和隔离。"},{"title":"平台部署与迁移","inputDifference":"新建/重建资源或平台版本切换。","assessmentChange":"按配额/兼容/故障域阶段接收；数据切换另见W-C17。","acceptanceChange":"重建、配置漂移、故障切换和实际费用。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C20","domainId":"computing","title":"运行信号交付：SLI/SLO、检测与可行动告警","goal":"交付能反映指定用户目标的运行指标、目标与告警，使已知故障被检测而正常波动不触发无效操作。","trigger":"给关键用户路径、可接受失败/时延与取舍，明确SLI分母、时间窗、依赖和错误预算用途。","scope":"产物是信号定义、采集/规则与正反例校准，不是故障修复或容量提升；业务目标与错误预算取舍需有实际依据。","inputs":[{"name":"用户目标","detail":"给关键用户路径、可接受失败/时延与取舍，明确SLI分母、时间窗、依赖和错误预算用途。"},{"name":"遥测与历史","detail":"提供日志/指标/trace、正常和已知故障窗口、采样/缺失条件及采集权限。"},{"name":"告警行动","detail":"每个告警对应可核查症状、责任/升级与处置入口；冻结误报/漏报和成本容忍。"}],"deliverables":["指标/目标/错误预算定义和可执行采集/规则","正常/故障/采集缺失/抖动控制结果","行动手册、阈值理由、观测缺口和版本管理"],"acceptance":[{"criterion":"信号代表用户目标","method":"把分母、窗口和路径与真实请求对照，检查依赖/采样遗漏。","failure":"只测机器存活，用户付款失败却SLI健康。"},{"criterion":"规则可检测可拒绝","method":"重放已知故障与正常波动，记录误报/漏报、检测延迟和采集失败。","failure":"只加一条能报警的规则，没有正常负样本。"},{"criterion":"告警可行动","method":"另一接收者按规则和指引定位受影响路径，核对升级/静默/恢复及版本。","failure":"告警没有对象、含义或动作，导致反复无效值班。"}],"assessment":{"recommendation":"可委派指标/规则和正反例校准；先核对用户口径，再让程序持续测量，异常可交Agent解释与定位。","conditions":["提供日志/指标/trace、正常和已知故障窗口、采样/缺失条件及采集权限。","每个告警对应可核查症状、责任/升级与处置入口；冻结误报/漏报和成本容忍。"],"evidenceBasis":"AIOpsLab的故障、负载和评估接口可用于校准运行信号与检测，规则正确性应以独立故障/正常控制检查。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["用户SLI分母/缺测、真实新故障与长期误报/漏报分布尚需校准。"],"mode":"staged-delivery","rationale":"公开故障/负载试验床能校准信号，但无法替组织确定风险容忍。把指标和告警本身当产物，可避免Agent靠“监控正常”自证。","frontier":"AIOpsLab提供检测任务、故障/负载和评估接口，可用正常与故障重放验证告警；公开接口没有替本团队选择用户SLI/SLO，也未给本规则长期漏报/误报成绩。","nextValidation":"选一个关键用户路径和历史故障，冻结口径，用正常/故障/遥测缺失控制验证规则，再在观察窗记录行动与噪声。","futureSignals":["真实新故障被及时检测而正常波动与缺测被正确区分，可考虑扩大相近输入范围。","用户失败未进分母、告警噪声升高或采集缺失被写成健康，应收窄或暂停相应采用。"]},"workflow":{"before":"资料、实现/处理与验证分散交接，完成记录易脱离实际结果。","now":"Agent读服务定义和历史，生成/调整采集和告警，程序重放控制；负责人决定SLO和升级取舍。","retained":["设定用户风险与错误预算政策，确认告警责任和可采取动作。"],"added":["另一接收者按规则和指引定位受影响路径，核对升级/静默/恢复及版本。"]},"humanResponsibility":"设定用户风险与错误预算政策，确认告警责任和可采取动作。","failureHandling":"目标不明确先给可选口径及影响；采集缺失应暴露为缺测，不能写成零错误。","example":{"input":"付款成功率看起来100%，但超时请求没进分母。","output":"修正SLI、超时/正常/日志缺失控制与升级说明。","failure":"删去超时事件，让仪表板变绿。"},"sourceIds":["WU-C-AIOPS"],"evidenceLinks":[{"sourceId":"WU-C-AIOPS","supports":"提供可构造症状与结果的环境。","notCovered":"方法不证明目标SLO恰当或长期告警可靠。"}],"resourceRefs":[],"career":{"direction":"运行信号交付：SLI/SLO、检测与可行动告警","evidence":"指标/目标/错误预算定义和可执行采集/规则","nextStep":"选一个关键用户路径和历史故障，冻结口径，用正常/故障/遥测缺失控制验证规则，再在观察窗记录行动与噪声。"},"team":{"bottleneck":"公开故障/负载试验床能校准信号，但无法替组织确定风险容忍。把指标和告警本身当产物，可避免Agent靠“监控正常”自证。","allocation":"Agent读服务定义和历史，生成/调整采集和告警，程序重放控制；负责人决定SLO和升级取舍。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C20-SIGNALS","statement":"AIOpsLab的故障、负载和评估接口可用于校准运行信号与检测，规则正确性应以独立故障/正常控制检查。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-AIOPS","locator":"Detection tasks/Workload/Evaluator","relation":"method","supports":"提供可构造症状与结果的环境。","limit":"方法不证明目标SLO恰当或长期告警可靠。"}],"configuration":null,"progressIds":[],"limitation":"方法不证明目标SLO恰当或长期告警可靠。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"框架公开；目标云权限、遥测与可回退环境由实际任务提供，不要求先接生产。","tools":[{"name":"AIOpsLab","url":"https://github.com/microsoft/AIOpsLab","version":"2026-10-20公开仓库HEAD ccf08d0d1d5fa5b30f120e2e8549662d44411b35；实际运行另固定依赖/镜像/模型","access":"开源；独立集群/命名空间、遥测与模型API自备"}]},"variants":[{"title":"用户SLI与目标","inputDifference":"需定义用户成功、延迟与错误预算。","assessmentChange":"先对照真实路径再自动化采集。","acceptanceChange":"分母/窗口/采样、依赖、目标取舍和缺测。"},{"title":"检测规则与值班噪声","inputDifference":"已有遥测但告警漏报或过多。","assessmentChange":"按可行动症状重做规则校准。","acceptanceChange":"故障/正常/抖动/采集缺失与升级路径。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C21","domainId":"computing","title":"事故复盘与整改闭环：结论可追溯、行动可验证","goal":"交付从实际事件证据到原因/促成条件、改进行动及验证状态的复盘包，区分已证实原因、假设与未完成整改。","trigger":"给事故影响、时间窗、用户路径、操作和恢复记录及需要回答的复发/值班负担问题。","scope":"产物是可检验复盘与行动闭环；不能把生成根因故事当已恢复服务，整改代码/配置另由相应任务接收。","inputs":[{"name":"事件目标","detail":"给事故影响、时间窗、用户路径、操作和恢复记录及需要回答的复发/值班负担问题。"},{"name":"有身份的证据","detail":"固定日志/trace/工单/配置版本、谁何时做了什么、缺失/矛盾以及真实负责人反馈。"},{"name":"闭环标准","detail":"每条行动定义要改变的可测结果、责任/期限、验证材料和复发/运行负担观察窗。"}],"deliverables":["可追溯时间线、影响与证据/假设分离的原因分析","候选原因的反证或重放检查、促成条件和未解决矛盾","有结果判据的整改/运行负担清单与真实完成/未完成证据"],"acceptance":[{"criterion":"原因有因果链而非联想","method":"把每项关键推论连到事件与状态；可行时重放/移除条件挑战，保留不能证实部分。","failure":"最近一次发布就被写成根因。"},{"criterion":"行动解决对应问题","method":"逐项核对行动怎样防止/发现/减轻故障，有可验收结果而非“加强意识”。","failure":"只加一场培训，把系统性故障视作已整改。"},{"criterion":"完成状态真实","method":"读取代码/配置/试验/工单验收证据及复发/负担记录，区分计划、实施和验证。","failure":"工单关闭就推断措施有效或事故不再发生。"}],"assessment":{"recommendation":"可委派证据整合、反证设计和整改跟踪；真实原因与完成性按证据核对，AI不能补造同事反馈。","conditions":["固定日志/trace/工单/配置版本、谁何时做了什么、缺失/矛盾以及真实负责人反馈。","每条行动定义要改变的可测结果、责任/期限、验证材料和复发/运行负担观察窗。"],"evidenceBasis":"AIOpsLab公开动作/遥测/结果轨迹，可支撑事件材料和反证的组织方法；真实复盘/整改闭环建议是编辑推断，效果须由实际复验与观察确认。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["真实时间线的证据冲突、原因反证、整改验证及复发/值班负担尚需实际记录。"],"mode":"staged-delivery","rationale":"AIOpsLab记录动作与状态，支持可回放事件方法；从受控轨迹迁移到真实复盘是编辑判断。复盘结果应使下一次验证可执行，而非只是更流畅叙事。","frontier":"公开事故框架记录实际动作、遥测与结果，使事件时间线及恢复假设可以重放检查；把它迁移为复盘/行动材料是编辑方法判断，目前未核实AI减少本团队复发或值班负担的直接研究。","nextValidation":"选一个日志完整事故，把一个错误根因与未实施行动混入材料，检查Agent是否保留证据冲突并把行动连到真实验证。","futureSignals":["整改的独立复验与后续观察证实复发/运行负担降低，可考虑扩大相近输入范围。","根因无法反证、行动只关单未验证或同类事故复发，应收窄或暂停相应采用。"]},"workflow":{"before":"资料、实现/处理与验证分散交接，完成记录易脱离实际结果。","now":"Agent自主合并时间线、提出反证与行动，查实际工单/补丁状态并更新；执行整改另交工程任务。","retained":["确认影响、事实争议、行动责任及残余风险；真实沟通与反馈由实际参与者给出。"],"added":["读取代码/配置/试验/工单验收证据及复发/负担记录，区分计划、实施和验证。"]},"humanResponsibility":"确认影响、事实争议、行动责任及残余风险；真实沟通与反馈由实际参与者给出。","failureHandling":"证据缺失保留多假设和采集改进，不造单一原因；行动未验证就明确标注未完成。","example":{"input":"服务已恢复，怀疑最近发布但同时发生依赖超时。","output":"两假设的证据/反证、真实影响、整改验收与责任期限。","failure":"按最像的故事归责个人，关闭工单后宣布复发问题解决。"},"sourceIds":["WU-C-AIOPS"],"evidenceLinks":[{"sourceId":"WU-C-AIOPS","supports":"轨迹和结果关联提供方法基础。","notCovered":"没有目标真实复盘的AI成绩，不把模拟轨迹当组织事实。"}],"resourceRefs":[],"career":{"direction":"事故复盘与整改闭环：结论可追溯、行动可验证","evidence":"可追溯时间线、影响与证据/假设分离的原因分析","nextStep":"选一个日志完整事故，把一个错误根因与未实施行动混入材料，检查Agent是否保留证据冲突并把行动连到真实验证。"},"team":{"bottleneck":"AIOpsLab记录动作与状态，支持可回放事件方法；从受控轨迹迁移到真实复盘是编辑判断。复盘结果应使下一次验证可执行，而非只是更流畅叙事。","allocation":"Agent自主合并时间线、提出反证与行动，查实际工单/补丁状态并更新；执行整改另交工程任务。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C21-POSTMORTEM","statement":"可把公开事故试验床的动作/状态轨迹方法迁移为复盘材料，但原因和整改效果须由真实证据及反证验证。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-AIOPS","locator":"Session trace/results/evaluator","relation":"partial","supports":"轨迹和结果关联提供方法基础。","limit":"没有目标真实复盘的AI成绩，不把模拟轨迹当组织事实。"}],"configuration":null,"progressIds":[],"limitation":"没有目标真实复盘的AI成绩，不把模拟轨迹当组织事实。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"框架公开；目标云权限、遥测与可回退环境由实际任务提供，不要求先接生产。","tools":[{"name":"AIOpsLab","url":"https://github.com/microsoft/AIOpsLab","version":"2026-10-20公开仓库HEAD ccf08d0d1d5fa5b30f120e2e8549662d44411b35；实际运行另固定依赖/镜像/模型","access":"开源；独立集群/命名空间、遥测与模型API自备"}]},"variants":[{"title":"原因与促成条件","inputDifference":"恢复完成但成因/触发存在矛盾。","assessmentChange":"交假设、证据和最小反证，不强行单根因。","acceptanceChange":"时间一致、状态链、替代解释与重放结果。"},{"title":"整改与运行负担","inputDifference":"需要跟踪复发、值班和配置漂移措施。","assessmentChange":"按真实结果记录实施/验证，不自动关单。","acceptanceChange":"行动判据、版本/负载复验、复发和工作负担观察。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C22","domainId":"computing","title":"项目交付基线与状态重排：依赖、资源和验收差距","goal":"交付依据真实状态的范围/里程碑/资源基线或重排方案，并记录风险、变更、决定和实际验收/移交差距。","trigger":"给实际范围、里程碑、预算、资源、验收目标及已有承诺；明确不能改变的约束。","scope":"产物为可执行交付计划与可追溯状态；AI不能以模拟同事沟通伪造承诺，系统技术方案由W-C11另接收。","inputs":[{"name":"交付目标与现有基线","detail":"给实际范围、里程碑、预算、资源、验收目标及已有承诺；明确不能改变的约束。"},{"name":"状态证据","detail":"冻结工单/版本/验收、依赖、资源日历、费用与更新日期；每个进度分开开发、部署、验收。"},{"name":"决定权限","detail":"明确谁可批准范围/预算/时间承诺与对外发送；缺真实确认则交待确认影响，不静默承诺。"}],"deliverables":["依赖/资源可执行的基线与至少一种冲突/风险备选","状态/风险/范围变更影响、真实决定和待确认责任期限","交付验收差距、运营移交及证据到承诺的记录"],"acceptance":[{"criterion":"状态可追溯","method":"关键进度连到实际工单/产物/验收记录并标时效；冲突并列。","failure":"一条提交即被写成客户验收完成。"},{"criterion":"重排可执行","method":"程序检查依赖、日历/资源冲突与成本，列出范围/时间/风险取舍。","failure":"同一人同时承担互斥工作，依赖未就绪仍承诺日期。"},{"criterion":"验收与决定真实","method":"记录实际批准/未确认、范围前后与验收差距；移交需有运营任务可接续。","failure":"把会议摘要建议写成各方已同意或把关单当完成。"}],"assessment":{"recommendation":"可整包委派状态核对、计划计算和备选生成；实际承诺、预算和冲突取舍由对应授权接收，闭环以验收/移交事实检查。","conditions":["冻结工单/版本/验收、依赖、资源日历、费用与更新日期；每个进度分开开发、部署、验收。","明确谁可批准范围/预算/时间承诺与对外发送；缺真实确认则交待确认影响，不静默承诺。"],"evidenceBasis":"TheAgentCompany在模拟公司网站、数据与同事沟通中测试跨工具工作；可执行任务存在，但困难长流程不等于自动完成整职业；OSWorld2.0的长流程评估指出约束、动态信息和隐藏状态会造成失败；部分得分不能代替完整流程。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["真实团队状态时效、资源冲突、承诺与验收事实尚需独立抽核，模拟沟通不能替代。"],"mode":"staged-delivery","rationale":"公司任务已有跨工具研究，而OSWorld2显示动态信息/隐藏状态会造成长期失败。计划与状态可以自动重算，但必须保留身份、时效和承诺边界。","frontier":"TheAgentCompany实测模拟公司跨工具任务；OSWorld2.0进一步实测108个长流程，指出动态信息、约束和隐藏状态导致失败。它们支持状态化工作可评估，但没有给真实组织按期交付/协商承诺成功率。","nextValidation":"冻结一个依赖延期的项目，混入过期状态/资源冲突，要求Agent给两种可执行方案；用工单与日历程序核对，并独立检查验收差距。","futureSignals":["新状态到来后依赖/资源计划仍可执行，承诺与实际决定一致，可考虑扩大相近输入范围。","过期/推断状态写成完成，或范围变化未进入验收与预算，应收窄或暂停相应采用。"]},"workflow":{"before":"资料、实现/处理与验证分散交接，完成记录易脱离实际结果。","now":"Agent收集状态、计算依赖/资源并更新方案，直接从产物/验收区分完成阶段；负责人处理范围和真实承诺。","retained":["确认目标、资源与预算授权，主持真实参与方的冲突和决定；未确认事项不能由AI补造。"],"added":["记录实际批准/未确认、范围前后与验收差距；移交需有运营任务可接续。"]},"humanResponsibility":"确认目标、资源与预算授权，主持真实参与方的冲突和决定；未确认事项不能由AI补造。","failureHandling":"状态冲突保留证据和影响备选；不靠删范围或改验收口径把延期藏起来。","example":{"input":"上游晚两周，客户日期固定，两个关键工程师重叠。","output":"资源/依赖冲突、范围/阶段交付备选与真实决定待办、验收缺口。","failure":"把未完成接口写已完成，静默减少范围以填绿状态表。"},"sourceIds":["WU-C-COMPANY","WU-C-OS2","WU-C-COMPANY-REPO"],"evidenceLinks":[{"sourceId":"WU-C-COMPANY","supports":"公司任务有直接研究对象。","notCovered":"模拟组织；本文不采用总体比例推断项目管理能力。"},{"sourceId":"WU-C-OS2","supports":"提醒状态任务与最终完成分别验证。","notCovered":"通用跨应用基准，不是系统方案或项目绩效专属试验。"},{"sourceId":"WU-C-COMPANY-REPO","supports":"公司工作环境有可验结果方法。","notCovered":"真实协调/计划实现不由基准保证。"}],"resourceRefs":[],"career":{"direction":"项目交付基线与状态重排：依赖、资源和验收差距","evidence":"依赖/资源可执行的基线与至少一种冲突/风险备选","nextStep":"冻结一个依赖延期的项目，混入过期状态/资源冲突，要求Agent给两种可执行方案；用工单与日历程序核对，并独立检查验收差距。"},"team":{"bottleneck":"公司任务已有跨工具研究，而OSWorld2显示动态信息/隐藏状态会造成长期失败。计划与状态可以自动重算，但必须保留身份、时效和承诺边界。","allocation":"Agent收集状态、计算依赖/资源并更新方案，直接从产物/验收区分完成阶段；负责人处理范围和真实承诺。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C22-COMPANY","statement":"TheAgentCompany在模拟公司网站、数据与同事沟通中测试跨工具工作；可执行任务存在，但困难长流程不等于自动完成整职业。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-C-COMPANY","locator":"Abstract；environment and baseline evaluation","relation":"partial","supports":"公司任务有直接研究对象。","limit":"模拟组织；本文不采用总体比例推断项目管理能力。"}],"configuration":"论文v3闭源API和开放权重agent基线、模拟小型软件公司与结果验收；本页不声称某个模型在项目子集上的成绩。","progressIds":[],"limitation":"模拟组织；本文不采用总体比例推断项目管理能力。","evidenceStatus":"active"},{"id":"C22-STATE","statement":"OSWorld2.0的长流程评估指出约束、动态信息和隐藏状态会造成失败；部分得分不能代替完整流程。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-C-OS2","locator":"Abstract；binary-completion metric","relation":"partial","supports":"提醒状态任务与最终完成分别验证。","limit":"通用跨应用基准，不是系统方案或项目绩效专属试验。"}],"configuration":"v2：108流程；论文500步主口径，二元完成/部分分分开，具体模型配置按论文，未在本站运行。","progressIds":["P-C-LONG-STATE"],"limitation":"通用跨应用基准，不是系统方案或项目绩效专属试验。","evidenceStatus":"active"},{"id":"C22-ADOPT","statement":"可让Agent交付可追溯项目状态与可执行重排，但承诺与验收必须连到实际参与方和产物。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-COMPANY-REPO","locator":"task roles/result-based grading","relation":"partial","supports":"公司工作环境有可验结果方法。","limit":"真实协调/计划实现不由基准保证。"}],"configuration":null,"progressIds":[],"limitation":"真实协调/计划实现不由基准保证。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开可重建公司环境可研究工作法；真实工单/看板/日历访问与写权限需组织提供，未接入本站。","tools":[{"name":"TheAgentCompany","url":"https://github.com/TheAgentCompany/TheAgentCompany","version":"2026-10-20公开仓库HEAD 98b68ef82a47690c316f42fddb05baafaab56851；实际运行另固定依赖/镜像/模型","access":"开源；隔离部署和模型账户自备"}]},"variants":[{"title":"基线与资源计划","inputDifference":"项目初期需要范围/里程碑与资源预算。","assessmentChange":"按依赖和容量可执行性接收，不把估算当承诺。","acceptanceChange":"资源日历、关键路径、风险缓冲和验收目标。"},{"title":"风险与范围变更","inputDifference":"发生延期、问题、费用/需求变化。","assessmentChange":"生成影响/备选与决定记录，不静默变更基线。","acceptanceChange":"范围/时间/预算前后、真实批准和依赖传播。"},{"title":"验收移交与沟通","inputDifference":"进入交付阶段，需要状态与冲突/运营材料。","assessmentChange":"可整合信息与待发送内容，实际反馈/接收需真实证据。","acceptanceChange":"产物/测试/运营接手、未完成差距、已确认决定和责任期限。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C23","domainId":"computing","title":"终端配置与迁移：交付可用设备及完整用户资料","goal":"交付按目标账户、策略和软件版本配置的终端/应用，完成必要数据迁移并证明代表用户工作可用。","trigger":"给用户任务、设备/OS、软件版本、账户/策略、源资料与需保留内容，明确新配置应允许/拒绝什么。","scope":"对象是新配置/实施迁移后的可用状态，非故障修复；物理安装/身份核验若无法远程执行需明确实际接收条件。","inputs":[{"name":"配置/迁移目标","detail":"给用户任务、设备/OS、软件版本、账户/策略、源资料与需保留内容，明确新配置应允许/拒绝什么。"},{"name":"执行条件","detail":"提供获准管理入口、安装许可/资源、源目标访问、备份和回退；敏感用户内容控制在必要范围。"},{"name":"接收任务","detail":"固定代表用户身份、文件/应用任务和权限负例、迁移核对及重启/升级行为。"}],"deliverables":["可重放安装/配置/迁移与版本/策略差异","源目标资料/账户/权限对照及真实用户任务结果","恢复/回退、未支持应用/设备和运营交接记录"],"acceptance":[{"criterion":"配置按角色真实可用","method":"以用户身份执行应用/连接/文件任务，重启后检查，不能只看安装成功。","failure":"管理员安装完成但用户无法访问应用。"},{"criterion":"资料与保护不丢","method":"独立核对源目标清单/内容/权限与策略，失败可回退。","failure":"迁移只拷文件名，内容/权限错误。"},{"criterion":"维护可接续","method":"另一支持者按记录重放差异和回退，标出许可/硬件未完成项。","failure":"脚本只在当前设备碰巧生效且没有版本记录。"}],"assessment":{"recommendation":"可委派获准安装、策略配置和数据核对，按实际用户任务、资料完整与权限接收；缺设备/许可时交明确实施阻塞。","conditions":["提供获准管理入口、安装许可/资源、源目标访问、备份和回退；敏感用户内容控制在必要范围。","固定代表用户身份、文件/应用任务和权限负例、迁移核对及重启/升级行为。"],"evidenceBasis":"微软IT管理员随机试验在登录排障、策略与设备故障情景观察到Copilot辅助的准确性/耗时改善；这不是无人维修实测；Security Copilot在Entra/Intune等工具中的接入受租户、产品权限和容量影响；不能由产品文档推定已经在本用户可用。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["目标终端/软件/许可、源目标资料完整与用户权限/重启行为尚未实际配置验证。"],"mode":"staged-delivery","rationale":"IT管理员RCT含设备策略/故障情景，对信息整合有直接邻近依据；配置完成仍要目标设备/用户检查，不能由回答准确推断迁移成功。","frontier":"IT管理员RCT实际在Entra/Intune的登录、设备策略/排障任务比较人加Copilot与对照；研究对象是辅助信息工作，不是自主新机迁移。现行文档给出产品/权限接入条件，真实设备资料与用户任务仍需另验。","nextValidation":"给一个测试终端和源资料快照，用用户/越权角色验收；迁移后重启，核对资料内容、应用任务与回退。","futureSignals":["新终端/软件版本下用户任务、资料与权限核对都保持正确，可考虑扩大相近输入范围。","安装成功但用户无法使用，或内容/权限在迁移中丢失，应收窄或暂停相应采用。"]},"workflow":{"before":"资料、实现/处理与验证分散交接，完成记录易脱离实际结果。","now":"Agent整合配置目标、操作获准管理入口并做资料/策略核对；确定性安装/复制可直接脚本执行。","retained":["确认身份/设备授权、数据保留、许可和不可逆账户操作。"],"added":["另一支持者按记录重放差异和回退，标出许可/硬件未完成项。"]},"humanResponsibility":"确认身份/设备授权、数据保留、许可和不可逆账户操作。","failureHandling":"源资料或许可不足不伪造实施；保留备份和未迁移清单，不清空用户资料制造干净状态。","example":{"input":"为新终端安装业务软件并迁移旧资料/策略。","output":"角色化配置、文件内容/权限核对与原用户代表任务。","failure":"安装成功但凭据/权限缺失，文件内容不完整。"},"sourceIds":["WU-C-ITRCT","WU-C-ITACCESS"],"evidenceLinks":[{"sourceId":"WU-C-ITRCT","supports":"IT支持工作有直接领域试验。","notCovered":"厂商相关、参与者加Copilot；本页不转用总体提升幅度。"},{"sourceId":"WU-C-ITACCESS","supports":"明确当前可用条件。","notCovered":"未在本站租户启用，不保证所有设备动作可执行。"}],"resourceRefs":[],"career":{"direction":"终端配置与迁移：交付可用设备及完整用户资料","evidence":"可重放安装/配置/迁移与版本/策略差异","nextStep":"给一个测试终端和源资料快照，用用户/越权角色验收；迁移后重启，核对资料内容、应用任务与回退。"},"team":{"bottleneck":"IT管理员RCT含设备策略/故障情景，对信息整合有直接邻近依据；配置完成仍要目标设备/用户检查，不能由回答准确推断迁移成功。","allocation":"Agent整合配置目标、操作获准管理入口并做资料/策略核对；确定性安装/复制可直接脚本执行。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C23-ITRCT","statement":"微软IT管理员随机试验在登录排障、策略与设备故障情景观察到Copilot辅助的准确性/耗时改善；这不是无人维修实测。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-C-ITRCT","locator":"Abstract；three scenarios","relation":"partial","supports":"IT支持工作有直接领域试验。","limit":"厂商相关、参与者加Copilot；本页不转用总体提升幅度。"}],"configuration":"v2：Entra/Intune中有/无Security Copilot的随机组；三情景，问答/自由回答口径区分；v2修正时间数字。","progressIds":["P-C-IT-RCT"],"limitation":"厂商相关、参与者加Copilot；本页不转用总体提升幅度。","evidenceStatus":"active"},{"id":"C23-ACCESS","statement":"Security Copilot在Entra/Intune等工具中的接入受租户、产品权限和容量影响；不能由产品文档推定已经在本用户可用。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-ITACCESS","locator":"Eligibility；permissions；agents deployment","relation":"method","supports":"明确当前可用条件。","limit":"未在本站租户启用，不保证所有设备动作可执行。"}],"configuration":null,"progressIds":[],"limitation":"未在本站租户启用，不保证所有设备动作可执行。","evidenceStatus":"active"},{"id":"C23-ADOPT","statement":"可委派终端配置/迁移，但实际可用性与资料完整必须在源目标和用户身份下核对。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-ITRCT","locator":"device policy/device troubleshooting scenarios","relation":"partial","supports":"领域任务研究提供邻近整合依据。","limit":"设备策略诊断研究不代表本迁移已实测。"}],"configuration":null,"progressIds":[],"limitation":"设备策略诊断研究不代表本迁移已实测。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"有条件商业助手与开源桌面试验环境；未连接任何用户设备。","tools":[{"name":"Security Copilot","url":"https://learn.microsoft.com/en-us/copilot/security/security-copilot-inclusion","version":"2026-10-20核查；租户实际版本/启用另记","access":"需产品许可、容量、角色与组织设置"},{"name":"OSWorld","url":"https://github.com/xlang-ai/OSWorld","version":"2026-10-20公开仓库HEAD 9f4b4d34319589f3ba6947da6045b0955b08900e；实际运行另固定依赖/镜像/模型","access":"开源隔离桌面环境，模型与系统资源另备"}]},"variants":[{"title":"软件与账户配置","inputDifference":"新装应用、身份、连接和策略。","assessmentChange":"按角色任务和拒绝矩阵接收。","acceptanceChange":"重启、版本/许可、访问与越权负例。"},{"title":"用户数据迁移","inputDifference":"源目标设备/软件版本不同。","assessmentChange":"先固定保留范围、备份与回退再实施。","acceptanceChange":"内容/权限/格式、应用可读取性、重复/中断和回退。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C24","domainId":"computing","title":"系统采用与用户支持材料：让使用者完成真实任务","goal":"交付针对角色任务的指引/培训/交接材料及实际使用反馈，定位系统或知识缺口并给出可复验改进。","trigger":"给角色、真实工作任务、前置知识、现有系统/流程与需要的独立完成结果。","scope":"产物是可用支持材料与采用证据，不是生成教材本身；不能由模拟同事评价冒充真实用户完成效果。","inputs":[{"name":"使用目标","detail":"给角色、真实工作任务、前置知识、现有系统/流程与需要的独立完成结果。"},{"name":"材料与系统","detail":"冻结目标版本、权限、任务样例、已知错误/常见问题和可访问性条件，提供安全演练环境。"},{"name":"反馈与接收","detail":"准备代表用户或独立接收者、任务脚本与成功/阻塞判据；未经授权的外发只交待发送材料。"}],"deliverables":["按角色任务可操作的指引/培训/交接及版本","代表任务的真实完成/阻塞、反馈与修改记录","知识缺口/系统缺陷分离、升级入口与后续维护条件"],"acceptance":[{"criterion":"步骤在目标环境可执行","method":"独立接收者按材料完成任务，核对真实结果，不靠作者解释救场。","failure":"只有截图说明，关键权限/异常步骤无法执行。"},{"criterion":"缺口定位正确","method":"观察失败到底是材料、权限、产品缺陷还是前置知识；修订后重测。","failure":"把系统错误归咎用户，或把同意度评分当学会任务。"},{"criterion":"资料可维护且可访问","method":"记录版本、角色/语言、异常/恢复与升级条件，按键盘/缩放等目标条件查使用。","failure":"系统升级后旧步骤继续分发且无异常入口。"}],"assessment":{"recommendation":"可委派制作角色化支持材料并组织问题诊断；以独立使用者实际完成任务接收，缺真实用户时只能接收材料可执行性。","conditions":["冻结目标版本、权限、任务样例、已知错误/常见问题和可访问性条件，提供安全演练环境。","准备代表用户或独立接收者、任务脚本与成功/阻塞判据；未经授权的外发只交待发送材料。"],"evidenceBasis":"TheAgentCompany公开任务环境与结果检查可用于检验指引是否可执行；用户采用/培训收益没有直接任务试验，需以真实独立用户任务反馈验证。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["目标用户独立完成、材料/产品/权限失败区分和版本变化后的采用效果尚无现场反馈。"],"mode":"staged-delivery","rationale":"公司环境提供跨工具任务方法，可用于检查指引步骤；采用/培训收益没有本任务直接AI实测，因此不从办公基准外推。具体任务完成与反馈能使这一交付可验证。","frontier":"公开公司环境可执行多角色任务并按结果检查，可用于验证指引步骤是否可操作；它没有直接展示本系统用户培训或长期采用效果。当前可委派材料制作/环境重放，真实独立用户完成是下一项证据。","nextValidation":"选一项目标用户原先常失败的任务，冻结目标环境；让未看制作过程的接收者照材料执行，记录完成、阻塞、时间和修订后结果。","futureSignals":["不同目标用户可独立按材料完成任务，支持返工与升级负担下降，可考虑扩大相近输入范围。","满意度/材料数量掩盖实际任务失败，或系统更新使指引过期，应收窄或暂停相应采用。"]},"workflow":{"before":"资料、实现/处理与验证分散交接，完成记录易脱离实际结果。","now":"Agent按角色生成可执行材料、在环境重放并根据真实反馈修订；稳定步骤可自动版本化，使用效果仍从真实任务观察。","retained":["选代表用户/任务、确认外发和数据授权，决定材料可用范围和支持责任。"],"added":["记录版本、角色/语言、异常/恢复与升级条件，按键盘/缩放等目标条件查使用。"]},"humanResponsibility":"选代表用户/任务、确认外发和数据授权，决定材料可用范围和支持责任。","failureHandling":"没有真实反馈不虚构满意/解决率；若问题来自产品或权限则转具体修复任务并保留支持路径。","example":{"input":"新系统上线后用户不会处理导入错误。","output":"可执行错误诊断/恢复指引，独立用户完成与阻塞记录。","failure":"只发通用培训PDF，实际导入仍失败却说培训完成。"},"sourceIds":["WU-C-COMPANY-REPO"],"evidenceLinks":[{"sourceId":"WU-C-COMPANY-REPO","supports":"可执行任务方法提供邻近参照。","notCovered":"不证明本教材学习效果或真实同事反馈。"}],"resourceRefs":[],"career":{"direction":"系统采用与用户支持材料：让使用者完成真实任务","evidence":"按角色任务可操作的指引/培训/交接及版本","nextStep":"选一项目标用户原先常失败的任务，冻结目标环境；让未看制作过程的接收者照材料执行，记录完成、阻塞、时间和修订后结果。"},"team":{"bottleneck":"公司环境提供跨工具任务方法，可用于检查指引步骤；采用/培训收益没有本任务直接AI实测，因此不从办公基准外推。具体任务完成与反馈能使这一交付可验证。","allocation":"Agent按角色生成可执行材料、在环境重放并根据真实反馈修订；稳定步骤可自动版本化，使用效果仍从真实任务观察。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C24-METHOD","statement":"可把公开公司工作环境中的可执行任务方法用于检查支持材料是否引导到真实结果，但培训/采用收益需真实使用反馈。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-COMPANY-REPO","locator":"result-based evaluation/task environment","relation":"partial","supports":"可执行任务方法提供邻近参照。","limit":"不证明本教材学习效果或真实同事反馈。"}],"configuration":null,"progressIds":[],"limitation":"不证明本教材学习效果或真实同事反馈。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"材料可直接用普通文档工具制作；浏览器用于在目标系统重放指引，真实用户与系统权限需另提供。","tools":[{"name":"Playwright MCP","url":"https://github.com/microsoft/playwright-mcp","version":"2026-10-20公开仓库HEAD ab6d67cbf8b619338f096980806fdbe29432cc4d；实际运行另固定依赖/镜像/模型","access":"公开接口；目标系统/用户身份和独立接收者另备"}]},"variants":[{"title":"系统采用与交接","inputDifference":"新流程/系统角色需要独立使用或运营接手。","assessmentChange":"先验证代表任务，再扩大培训范围。","acceptanceChange":"实际任务、权限/异常、运营责任和未完成差距。"},{"title":"知识库与自助支持","inputDifference":"重复问题需要可维护解决路径。","assessmentChange":"生成材料后以真实恢复/升级结果迭代。","acceptanceChange":"版本、查询匹配、失败分流和用户独立完成。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C25","domainId":"computing","title":"论文复现交付：贡献到实际实验结果","goal":"针对固定论文贡献交付独立可运行实现及实际复现实验，逐项报告成功、差异与未复现范围。","trigger":"固定原论文版本、补充/澄清和目标贡献，列需要复现的实验/理论结论及允许差异。","scope":"对象为既有结果的复现，不是原创发现；部分子贡献得分不能写成整篇论文成功，理论证明另以对应论证接收。","inputs":[{"name":"复现目标","detail":"固定原论文版本、补充/澄清和目标贡献，列需要复现的实验/理论结论及允许差异。"},{"name":"数据与环境","detail":"准备许可数据、代码访问边界、硬件/依赖、基线和实际预算；区分从零实现与可复用原作者代码。"},{"name":"独立rubric","detail":"把贡献连到可验证产物和实验结果，预留独立复跑；LLM评价不能单独证明数值/数学正确。"}],"deliverables":["按贡献组织的实现、环境/依赖与数据来源","实际执行日志、基线/消融/结果表及预算/失败账本","逐贡献复现状态、与原文差异原因及可复用材料"],"acceptance":[{"criterion":"确实运行关键贡献","method":"从固定环境重跑核心实验并核对产物/数据，代码存在或论文文字不计运行。","failure":"只生成完整仓库而没有训练/实验输出。"},{"criterion":"比较口径对应原文","method":"检查数据划分、预处理、预算、硬件/模型版本与指标定义，差异显式报告。","failure":"换更容易数据或多用预算却宣称精确复现。"},{"criterion":"完成声明相称","method":"每项贡献有证据与失败/缺口，不将部分分或裁判摘要改为整论文成功。","failure":"模型judge说像原文，就忽略数值/理论错误。"}],"assessment":{"recommendation":"可委派理解论文、实现、实验与诊断循环；按每项贡献的真实产物接收，可部分完成但不能省略失败或预算。","conditions":["准备许可数据、代码访问边界、硬件/依赖、基线和实际预算；区分从零实现与可复用原作者代码。","把贡献连到可验证产物和实验结果，预留独立复跑；LLM评价不能单独证明数值/数学正确。"],"evidenceBasis":"PaperBench要求从论文到代码和实际实验，以贡献rubric分开检查；原实验报告部分复现能力且没有超过其人类参照。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["目标贡献的真实运行、原文口径/预算比较和独立干净环境复跑尚未完成。"],"mode":"candidate-exploration","rationale":"PaperBench直接把论文复现作为实际运行任务，提供贡献rubric与裁判校准；这比代码生成更贴近研究复用结果，但仍是有分布和评价误差的评估。","frontier":"PaperBench以20篇ICML2024论文、作者参与的贡献rubric和实际实验评估复现；其原实验展示部分贡献复现能力并有裁判局限。部分评分不等于整论文成功，本目标论文未运行。","nextValidation":"选一项可公开且预算可控的论文贡献，校准独立rubric；让Agent交完整复现账本，再从干净环境复跑关键输出并解释差异。","futureSignals":["独立环境能重复关键贡献，差异/预算与失败全部可追踪，可考虑扩大相近输入范围。","只生成代码/论文文本、隐藏未跑实验或把部分分写成完整复现，应收窄或暂停相应采用。"]},"workflow":{"before":"资料、实现/处理与验证分散交接，完成记录易脱离实际结果。","now":"Agent可检索澄清、实现和运行，按失败/统计差异修订；目标贡献与独立复跑保持固定。","retained":["确定复现价值、数据许可/预算，裁决何种差异仍算同一贡献；研究新颖性不在复现成功中偷渡。"],"added":["每项贡献有证据与失败/缺口，不将部分分或裁判摘要改为整论文成功。"]},"humanResponsibility":"确定复现价值、数据许可/预算，裁决何种差异仍算同一贡献；研究新颖性不在复现成功中偷渡。","failureHandling":"运行环境或预算不足交已实现/已运行/未运行分界，不能用预计成绩填结果。","example":{"input":"复现一篇图模型论文的两个关键实验与一个消融。","output":"可重跑实现、数据/种子/资源、真实表格和未复现贡献。","failure":"只画类似图，不记录训练与选择次数。"},"sourceIds":["WU-C-PAPER"],"evidenceLinks":[{"sourceId":"WU-C-PAPER","supports":"直接研究复现任务证据。","notCovered":"部分评分和LLM裁判均有局限；不代表完整复现成功率。"}],"resourceRefs":[],"career":{"direction":"论文复现交付：贡献到实际实验结果","evidence":"按贡献组织的实现、环境/依赖与数据来源","nextStep":"选一项可公开且预算可控的论文贡献，校准独立rubric；让Agent交完整复现账本，再从干净环境复跑关键输出并解释差异。"},"team":{"bottleneck":"PaperBench直接把论文复现作为实际运行任务，提供贡献rubric与裁判校准；这比代码生成更贴近研究复用结果，但仍是有分布和评价误差的评估。","allocation":"Agent可检索澄清、实现和运行，按失败/统计差异修订；目标贡献与独立复跑保持固定。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C25-REPLICATION","statement":"PaperBench要求从论文到代码和实际实验，以贡献rubric分开检查；原实验报告部分复现能力且没有超过其人类参照。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-C-PAPER","locator":"Abstract；rubrics/judge/baselines","relation":"partial","supports":"直接研究复现任务证据。","limit":"部分评分和LLM裁判均有局限；不代表完整复现成功率。"}],"configuration":"v3：20篇ICML2024论文，作者参与rubric，8316子项；原最佳Claude3.5 Sonnet(New)+开放脚手架，本站不转用评分。","progressIds":["P-C-REPRO-RESEARCH"],"limitation":"部分评分和LLM裁判均有局限；不代表完整复现成功率。","evidenceStatus":"active"},{"id":"C25-ADOPT","statement":"可让Agent交从论文到实际实验的复现包，但每项贡献应由独立结果核对，部分成功与完整复现分开。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-PAPER","locator":"rubrics and judge","relation":"partial","supports":"直接复现评测支持按贡献接收。","limit":"裁判和数据分布有限；本目标未试验。"}],"configuration":null,"progressIds":[],"limitation":"裁判和数据分布有限；本目标未试验。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"实际复现使用PaperBench与目标论文的贡献和执行入口；目标数据、环境与计算条件仍需准备。","tools":[{"name":"PaperBench论文及作者代码入口","url":"https://arxiv.org/abs/2504.01848v3","version":"论文v3和目标贡献rubric；代码/依赖/镜像执行时固定","access":"公开论文链接代码；数据许可、计算硬件和模型预算另备"}]},"variants":[{"title":"从零与复用原代码","inputDifference":"是否能读取作者仓库，是否需从零实现。","assessmentChange":"必须保留使用条件，不混合两种成绩。","acceptanceChange":"依赖/代码来源、同数据指标和公平预算。"},{"title":"数值与理论贡献","inputDifference":"论文同时有经验结果和定理。","assessmentChange":"经验按运行验收，理论另查完整论证。","acceptanceChange":"计算输出、统计不确定性与假设/证明分别有证据。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C26","domainId":"computing","title":"检索与工具Agent应用交付：任务结果和外部状态","goal":"交付能在规定数据、工具与权限内完成目标任务的检索/Agent应用，附引用/结果、拒绝行为、外部副作用和失败恢复证据。","trigger":"给真实用户目标、代表输入、正确结果/状态、无答案范围与不得做的动作；列需检索的信息来源和工具。","scope":"对象是实际任务与工具状态，不是模型聊天质量；可用确定性流程时不强制规划循环，内部路径允许动态选择。","inputs":[{"name":"任务契约","detail":"给真实用户目标、代表输入、正确结果/状态、无答案范围与不得做的动作；列需检索的信息来源和工具。"},{"name":"环境与权限","detail":"固定模型/提示/检索库/工具版本、账户/网络、许可和费用/步数预算，敏感工具隔离试验。"},{"name":"独立控制","detail":"准备保留任务、引用真值、提示注入/越权/无答案/超时/重试控制及真实状态查询，区分工具调用和任务完成。"}],"deliverables":["可运行检索/Agent应用、配置/工具/版本与权限边界","保留任务产物/引用支持、工具轨迹和外部状态核对","注入/越权/重试/依赖失败结果、成本与回退/停止路径"],"acceptance":[{"criterion":"结果和来源对应目标","method":"独立检查引用是否支持每项结论，查询/文件/动作是否改变正确状态；无答案可拒绝。","failure":"引用存在但不支持答案，或只调用工具就算完成。"},{"criterion":"边界与副作用受控","method":"用注入/越权/恶意检索和重试测试，确认未执行禁止动作、未重复写入。","failure":"从网页指令拿凭据或重复扣款。"},{"criterion":"失败和成本可接收","method":"截断依赖、改变schema/上下文，记录预算/步骤、失败解释与恢复；不能无止境循环。","failure":"超预算重试隐藏失败，最终产物也无法核对。"}],"assessment":{"recommendation":"可委派应用实现与任务评测；先冻结独立结果/状态控制，按效果、权限和失败路径接收，单条演示不能替代任务族表现。","conditions":["固定模型/提示/检索库/工具版本、账户/网络、许可和费用/步数预算，敏感工具隔离试验。","准备保留任务、引用真值、提示注入/越权/无答案/超时/重试控制及真实状态查询，区分工具调用和任务完成。"],"evidenceBasis":"公开内部数据Agent案例描述了查询/上下文与结果修正闭环，是专用数据系统的自报能力展示，可作邻近任务组织依据；Playwright MCP等公开接口可供Agent操作和观察，但工具存在不能代替任务完成/权限验证。","localStatus":"本次只核查公开材料和任务契约；未在本站运行该任务、工具或模型。","gaps":["目标检索/工具版本、任务分布、无答案/注入/越权及幂等外部状态尚未校准。"],"mode":"staged-delivery","rationale":"公开专用数据Agent案例表明检索/查询/结果反馈可以形成任务闭环，公开浏览器工具提供操作接口；迁移到新工具/权限必须用目标任务和副作用控制核验。","frontier":"公开内部数据Agent案例描述专用语义/查询/结果反馈闭环；Playwright MCP公开工具接口便于操作网页。前者是厂商专用案例、后者是方法，均不证明本应用写工具、对抗注入和任务族稳定性。","nextValidation":"从真实流程取一组脱敏保留任务及无答案/注入/越权/重试控制；让Agent交可运行应用，独立读出工具最终状态并记录实际费用。","futureSignals":["新保留任务下结果/来源和外部状态持续正确，预算与边界受控，可考虑扩大相近输入范围。","注入导致越权、引用不支持结论或重试造成重复副作用，应收窄或暂停相应采用。"]},"workflow":{"before":"资料、实现/处理与验证分散交接，完成记录易脱离实际结果。","now":"Agent可以检索、选工具与修订结果；稳定部分用普通程序，独立结果检查与权限门禁持续作用。","retained":["定义任务用途、工具授权、敏感数据及错误代价，明确外部写入/发送权限和停止条件。"],"added":["截断依赖、改变schema/上下文，记录预算/步骤、失败解释与恢复；不能无止境循环。"]},"humanResponsibility":"定义任务用途、工具授权、敏感数据及错误代价，明确外部写入/发送权限和停止条件。","failureHandling":"依赖或语义不足输出可定位阻塞与已完成状态；不能编造引用/工具结果，失败恢复须避免重复副作用。","example":{"input":"内部知识问答需要查版本文档并开一个获准工单，无答案应返回缺口。","output":"可运行应用、逐项来源与正确工单状态，注入/越权/重试负控制。","failure":"答对文字却错开工单，或把检索内容的恶意指令当系统授权。"},"sourceIds":["WU-DATA-AGENT","WU-C-PWMCP"],"evidenceLinks":[{"sourceId":"WU-DATA-AGENT","supports":"已有专用Agent任务闭环案例。","notCovered":"不是可下载同配置应用，也不评估本站工具写入/对抗条件。"},{"sourceId":"WU-C-PWMCP","supports":"提供实际工具接口。","notCovered":"不保证本应用效果或阻挡所有注入。"}],"resourceRefs":[],"career":{"direction":"检索与工具Agent应用交付：任务结果和外部状态","evidence":"可运行检索/Agent应用、配置/工具/版本与权限边界","nextStep":"从真实流程取一组脱敏保留任务及无答案/注入/越权/重试控制；让Agent交可运行应用，独立读出工具最终状态并记录实际费用。"},"team":{"bottleneck":"公开专用数据Agent案例表明检索/查询/结果反馈可以形成任务闭环，公开浏览器工具提供操作接口；迁移到新工具/权限必须用目标任务和副作用控制核验。","allocation":"Agent可以检索、选工具与修订结果；稳定部分用普通程序，独立结果检查与权限门禁持续作用。","nextStep":"在真实授权任务核对产物、失败、返工与接收时间，再据此配置工具、环境和人员。"},"claims":[{"id":"C26-DATAAGENT","statement":"公开内部数据Agent案例描述了查询/上下文与结果修正闭环，是专用数据系统的自报能力展示，可作邻近任务组织依据。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-DATA-AGENT","locator":"How it works；Context is everything","relation":"partial","supports":"已有专用Agent任务闭环案例。","limit":"不是可下载同配置应用，也不评估本站工具写入/对抗条件。"}],"configuration":"2026-01-29案例，GPT-5.2与专用数据/语义/权限；golden SQL/结果比较；未公开本任务分母/成本。","progressIds":[],"limitation":"不是可下载同配置应用，也不评估本站工具写入/对抗条件。","evidenceStatus":"active"},{"id":"C26-TOOLS","statement":"Playwright MCP等公开接口可供Agent操作和观察，但工具存在不能代替任务完成/权限验证。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-PWMCP","locator":"README tools/browser configuration","relation":"method","supports":"提供实际工具接口。","limit":"不保证本应用效果或阻挡所有注入。"}],"configuration":null,"progressIds":[],"limitation":"不保证本应用效果或阻挡所有注入。","evidenceStatus":"active"},{"id":"C26-ADOPT","statement":"可委派检索/工具应用的实现与评测，但接收需同时核对结果、来源与真实外部状态；任务预算与失败边界明确。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-DATA-AGENT","locator":"Lessons/contexts/results","relation":"partial","supports":"专用任务闭环提供邻近迁移依据。","limit":"外部状态、权限与真实任务效果必须另验。"}],"configuration":null,"progressIds":[],"limitation":"外部状态、权限与真实任务效果必须另验。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"数据Agent为专用内部系统；公开浏览器接口可用，目标模型/检索/工具环境需配置并固定版本。","tools":[{"name":"Playwright MCP","url":"https://github.com/microsoft/playwright-mcp","version":"2026-10-20公开仓库HEAD ab6d67cbf8b619338f096980806fdbe29432cc4d；实际运行另固定依赖/镜像/模型","access":"开源接口；应用身份/模型与目标工具权限另备"}]},"variants":[{"title":"检索与可引用回答","inputDifference":"结果为依据资料的回答/分析，不改变外部状态。","assessmentChange":"可整包交付，先校准来源支持、缺失/过期资料与无答案。","acceptanceChange":"逐项引用支持、版本/新鲜度、无答案和恶意资料。"},{"title":"工具与外部状态","inputDifference":"需要创建/修改工单、文件或执行获准操作。","assessmentChange":"按任务效果和副作用分阶段接收，权限不能由资料授予。","acceptanceChange":"工具参数、真实最终状态、注入/越权、幂等与失败恢复。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C27","domainId":"computing","title":"备份资产交付：持续产生受保护且可还原的备份集","goal":"建立并维护满足恢复目标的基础/增量备份与连续归档机制，交付真实备份资产、失败检测和保留依赖。","trigger":"给库/版本、关键数据/配置、RPO/RTO、恢复范围、保留期限与数据敏感性；明确可接受备份负载。","scope":"产物是可持续生成的可恢复资产，而非从已有备份还原实例；W-C16作为独立还原验收接口，两者可分别接收。","inputs":[{"name":"恢复目标与源库","detail":"给库/版本、关键数据/配置、RPO/RTO、恢复范围、保留期限与数据敏感性；明确可接受备份负载。"},{"name":"资产与执行环境","detail":"冻结备份/归档工具、调度、存储/加密/密钥与访问权限，配置文件需单独保护；列基础/增量/WAL依赖。"},{"name":"运行与验收","detail":"独立准备失败/落后/空间不足信号、保留删除规则、源库业务探针和W-C16恢复环境。"}],"deliverables":["可执行基础/增量备份、连续归档和调度/保护配置","真实备份/归档清单、链与配置/密钥依赖；受限访问/完整性结果","失败/落后告警、保留依赖与周期/负载取舍；独立还原所需接口和结果"],"acceptance":[{"criterion":"确实产生可用资产","method":"执行备份和归档，核对基础/增量/WAL依赖与配置、完整性、访问和密钥可获得性；交W-C16真还原。","failure":"只有脚本或文件名，没有可还原链。"},{"criterion":"失败被及时暴露","method":"模拟归档返回失败/存储满/落后，观察告警、WAL积压与源库影响，不把缺测写健康。","failure":"归档命令误报成功导致所需日志被回收。"},{"criterion":"保留与预算一致","method":"演练保留清理，确认不删仍被引用的基础/WAL，测备份负载和实际周期，复核目标恢复范围。","failure":"删旧基础后增量仍在，或备份拖垮源库却声称满足目标。"}],"assessment":{"recommendation":"可委派备份/归档配置与受限执行，按真实受保护资产、运行失败控制和独立还原接收；不由一般DB运维成绩推断备份成功。","conditions":["冻结备份/归档工具、调度、存储/加密/密钥与访问权限，配置文件需单独保护；列基础/增量/WAL依赖。","独立准备失败/落后/空间不足信号、保留删除规则、源库业务探针和W-C16恢复环境。"],"evidenceBasis":"PostgreSQL18区分基础/增量备份、连续归档和还原，并说明归档失败/落后后果。这支持可执行接收设计；AI组织与调试是工具方法上的迁移判断。","localStatus":"本次核查公开资料与任务契约，未在本站运行目标任务或新增模型试次。","gaps":["在授权测试库建立一个备份周期，注入归档失败/空间不足与保留清理，再由独立W-C16恢复到指定事务点。"],"mode":"staged-delivery","rationale":"PostgreSQL18区分基础/增量备份、连续归档和还原，并说明归档失败/落后后果。这支持可执行接收设计；AI组织与调试是工具方法上的迁移判断。","frontier":"原生备份/归档机制明确且可获得；尚未核实本备份任务的直接AI试验，脚本生成与一般运维结果均不能证明资产可还原。","nextValidation":"在授权测试库建立一个备份周期，注入归档失败/空间不足与保留清理，再由独立W-C16恢复到指定事务点。","futureSignals":["多个周期后备份链、失败信号与独立还原持续满足恢复预算，可扩大自动维护。","资产链断裂、落后未报警或清理破坏可恢复性时停止相应自动删除/备份承诺。"]},"workflow":{"before":"备份脚本、存储、告警和恢复演练分别维护，成功状态容易只看任务退出码。","now":"Agent联合配置、执行、资产核对与失败诊断；稳定调度直接程序运行，独立还原挑战资产。","retained":["确定恢复/保留/保护目标、密钥与生产负载/删除权限，裁决RPO/RTO与成本取舍。"],"added":["演练保留清理，确认不删仍被引用的基础/WAL，测备份负载和实际周期，复核目标恢复范围。"]},"humanResponsibility":"确定恢复/保留/保护目标、密钥与生产负载/删除权限，裁决RPO/RTO与成本取舍。","failureHandling":"无法获得所需配置/日志/密钥时交准确缺口；不把无法还原资产标完成，保留原链避免误删。","example":{"input":"库需持续归档且保留一周恢复范围，备份窗口有限。","output":"真实资产/链/配置保护，落后/失败告警与独立恢复结果。","failure":"备份作业绿色，但归档误报成功且日志缺段。"},"sourceIds":["WU-C-PITR"],"evidenceLinks":[{"sourceId":"WU-C-PITR","supports":"支持备份产生/维护的独立结果契约。","notCovered":"原生方法不是AI备份试验。"}],"resourceRefs":[],"career":{"direction":"数据库备份与保护","evidence":"可执行基础/增量备份、连续归档和调度/保护配置","nextStep":"在授权测试库建立一个备份周期，注入归档失败/空间不足与保留清理，再由独立W-C16恢复到指定事务点。"},"team":{"bottleneck":"PostgreSQL18区分基础/增量备份、连续归档和还原，并说明归档失败/落后后果。这支持可执行接收设计；AI组织与调试是工具方法上的迁移判断。","allocation":"Agent联合配置、执行、资产核对与失败诊断；稳定调度直接程序运行，独立还原挑战资产。","nextStep":"在授权测试库建立一个备份周期，注入归档失败/空间不足与保留清理，再由独立W-C16恢复到指定事务点。"},"claims":[{"id":"C27-BACKUP","statement":"PostgreSQL18提供基础/增量备份与WAL归档方法，归档成功语义、失败积压和链依赖必须由实际资产检查。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-PITR","locator":"25.3.1—25.3.3","relation":"method","supports":"支持备份产生/维护的独立结果契约。","limit":"原生方法不是AI备份试验。"}],"configuration":null,"progressIds":[],"limitation":"原生方法不是AI备份试验。","evidenceStatus":"active"},{"id":"C27-ADOPT","statement":"可把备份配置、受限运行和资产诊断委派给Agent，以真实链、保护/失败控制与独立还原接收。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-PITR","locator":"归档/备份/恢复接口","relation":"partial","supports":"可执行原生机制使结果可检验。","limit":"没有本库AI成功率，备份创建与还原分别接收。"}],"configuration":null,"progressIds":[],"limitation":"没有本库AI成功率，备份创建与还原分别接收。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"原生工具可用；调度/存储/密钥和访问保护由目标环境提供，未在本站创建备份。","tools":[{"name":"PostgreSQL18备份/归档","url":"https://www.postgresql.org/docs/18/continuous-archiving.html","version":"文档18；实际源库、工具、格式与链固定","access":"需授权库、隔离存储/密钥和还原空间"}]},"variants":[{"title":"基础增量与归档链","inputDifference":"使用增量依赖或连续恢复目标。","assessmentChange":"先核对实际依赖与归档完整性，再接收恢复范围。","acceptanceChange":"基础/增量/WAL依赖、缺段与恢复点；配置文件另备。"},{"title":"保留与负载约束","inputDifference":"资产长期增长、备份争用源库资源。","assessmentChange":"清理和周期可委派但不能破坏可恢复窗口。","acceptanceChange":"保留删除演练、存储满/落后、业务负载与实际恢复范围。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C28","domainId":"computing","title":"复制与HA切换：单主、事务连续与冗余恢复","goal":"交付可验证的复制/故障切换机制及切换后业务连续状态，确保单主、事务边界、连接端点和备机冗余重新成立。","trigger":"固定DB版本、主/备/见证/客户端端点、同步语义、复制延迟与允许丢事务、RPO/RTO及读写一致性。","scope":"工作对象是复制拓扑及故障后的持续服务，不是schema/数据迁移。提升备机不等于完整恢复；旧主隔离和重建冗余都在接收范围。","inputs":[{"name":"拓扑与事务契约","detail":"固定DB版本、主/备/见证/客户端端点、同步语义、复制延迟与允许丢事务、RPO/RTO及读写一致性。"},{"name":"故障与权限","detail":"提供隔离HA环境、检测/通知/fencing工具、网络分区/主故障/旧主回归控制和授权操作；不假定原生DB包揽自动检测。"},{"name":"业务和重建探针","detail":"记录切换前已确认事务/复制位点，准备客户端重连/重试、双主写拒绝、备机重建及第二故障验收。"}],"deliverables":["复制/检测/隔离/切换与连接端点配置、运行说明","故障/分区/旧主重入下单主和事务/客户端连续性证据、实际恢复时间","备机重新建立、复制追赶/读服务与第二次切换验证；未满足冗余说明"],"acceptance":[{"criterion":"单主与事务边界成立","method":"故障/分区后确认唯一可写主，旧主恢复仍被隔离；按已确认事务与复制延迟核对丢失/重复。","failure":"新主可写但旧主也重回写入，形成分裂脑。"},{"criterion":"业务连接真实恢复","method":"从正常客户端端点重连/重试读写，测实际中断与RPO/RTO，核对会话/事务语义。","failure":"手动连新主成功，却客户端仍指旧端点。"},{"criterion":"冗余重新成立","method":"重建/追赶备机并验证读/复制服务，重新施加一个故障，核对单主和事务状态。","failure":"只恢复主库业务，副本未恢复却宣告完整HA。"}],"assessment":{"recommendation":"可委派复制/切换配置和受控演练；按单主、事务/客户端与冗余重建阶段接收，高影响切换有明确权限。","conditions":["提供隔离HA环境、检测/通知/fencing工具、网络分区/主故障/旧主回归控制和授权操作；不假定原生DB包揽自动检测。","记录切换前已确认事务/复制位点，准备客户端重连/重试、双主写拒绝、备机重建及第二故障验收。"],"evidenceBasis":"PostgreSQL18要求旧主隔离和新备机重建，原生DB不提供完整故障检测。DBA-Bench持续场景实际出现主事务恢复而副本未恢复，说明不能只验主库。","localStatus":"本次核查公开资料与任务契约，未在本站运行目标任务或新增模型试次。","gaps":["在隔离主备拓扑冻结事务/端点，测试主故障、网络分区、旧主回归与备机重建，再执行第二次切换。"],"mode":"staged-delivery","rationale":"PostgreSQL18要求旧主隔离和新备机重建，原生DB不提供完整故障检测。DBA-Bench持续场景实际出现主事务恢复而副本未恢复，说明不能只验主库。","frontier":"原生机制与状态型实验共同给出完整结果边界；DBA-Bench只是有限持续场景，不是此HA工具/拓扑的生产切换可靠率。","nextValidation":"在隔离主备拓扑冻结事务/端点，测试主故障、网络分区、旧主回归与备机重建，再执行第二次切换。","futureSignals":["不同故障/延迟条件下单主、事务、客户端和冗余都持续可复验，可扩大受限自动切换。","旧主重入可写、客户端未恢复或备机长期退化时停止完整HA声明。"]},"workflow":{"before":"运维分别配置复制、故障检测、切换和客户端，常在主恢复后忽略冗余。","now":"Agent可联合诊断拓扑/位点、配置授权机制并驱动故障测试；独立事务/单主/端点探针决定接收。","retained":["确定复制丢失/可用性取舍、fencing与切换授权，确认业务和冗余均已恢复。"],"added":["重建/追赶备机并验证读/复制服务，重新施加一个故障，核对单主和事务状态。"]},"humanResponsibility":"确定复制丢失/可用性取舍、fencing与切换授权，确认业务和冗余均已恢复。","failureHandling":"不能证明唯一主或链状态时停止扩大写入，保留证据并升级；不可用盲提升/双写试探。","example":{"input":"主库故障后需切备，旧主随后重启，客户端必须继续且不双主。","output":"事务/连接/单主证据、实际时间、重建副本和第二次故障结果。","failure":"新主可交易，旧主也写入且没有可用副本。"},"sourceIds":["WU-C-PG-HA","WU-C-DBA"],"evidenceLinks":[{"sourceId":"WU-C-PG-HA","supports":"定义单主与冗余的实际接收条件。","notCovered":"方法不是AI切换成绩。"},{"sourceId":"WU-C-DBA","supports":"有直接副本未恢复的失败边界。","notCovered":"受控场景不估生产频率，也不覆盖本拓扑。"}],"resourceRefs":[],"career":{"direction":"数据库高可用与复制","evidence":"复制/检测/隔离/切换与连接端点配置、运行说明","nextStep":"在隔离主备拓扑冻结事务/端点，测试主故障、网络分区、旧主回归与备机重建，再执行第二次切换。"},"team":{"bottleneck":"PostgreSQL18要求旧主隔离和新备机重建，原生DB不提供完整故障检测。DBA-Bench持续场景实际出现主事务恢复而副本未恢复，说明不能只验主库。","allocation":"Agent可联合诊断拓扑/位点、配置授权机制并驱动故障测试；独立事务/单主/端点探针决定接收。","nextStep":"在隔离主备拓扑冻结事务/端点，测试主故障、网络分区、旧主回归与备机重建，再执行第二次切换。"},"claims":[{"id":"C28-HA","statement":"PostgreSQL18说明故障切换须避免旧主重入形成双主，切换后需重建备机恢复冗余；检测/通知软件另配。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-PG-HA","locator":"26.3","relation":"method","supports":"定义单主与冗余的实际接收条件。","limit":"方法不是AI切换成绩。"}],"configuration":null,"progressIds":[],"limitation":"方法不是AI切换成绩。","evidenceStatus":"active"},{"id":"C28-PARTIAL","statement":"DBA-Bench持续PostgreSQL复制场景出现主库事务达标但副本服务未恢复，局部恢复不满足完整结果。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-C-DBA","locator":"6.6.1 Partial business outcomes","relation":"partial","supports":"有直接副本未恢复的失败边界。","limit":"受控场景不估生产频率，也不覆盖本拓扑。"}],"configuration":"v2持续评估：GPT5.5/Opus4.8/GLM5.1等，PostgreSQL/TiDB各5场景，每模型每场景1次；本主张仅该复制反例。","progressIds":["P-C-DBA-STATE"],"limitation":"受控场景不估生产频率，也不覆盖本拓扑。","evidenceStatus":"active"},{"id":"C28-ADOPT","statement":"可委派HA配置/故障演练，以单主、客户端事务和冗余恢复联合接收，不能只按提升命令成功。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-PG-HA","locator":"26.3 failover and standby recreation","relation":"partial","supports":"原生可检查机制支持演练。","limit":"需目标工具、拓扑和业务负载核验。"}],"configuration":null,"progressIds":[],"limitation":"需目标工具、拓扑和业务负载核验。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"原生复制/提升/重建接口公开，完整故障检测与fencing工具需按目标部署；本站未演练。","tools":[{"name":"PostgreSQL18 HA机制","url":"https://www.postgresql.org/docs/18/warm-standby-failover.html","version":"文档18；实际DB/检测/fencing/端点工具版本固定","access":"隔离主备/网络控制与业务客户端；完整检测工具另备"}]},"variants":[{"title":"复制延迟与丢失边界","inputDifference":"同步/异步与不同延迟条件。","assessmentChange":"按事先约定事务损失和等待/可用性接收。","acceptanceChange":"确认事务/位点、丢失/重复、追赶与中断时间。"},{"title":"分区与旧主回归","inputDifference":"主未真停机或随后恢复。","assessmentChange":"必须先证明单主与fencing，不能用提升掩盖双主。","acceptanceChange":"双方写入负例、客户端路由、重建副本及第二故障。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C09","domainId":"computing","title":"授权风险发现：威胁路径与可证实漏洞","goal":"在授权资产范围交付可追溯风险/漏洞发现及触发、影响和证据；缺证据的假设与未覆盖范围分开。","trigger":"冻结资产/接口/代码、身份、访问范围与禁止动作，给关键数据/业务目标和关注威胁。","scope":"结果是发现/风险判断，不包含修复、告警队列结案或事件遏制。威胁分析可交可反驳路径，已验证漏洞须给真实触发依据。","inputs":[{"name":"资产与授权","detail":"冻结资产/接口/代码、身份、访问范围与禁止动作，给关键数据/业务目标和关注威胁。"},{"name":"验证环境","detail":"提供隔离副本、版本/依赖、日志/配置、可重放输入及正常行为，不在授权外目标验证。"},{"name":"发现接收","detail":"定义需证实的输入→路径→影响与严重性条件，允许输出待证/误报排除；避免把理论可能写成已入侵。"}],"deliverables":["带位置/资产、触发前提和影响的发现/威胁路径","可重放或严格静态推导的证据、反例与误报排除","风险优先级、未查/不可验证范围及修复验证所需接口"],"acceptance":[{"criterion":"发现沿真实路径成立","method":"沿输入/身份到影响核对，静态发现给完整依赖前提；隔离重放与正常负例区分。","failure":"仅泛化风险标签或把普通登录失败说成入侵。"},{"criterion":"范围与影响相称","method":"核对版本、身份和数据/权限后果，保留反证与不适用条件。","failure":"靶场可利用就称生产一定可利用。"},{"criterion":"结果可接续","method":"另一接收者能重放/检查，发现与偏好/待证分开，修复或处置另接任务。","failure":"报告只有告警摘要，没有触发和证据。"}],"assessment":{"recommendation":"可委派风险路径搜索、受控验证和误报排除；按可证实发现接收，未发现不等于全安全。","conditions":["提供隔离副本、版本/依赖、日志/配置、可重放输入及正常行为，不在授权外目标验证。","定义需证实的输入→路径→影响与严重性条件，允许输出待证/误报排除；避免把理论可能写成已入侵。"],"evidenceBasis":"Cybench有受控攻击任务的直接实测，适合说明有限场景的执行能力；真实资产/威胁模型仍需环境与影响适配，不能从钓鱼产品推导通用核查。","localStatus":"本次核查公开资料与任务契约，未在本站运行目标任务或新增模型试次。","gaps":["在授权隔离资产中混入已知缺陷与正常替代，让Agent交可重放发现和排除，独立检查范围/误报/影响。"],"mode":"staged-delivery","rationale":"Cybench有受控攻击任务的直接实测，适合说明有限场景的执行能力；真实资产/威胁模型仍需环境与影响适配，不能从钓鱼产品推导通用核查。","frontier":"受控CTF和可执行环境已有研究；本资产的隐含依赖、真实攻击分布与未知漏洞查全率没有直接成绩。","nextValidation":"在授权隔离资产中混入已知缺陷与正常替代，让Agent交可重放发现和排除，独立检查范围/误报/影响。","futureSignals":["新资产/输入下有效发现可复验且误报确认时间可控，可扩大相近核查。","越范围验证、把待证当已利用或严重漏报时收窄采用。"]},"workflow":{"before":"分析师分开查资产、读代码、提出假设和手动复现。","now":"Agent可动态追上下文、构造隔离输入并据反证修正，独立接收者核对影响与边界。","retained":["确定资产授权、业务影响与风险接受，批准有影响验证；不自动合并到整改/处置完成。"],"added":["另一接收者能重放/检查，发现与偏好/待证分开，修复或处置另接任务。"]},"humanResponsibility":"确定资产授权、业务影响与风险接受，批准有影响验证；不自动合并到整改/处置完成。","failureHandling":"缺环境或证据时交待证假设与具体所需材料；不越权或破坏证据求证。","example":{"input":"租户读取接口疑似越权，需确定真实数据访问路径。","output":"两身份/数据对照、触发到影响证据、修复接收条件。","failure":"仅根据函数名报越权，未核对实际鉴权链。"},"sourceIds":["WU-C-CYBENCH","WU-C-IR-NIST"],"evidenceLinks":[{"sourceId":"WU-C-CYBENCH","supports":"有直接安全任务实验。","notCovered":"任务族、历史模型与靶场条件限定；不声称实际系统已被检查。"},{"sourceId":"WU-C-IR-NIST","supports":"提供威胁建模、攻击面/路径与影响分析的接收方法。","notCovered":"风险管理建议，不是模型完成威胁建模的成绩。"}],"resourceRefs":[],"career":{"direction":"安全风险与漏洞核查","evidence":"带位置/资产、触发前提和影响的发现/威胁路径","nextStep":"在授权隔离资产中混入已知缺陷与正常替代，让Agent交可重放发现和排除，独立检查范围/误报/影响。"},"team":{"bottleneck":"Cybench有受控攻击任务的直接实测，适合说明有限场景的执行能力；真实资产/威胁模型仍需环境与影响适配，不能从钓鱼产品推导通用核查。","allocation":"Agent可动态追上下文、构造隔离输入并据反证修正，独立接收者核对影响与边界。","nextStep":"在授权隔离资产中混入已知缺陷与正常替代，让Agent交可重放发现和排除，独立检查范围/误报/影响。"},"claims":[{"id":"C09-CTF","statement":"Cybench在命令可执行的隔离CTF中观察到Agent解决部分完整任务；这是受控攻击任务证据，不能直接评价企业防守。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-C-CYBENCH","locator":"Abstract；agent scaffolds与unguided tasks","relation":"partial","supports":"有直接安全任务实验。","limit":"任务族、历史模型与靶场条件限定；不声称实际系统已被检查。"}],"configuration":"v4：40个CTF、4竞赛、8模型与多脚手架；有/无子任务提示分开，本文不转用分数。","progressIds":[],"limitation":"任务族、历史模型与靶场条件限定；不声称实际系统已被检查。","evidenceStatus":"active"},{"id":"C09-ADOPT","statement":"可以委派授权发现/验证任务，但以实际路径、前提、影响和反证接收，不将CTF结果外推完整企业安全。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-CYBENCH","locator":"任务环境与unguided task边界","relation":"partial","supports":"受控执行提供邻近验证依据。","limit":"真实威胁/漏洞仍需目标输入和独立影响核验。"}],"configuration":null,"progressIds":[],"limitation":"真实威胁/漏洞仍需目标输入和独立影响核验。","evidenceStatus":"active"},{"id":"C09-THREAT-METHOD","statement":"NIST将资产漏洞、内外威胁、潜在影响与攻击向量/攻击面/横向路径纳入风险识别；这些方法可用于检查威胁路径交付是否具有明确前提和影响。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-IR-NIST","locator":"Table2：ID.RA-01、ID.RA-03、ID.RA-04、ID.RA-05","relation":"method","supports":"提供威胁建模、攻击面/路径与影响分析的接收方法。","limit":"风险管理建议，不是模型完成威胁建模的成绩。"}],"configuration":null,"progressIds":[],"limitation":"资产/信任边界与威胁输入需实际核实；文档方法不证明AI遗漏率。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开靶场可检查；目标资产/代码与验证权限另备，不包含钓鱼或生产响应服务。","tools":[{"name":"Cybench","url":"https://cybench.github.io/","version":"论文v4与具体challenge镜像/版本固定","access":"公开靶场；仅隔离授权目标与自备模型"}]},"variants":[{"title":"威胁与攻击路径","inputDifference":"结果为资产/信任边界/攻击前提与风险路径，不一定可执行利用。","assessmentChange":"可交可反驳路径，未验证部分明确列出。","acceptanceChange":"核对资产/身份/边界、成立前提、反证与影响优先级。"},{"title":"授权漏洞验证","inputDifference":"候选缺陷可在隔离目标重放。","assessmentChange":"以触发/真实影响接收，修复另交W-C29或W-C01。","acceptanceChange":"版本/身份/输入、正反例、证据位置与误报排除。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C29","domainId":"computing","title":"安全控制变更：风险降低且合法业务仍成立","goal":"交付一个限定风险的代码/配置/访问/检测控制变更，并用原风险、绕过和正常行为证明有效及可回退。","trigger":"给已证实风险/明确策略、资产/角色、期望允许/拒绝行为和不可牺牲的业务。","scope":"产物是生效的控制变更，不是发现列表或告警结论；治理可组合这些实际控制及整改跟踪。风险接受与合规认定另有组织依据。","inputs":[{"name":"控制目标","detail":"给已证实风险/明确策略、资产/角色、期望允许/拒绝行为和不可牺牲的业务。"},{"name":"变更环境","detail":"固定代码/配置/身份/规则、目标系统权限与旧状态、可逆测试环境，明确授权动作。"},{"name":"独立测试","detail":"准备原风险/合法行为/绕过和规则漂移控制，定义风险/业务/资源接收及回退。"}],"deliverables":["可审查并可部署的控制补丁/配置与实际差异","原风险被拒绝、正常业务被允许及绕过/副作用结果","版本/审计、回退与后续漂移复核和未覆盖风险"],"acceptance":[{"criterion":"控制达到具体目标","method":"独立重放原路径/策略，直接查询真实权限或业务状态。","failure":"仅配置文件存在，没有生效检查。"},{"criterion":"正常与绕过都检查","method":"同角色/对象的正反矩阵、替代路径、错误配置和资源副作用控制。","failure":"全拒绝让风险消失，合法业务也坏了。"},{"criterion":"变更可维护","method":"核对版本/审计与回退后状态；用漂移反例验证持续复核入口。","failure":"临时手工改动不在版本中，重建后控制丢失。"}],"assessment":{"recommendation":"可委派限定控制实现、获准部署与验证，按风险降低/合法行为和回退接收；不能由安全产品接入文档证明效果。","conditions":["固定代码/配置/身份/规则、目标系统权限与旧状态、可逆测试环境，明确授权动作。","准备原风险/合法行为/绕过和规则漂移控制，定义风险/业务/资源接收及回退。"],"evidenceBasis":"NIST53A把检查对象与测试结合，kubectl提供可执行授权查询；AI整合控制变更属于方法基础上的迁移，不是独立性能成绩。","localStatus":"本次核查公开资料与任务契约，未在本站运行目标任务或新增模型试次。","gaps":["选一项访问/配置风险，冻结允许/拒绝矩阵，让Agent交变更，再独立测试原风险、合法路径、绕过及回退。"],"mode":"staged-delivery","rationale":"NIST53A把检查对象与测试结合，kubectl提供可执行授权查询；AI整合控制变更属于方法基础上的迁移，不是独立性能成绩。","frontier":"有公开控制评估方法和权限检查接口；目标控制、绕过及持续漂移没有本站AI试验。","nextValidation":"选一项访问/配置风险，冻结允许/拒绝矩阵，让Agent交变更，再独立测试原风险、合法路径、绕过及回退。","futureSignals":["新角色/对象变化下风险控制和合法业务仍可复验，可扩大同类变更。","依靠全拒绝、扩大其他权限或绕过未测时拒绝接收。"]},"workflow":{"before":"安全指出风险、工程改配置、运维核实生效，结果常分散。","now":"Agent可联合实现、部署到授权环境并重测，程序检查正反矩阵，责任人核验风险与业务取舍。","retained":["确定策略和授权、业务容忍与风险接受，合规结论依据真实组织要求。"],"added":["核对版本/审计与回退后状态；用漂移反例验证持续复核入口。"]},"humanResponsibility":"确定策略和授权、业务容忍与风险接受，合规结论依据真实组织要求。","failureHandling":"验证失败回退或交明确未生效原因；不以关闭告警或禁用全部功能充作整改。","example":{"input":"服务账号权限过宽，目标仅能读指定namespace。","output":"最小权限补丁、生效授权矩阵、正常请求/越权拒绝与回退。","failure":"删掉全部权限导致业务中断，但报告说风险已消除。"},"sourceIds":["WU-C-CONTROL-NIST","WU-C-K8S-AUTH"],"evidenceLinks":[{"sourceId":"WU-C-CONTROL-NIST","supports":"控制验收方法基础。","notCovered":"专业方法不是模型成绩。"},{"sourceId":"WU-C-K8S-AUTH","supports":"实际权限查询接口。","notCovered":"单一权限查询不排除所有业务/网络绕过。"}],"resourceRefs":[],"career":{"direction":"安全控制工程","evidence":"可审查并可部署的控制补丁/配置与实际差异","nextStep":"选一项访问/配置风险，冻结允许/拒绝矩阵，让Agent交变更，再独立测试原风险、合法路径、绕过及回退。"},"team":{"bottleneck":"NIST53A把检查对象与测试结合，kubectl提供可执行授权查询；AI整合控制变更属于方法基础上的迁移，不是独立性能成绩。","allocation":"Agent可联合实现、部署到授权环境并重测，程序检查正反矩阵，责任人核验风险与业务取舍。","nextStep":"选一项访问/配置风险，冻结允许/拒绝矩阵，让Agent交变更，再独立测试原风险、合法路径、绕过及回退。"},"claims":[{"id":"C29-METHOD","statement":"NIST53A以检查、访谈与测试及明确评估对象核实控制；实际允许/拒绝行为应独立验证。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-CONTROL-NIST","locator":"assessment methods and objects","relation":"method","supports":"控制验收方法基础。","limit":"专业方法不是模型成绩。"}],"configuration":null,"progressIds":[],"limitation":"专业方法不是模型成绩。","evidenceStatus":"active"},{"id":"C29-AUTH","statement":"kubectl auth can-i公开提供授权检查，可为K8s角色/资源矩阵提供一项生效证据。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-K8S-AUTH","locator":"command synopsis/examples","relation":"method","supports":"实际权限查询接口。","limit":"单一权限查询不排除所有业务/网络绕过。"}],"configuration":null,"progressIds":[],"limitation":"单一权限查询不排除所有业务/网络绕过。","evidenceStatus":"active"},{"id":"C29-ADOPT","statement":"可委派限定控制变更，以独立风险/正常/绕过/回退结果接收，其他控制类型须配置对应测试。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-CONTROL-NIST","locator":"assessment planning/testing","relation":"partial","supports":"可检查对象/测试支持目标适配。","limit":"无本组织AI控制有效率。"}],"configuration":null,"progressIds":[],"limitation":"无本组织AI控制有效率。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"有公开权限检查工具；代码、网络与其他控制需目标配置/独立测试接口，没有本站部署。","tools":[{"name":"kubectl auth can-i","url":"https://kubernetes.io/docs/reference/kubectl/generated/kubectl_auth/kubectl_auth_can-i/","version":"实际K8s/kubectl/角色版本冻结","access":"公开工具；需目标集群授权，不能据此声称有权修改"}]},"variants":[{"title":"访问与配置","inputDifference":"结果是角色/网络/系统配置生效。","assessmentChange":"按具体允许/拒绝矩阵和业务回归接收。","acceptanceChange":"实际身份/对象/旁路、漂移和回退。"},{"title":"代码与检测控制","inputDifference":"结果是漏洞修复或规则生效。","assessmentChange":"组合W-C01/C12的复现与验收方法，不继承其他工具成绩。","acceptanceChange":"原风险、合法输入、误报漏报/绕过及规则版本。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C30","domainId":"computing","title":"告警调查结论：证据、判定与可继续的升级包","goal":"针对给定告警/队列交付可追溯的真假/待证判定、影响/优先级和升级依据，不默认改变资产状态。","trigger":"给邮件/身份/端点等告警、资产/用户、时间窗与需回答的真实风险/优先级。","scope":"结果为调查结论与队列排序；检测规则变更另W-C29，安全事件遏制另W-C31。钓鱼RCT仅支持其问卷队列条件，其他告警为明确方法迁移。","inputs":[{"name":"告警与问题","detail":"给邮件/身份/端点等告警、资产/用户、时间窗与需回答的真实风险/优先级。"},{"name":"证据与基线","detail":"冻结日志/消息/上下文及获取权限，保存原始证据与真/假/待证参照，恶意内容只在隔离环境处理。"},{"name":"判定与升级","detail":"给处置代价/漏报成本、证据不足标准、升级对象与所需诊断字段；区分判定和执行。"}],"deliverables":["逐告警证据定位、真假/待证及影响前提","优先级/关联时间线、反证与误报排除","仍需信息、升级包和后续控制/处置的明确接口"],"acceptance":[{"criterion":"判定由证据支持","method":"原始消息/身份/时间线与资产交叉核对，正常负例和替代解释挑战结论。","failure":"仅品牌词/风险标签就认恶意。"},{"criterion":"未知和错误可见","method":"用本地真/假与有意错误解释检查自动同意，保留缺日志/矛盾与查不到部分。","failure":"Agent判良性就无条件关单。"},{"criterion":"输出可接续且不冒充处置","method":"独立接收者能重查证据和升级，结论/规则/执行状态分开记录。","failure":"给出恶意判定便写资产已经隔离/清除。"}],"assessment":{"recommendation":"可委派证据聚合、排序和调查结论，先以本地真假/待证控制校准；高影响处置从调查结果另开实际授权交付。","conditions":["冻结日志/消息/上下文及获取权限，保存原始证据与真/假/待证参照，恶意内容只在隔离环境处理。","给处置代价/漏报成本、证据不足标准、升级对象与所需诊断字段；区分判定和执行。"],"evidenceBasis":"钓鱼RCT提供特定队列直接研究，NIST61r3给检测/分析与响应边界；其他告警不借钓鱼实验冒充同配置成绩。","localStatus":"本次核查公开资料与任务契约，未在本站运行目标任务或新增模型试次。","gaps":["用本地脱敏真假/待证告警盲测，核对逐条证据、漏报/误报与确认时间，再决定优先级和复核范围。"],"mode":"staged-delivery","rationale":"钓鱼RCT提供特定队列直接研究，NIST61r3给检测/分析与响应边界；其他告警不借钓鱼实验冒充同配置成绩。","frontier":"邮件材料/问卷的排序与解释效应已有试验；原语料无自然FP/FN、合成错误和反事实协议限制迁移，不能证明实际邮箱长期无人研判。","nextValidation":"用本地脱敏真假/待证告警盲测，核对逐条证据、漏报/误报与确认时间，再决定优先级和复核范围。","futureSignals":["新队列基率/对抗条件下证据与误报漏报仍可控，可扩大辅助研判。","合成控制或真实错误被盲目附和、证据不足仍关单时收窄范围。"]},"workflow":{"before":"分析师逐条查告警并在多处整理证据和升级。","now":"Agent自主关联证据、排序和构造反证，程序重放已知控制；调查结论独立于外部处置。","retained":["定义调查授权、风险/漏报代价和升级优先级，决定真实事件处置范围。"],"added":["独立接收者能重查证据和升级，结论/规则/执行状态分开记录。"]},"humanResponsibility":"定义调查授权、风险/漏报代价和升级优先级，决定真实事件处置范围。","failureHandling":"缺证据输出待证及具体请求，不用恶意内容授予工具权限，不越过调查范围执行隔离。","example":{"input":"可疑邮件和异常登录同时出现，但也有正常内部转发。","output":"逐条真假/待证、消息/身份证据、时间线和升级条件。","failure":"把排序靠前等同已入侵，并声称自动清除。"},"sourceIds":["WU-C-PHISH","WU-C-IR-NIST"],"evidenceLinks":[{"sourceId":"WU-C-PHISH","supports":"直接邮件研判研究，保留实际载体与干预身份。","notCovered":"原语料零FP/FN后注入合成判定错误，跳过良性协议为反事实；不能推断真实误报分布或无人遏制。"},{"sourceId":"WU-C-IR-NIST","supports":"专业方法支持调查结果分界。","notCovered":"非邮件任务没有由RCT直接证明的AI表现。"}],"resourceRefs":[],"career":{"direction":"安全告警调查","evidence":"逐告警证据定位、真假/待证及影响前提","nextStep":"用本地脱敏真假/待证告警盲测，核对逐条证据、漏报/误报与确认时间，再决定优先级和复核范围。"},"team":{"bottleneck":"钓鱼RCT提供特定队列直接研究，NIST61r3给检测/分析与响应边界；其他告警不借钓鱼实验冒充同配置成绩。","allocation":"Agent自主关联证据、排序和构造反证，程序重放已知控制；调查结论独立于外部处置。","nextStep":"用本地脱敏真假/待证告警盲测，核对逐条证据、漏报/误报与确认时间，再决定优先级和复核范围。"},"claims":[{"id":"C30-PHISH","statement":"钓鱼RCT在标准化材料/问卷中的邮件队列比较排序和解释对专业分析师的影响，并另作resolve-benign反事实；不是持续生产邮箱无人处置。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-C-PHISH","locator":"v1 §1—§3 methodology/behavior/counterfactuals","relation":"partial","supports":"直接邮件研判研究，保留实际载体与干预身份。","limit":"原语料零FP/FN后注入合成判定错误，跳过良性协议为反事实；不能推断真实误报分布或无人遏制。"}],"configuration":"v1 §2：招募167名专业分析师，13低分回答被剔除；93封原语料中抽25邮件队列，在Qualtrics以标准化只读材料判定。Control52随机无AI、Aware56优先队列+判定/解释、Blind59同排序但不告知AI；原语料Agent零FP/FN，另造判定/解释错误，每队列1SFP+1SFN。resolve-benign为由任务结果组合的反事实分析。","progressIds":["P-C-PHISH-RCT"],"limitation":"原语料零FP/FN后注入合成判定错误，跳过良性协议为反事实；不能推断真实误报分布或无人遏制。","evidenceStatus":"active"},{"id":"C30-INVESTIGATE","statement":"可用NIST61r3的检测/分析与证据要求组织其他告警调查，以独立证据和待证/升级条件接收。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-IR-NIST","locator":"Table3 detection/analysis","relation":"partial","supports":"专业方法支持调查结果分界。","limit":"非邮件任务没有由RCT直接证明的AI表现。"}],"configuration":null,"progressIds":[],"limitation":"非邮件任务没有由RCT直接证明的AI表现。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"商业钓鱼工具受许可/角色/容量约束；本包没有进入租户。其他告警需目标日志/调查接口。","tools":[{"name":"Phishing Triage Agent","url":"https://learn.microsoft.com/en-us/defender-xdr/phishing-triage-agent","version":"2026-10-20核查文档；租户实际部署版本需记录","access":"需相应Microsoft安全产品权限/许可与SCU容量"}]},"variants":[{"title":"钓鱼队列","inputDifference":"邮件/URL/附件材料与本地真假基率。","assessmentChange":"按本地错误控制校准，不照搬问卷反事实收益。","acceptanceChange":"真实/合成错误解释、恶意/良性/待证、排序和确认成本。"},{"title":"身份端点及关联告警","inputDifference":"结果是跨来源调查结论，非邮件分类。","assessmentChange":"NIST方法迁移，先给可证伪证据链；不继承钓鱼成绩。","acceptanceChange":"时间/身份/资产对应、替代解释、缺日志与升级包。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C31","domainId":"computing","title":"安全事件处置：证据保全、遏制与业务恢复","goal":"在明确授权事件范围内交付证据保全、有效遏制/清除和合法业务恢复的真实状态，保留攻击持续性与未解决风险。","trigger":"给已确认/疑似事件、资产/身份、攻击线索、业务重要性、处置范围与不能毁损证据/业务的约束。","scope":"结果为事件处置状态，不是告警判定或一般服务恢复；遏制、证据和业务恢复可阶段接收，但完整完成必须联合满足。","inputs":[{"name":"事件与目标","detail":"给已确认/疑似事件、资产/身份、攻击线索、业务重要性、处置范围与不能毁损证据/业务的约束。"},{"name":"证据和授权工具","detail":"冻结原始日志/镜像/时间线及保全要求，提供获准端点/身份/网络处置与恢复工具，动作留审计。"},{"name":"独立终态","detail":"准备攻击持续性/正常业务/侧向影响探针、恢复基线、回滚/升级和实际沟通/通知权限。"}],"deliverables":["证据清单/保全与事件时间线、实际处置动作审计","遏制/清除有效、攻击持续性和侧向影响检查","合法业务/身份/数据恢复、残余风险与后续观察/升级记录"],"acceptance":[{"criterion":"证据未被处置破坏","method":"按组织要求保存原始证据、时间/完整性/访问和动作记录，区分可逆隔离与不可逆清除。","failure":"删消息/日志清理痕迹后宣告解决。"},{"criterion":"攻击与影响受控","method":"在授权环境检查原攻击路径/持续性、侧向行为与隔离边界，不以告警消失作证。","failure":"停全部服务但攻击凭据仍有效。"},{"criterion":"合法业务真的恢复","method":"从合法用户/数据状态验证恢复，审计旧凭据/端点与后续观察；残余风险/未恢复明示。","failure":"只让服务启动，用户/数据未恢复或再入侵未检查。"}],"assessment":{"recommendation":"可委派证据整合与获准可逆处置，按保全、攻击遏制和合法业务恢复阶段接收；不可逆动作与通知有明确权限。","conditions":["冻结原始日志/镜像/时间线及保全要求，提供获准端点/身份/网络处置与恢复工具，动作留审计。","准备攻击持续性/正常业务/侧向影响探针、恢复基线、回滚/升级和实际沟通/通知权限。"],"evidenceBasis":"NIST61r3明确检测、响应、恢复及事件证据要求，Defender公开实际处置/调查包接口；这是工具/专业方法上的AI迁移，不是自主处置成功率。","localStatus":"本次核查公开资料与任务契约，未在本站运行目标任务或新增模型试次。","gaps":["在授权演练环境放入已知事件和正常业务，先保全证据再执行获准隔离/恢复，独立验证持续性、数据/用户与审计。"],"mode":"staged-delivery","rationale":"NIST61r3明确检测、响应、恢复及事件证据要求，Defender公开实际处置/调查包接口；这是工具/专业方法上的AI迁移，不是自主处置成功率。","frontier":"有公开响应接口与专业方法；CTF和钓鱼问卷均不证明本端点/身份事件已安全处置，没有本站真实攻击试验。","nextValidation":"在授权演练环境放入已知事件和正常业务，先保全证据再执行获准隔离/恢复，独立验证持续性、数据/用户与审计。","futureSignals":["新事件条件下证据、遏制与业务恢复都被独立复验，可扩大受限自动动作。","证据丢失、攻击持续或业务未恢复时拒绝完整结案。"]},"workflow":{"before":"调查、隔离和恢复分散交接，告警关闭可能掩盖攻击或业务未恢复。","now":"Agent可联合证据/状态、执行明确授权动作并持续验证，程序读真实终态；未知或超权限及时升级。","retained":["确定事件范围、保全/法律/沟通要求、隔离代价与残余风险接受，授权不可逆动作。"],"added":["从合法用户/数据状态验证恢复，审计旧凭据/端点与后续观察；残余风险/未恢复明示。"]},"humanResponsibility":"确定事件范围、保全/法律/沟通要求、隔离代价与残余风险接受，授权不可逆动作。","failureHandling":"证据/权限不足先保留状态与升级包；不能破坏性试探或把调查结论写成已处置。","example":{"input":"疑似盗用账号访问端点，需保留证据并恢复受影响用户工作。","output":"证据/时间线、受限凭据/端点处置、持续性负例和合法用户恢复。","failure":"清日志并禁用所有账号，既没证明清除也没恢复业务。"},"sourceIds":["WU-C-IR-NIST","WU-C-EDR-ACTIONS"],"evidenceLinks":[{"sourceId":"WU-C-IR-NIST","supports":"支持联合处置结果和证据边界。","notCovered":"专业建议不是AI实验，也不代替具体组织授权。"},{"sourceId":"WU-C-EDR-ACTIONS","supports":"有实际产品动作入口。","notCovered":"需目标产品/角色/许可；工具存在不证明事件已处置。"}],"resourceRefs":[],"career":{"direction":"安全事件响应","evidence":"证据清单/保全与事件时间线、实际处置动作审计","nextStep":"在授权演练环境放入已知事件和正常业务，先保全证据再执行获准隔离/恢复，独立验证持续性、数据/用户与审计。"},"team":{"bottleneck":"NIST61r3明确检测、响应、恢复及事件证据要求，Defender公开实际处置/调查包接口；这是工具/专业方法上的AI迁移，不是自主处置成功率。","allocation":"Agent可联合证据/状态、执行明确授权动作并持续验证，程序读真实终态；未知或超权限及时升级。","nextStep":"在授权演练环境放入已知事件和正常业务，先保全证据再执行获准隔离/恢复，独立验证持续性、数据/用户与审计。"},"claims":[{"id":"C31-IR","statement":"NIST61r3将检测、响应/遏制/根除与恢复联系，并要求按组织政策保全事件证据；终态不能只有告警关闭。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-IR-NIST","locator":"Table3；证据保留与响应/恢复","relation":"method","supports":"支持联合处置结果和证据边界。","limit":"专业建议不是AI实验，也不代替具体组织授权。"}],"configuration":null,"progressIds":[],"limitation":"专业建议不是AI实验，也不代替具体组织授权。","evidenceStatus":"active"},{"id":"C31-EDR","statement":"Defender for Endpoint公开设备隔离、调查包等响应接口与条件，可承接获准动作和证据。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-EDR-ACTIONS","locator":"device response actions/investigation package","relation":"method","supports":"有实际产品动作入口。","limit":"需目标产品/角色/许可；工具存在不证明事件已处置。"}],"configuration":null,"progressIds":[],"limitation":"需目标产品/角色/许可；工具存在不证明事件已处置。","evidenceStatus":"active"},{"id":"C31-ADOPT","statement":"可以委派受限事件处置循环，但以证据、攻击持续性及合法业务恢复联合接收，不能继承CTF/钓鱼研判成绩。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-IR-NIST","locator":"Response/Recovery","relation":"partial","supports":"明确专业终态使验证可执行。","limit":"真实事件和高影响动作需目标环境/授权。"}],"configuration":null,"progressIds":[],"limitation":"真实事件和高影响动作需目标环境/授权。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"有条件商业响应工具；其他EDR/SIEM/身份平台需实际接口，本文没有自动执行或复现。","tools":[{"name":"Defender for Endpoint响应接口","url":"https://learn.microsoft.com/en-us/defender-endpoint/respond-machine-alerts","version":"2026-10-20核查；目标设备/产品/角色版本另固定","access":"商业产品、设备加入及响应权限/许可；未在本站启用"}]},"variants":[{"title":"端点与身份事件","inputDifference":"涉及端点隔离、账号/凭据与用户恢复。","assessmentChange":"先明确证据/隔离/恢复边界再执行获准动作。","acceptanceChange":"持续性、旧凭据、合法用户/数据和端点恢复。"},{"title":"跨资产与业务恢复","inputDifference":"多系统、网络或第三方参与，可能超现有权限。","assessmentChange":"可分阶段交结果，缺责任方/日志不能完整结案。","acceptanceChange":"证据时间一致、侧向影响、业务探针、沟通/升级与残余风险。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-C32","domainId":"computing","title":"研究材料交付：陈述到证据与可复用成果","goal":"针对实际已有研究结果交付陈述、证明或实验相互对应的论文/报告与复用材料；按理论、代码或数据产物选择核查方式，不强制纯理论成果制造代码或数据。","trigger":"冻结真实研究问题、已得到的证明/实验/失败、原始产物与文献；明确受众、打包用途和可公开边界。","scope":"对象可以是原创或复现成果的传播/复用材料；不在此重新宣称发现或复现成功。投稿/公开发送需相应授权，材料可本地接收。","inputs":[{"name":"已有结果与受众","detail":"冻结真实研究问题、已得到的证明/实验/失败、原始产物与文献；明确受众、打包用途和可公开边界。"},{"name":"复用条件","detail":"给出实际产物及其获取、引用和许可条件；有代码、数据、权重或硬件时再提供依赖、预算与隐私条件，纯理论给完整命题、假设、证明和必要引文。"},{"name":"陈述核对与外部复算","detail":"关键陈述连接证明、引文或真实运行/数据/图表；理论按假设与论证核查，有计算资产时准备独立环境与最小复算结果；区分新颖性、意义和可复用性。"}],"deliverables":["论文/报告和逐关键陈述的证明/运行/引用定位","理论成果的完整命题、假设、证明/引文材料；有代码或数据时另附可运行环境、数据字典/版本与获取/许可条件","与产物相适用的独立理论核查或运行复算记录，以及限制、负结果和无法公开项说明"],"acceptance":[{"criterion":"陈述不扩大结果","method":"每项关键结论反查证明/日志/数据与引用，不把未跑、探索或部分结果写成确认/完整成功。","failure":"正文声称所有实验通过，实际只运行一组。"},{"criterion":"材料确可复用","method":"理论成果核对完整命题、假设、证明与引文是否足以复查；选择形式交付时追加内核及依赖检查。有代码/数据时才在干净环境获取、安装并复算核心结果，核对版本、图表与预算。","failure":"理论结论缺关键假设、证明或引文；或声称可运行的仓库缺数据/脚本，图表无法重建。"},{"criterion":"权利与限制透明","method":"核对许可/引用/敏感数据与不可公开边界，说明失败/限制/人介入和适用范围。","failure":"用占位许可证或泄露数据，隐藏复用必需私有资产。"}],"assessment":{"recommendation":"可委派真实成果的材料整理与打包；理论按命题、假设、证明和引文接收，有代码/数据时再独立运行复算，发表新颖性与意义另审。","conditions":["给出实际产物及其获取、引用和许可条件；有代码、数据、权重或硬件时再提供依赖、预算与隐私条件，纯理论给完整命题、假设、证明和必要引文。","关键陈述连接证明、引文或真实运行/数据/图表；理论按假设与论证核查，有计算资产时准备独立环境与最小复算结果；区分新颖性、意义和可复用性。"],"evidenceBasis":"NeurIPS清单提供陈述、限制、复现及代码/数据权利检查方法；AI写作/打包建议为编辑迁移，不用PaperBench部分复现成绩证明科研传播。","localStatus":"本次核查公开资料与任务契约，未在本站运行目标任务或新增模型试次。","gaps":["分别给一份理论结果和一份实验结果：理论注入漏假设/缺引文，实验保留未运行项/缺依赖，检查材料是否如实标注；只有计算产物需要独立环境复算。"],"mode":"staged-delivery","rationale":"NeurIPS清单提供陈述、限制、复现及代码/数据权利检查方法；AI写作/打包建议为编辑迁移，不用PaperBench部分复现成绩证明科研传播。","frontier":"公开有研究材料与复现自查要求；没有本成果传播效果或AI论文可靠率。可检查的包与外部复算比语言流畅更直接。","nextValidation":"分别给一份理论结果和一份实验结果：理论注入漏假设/缺引文，实验保留未运行项/缺依赖，检查材料是否如实标注；只有计算产物需要独立环境复算。","futureSignals":["外部环境复算且关键陈述均能回溯真实证据，可扩大同类材料整理。","陈述超证据、关键材料缺失或许可不明时拒绝发布/复用声明。"]},"workflow":{"before":"研究完成后另写论文、整理仓库，图表和陈述易与真实记录分离。","now":"Agent可把真实产物连到正文、证明、图表和引用；理论按完整命题与论证反馈修订，有代码/数据时再编排环境并独立复算，缺证据就缩小陈述。","retained":["确认研究贡献、署名/许可、隐私和公开授权，独立裁决新颖性/意义，不让材料生成替代研究。"],"added":["核对许可/引用/敏感数据与不可公开边界，说明失败/限制/人介入和适用范围。"]},"humanResponsibility":"确认研究贡献、署名/许可、隐私和公开授权，独立裁决新颖性/意义，不让材料生成替代研究。","failureHandling":"缺原始产物只交缺口与可写范围，不编结果/引用；敏感或权利不明材料保持本地并说明获取条件。","example":{"input":"算法已有保留图实验与失败，需打包论文/代码/数据说明。","output":"关键结论到结果/证明，环境和独立复算日志，失败与许可。","failure":"把探索图改成确认结论，并发布无法复算的图表。"},"sourceIds":["WU-C-PAPER-CHECKLIST"],"evidenceLinks":[{"sourceId":"WU-C-PAPER-CHECKLIST","supports":"有公开研究材料检查方法。","notCovered":"作者清单不是AI写作成绩，也不保证录用。"}],"resourceRefs":[],"career":{"direction":"研究传播与可复用工程","evidence":"论文/报告和逐关键陈述的证明/运行/引用定位","nextStep":"分别给一份理论结果和一份实验结果：理论注入漏假设/缺引文，实验保留未运行项/缺依赖，检查材料是否如实标注；只有计算产物需要独立环境复算。"},"team":{"bottleneck":"NeurIPS清单提供陈述、限制、复现及代码/数据权利检查方法；AI写作/打包建议为编辑迁移，不用PaperBench部分复现成绩证明科研传播。","allocation":"Agent可把真实产物连到正文、证明、图表和引用；理论按完整命题与论证反馈修订，有代码/数据时再编排环境并独立复算，缺证据就缩小陈述。","nextStep":"分别给一份理论结果和一份实验结果：理论注入漏假设/缺引文，实验保留未运行项/缺依赖，检查材料是否如实标注；只有计算产物需要独立环境复算。"},"claims":[{"id":"C32-MATERIAL","statement":"NeurIPS Paper Checklist要求检查主张/限制、复现和代码/数据等信息，适用于把实际结果与可复用材料相连。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-C-PAPER-CHECKLIST","locator":"Claims/Limitations/Reproducibility/Code and Data","relation":"method","supports":"有公开研究材料检查方法。","limit":"作者清单不是AI写作成绩，也不保证录用。"}],"configuration":null,"progressIds":[],"limitation":"作者清单不是AI写作成绩，也不保证录用。","evidenceStatus":"active"},{"id":"C32-ADOPT","statement":"可委派真实成果的写作/打包，以逐陈述证据、相适用的理论核查或独立复算及许可/限制接收；不把已有论文复现任务等同原创成果传播。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-C-PAPER-CHECKLIST","locator":"claim consistency and reproducibility","relation":"partial","supports":"方法使材料验收可执行。","limit":"本成果外部复算尚待实施，传播影响未知。"}],"configuration":null,"progressIds":[],"limitation":"本成果外部复算尚待实施，传播影响未知。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开方法与普通文档/代码工具即可组织材料；关键数据、证明、运行资产和权限必须真实可得。","tools":[{"name":"研究材料检查清单","url":"https://neurips.cc/public/guides/PaperChecklist","version":"访问版2026-10-20；目标论文/代码/环境版本固定","access":"公开方法；实际写作/构建/复算工具由成果环境提供"}]},"variants":[{"title":"原创研究传播","inputDifference":"对象为本研究新算法/模型/理论/系统的已有结果。","assessmentChange":"先核对贡献及证据，材料包不替代新颖性审阅。","acceptanceChange":"关键陈述对应完整证明或真实实验；理论按假设、论证和引用核查，有代码/数据时再运行复算，权利和限制均要明确。"},{"title":"受限资产与复用","inputDifference":"代码/数据/硬件存在许可、隐私或私有条件。","assessmentChange":"可交诚实限制/获取接口，不伪装完全开放。","acceptanceChange":"可公开最小材料、授权获取、依赖与外部可复算范围。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-A04","domainId":"analysis","title":"探索分析交付：分布、关系与时间变化的可重放诊断","goal":"针对明确分析问题交付可重放的分布/异常/关系/时间探索和备选解释，保留反证与尚不能归因部分。","trigger":"明确决策问题、对象/粒度、指标定义、时间/单位与数据生成/采样机制，提供原始表/日志及版本。","scope":"不限组成率分解；W-A02只承担其特定组成贡献。探索可生成假设，但不自动给因果或确认性推断，实验/识别另W-A03。","inputs":[{"name":"问题与记录","detail":"明确决策问题、对象/粒度、指标定义、时间/单位与数据生成/采样机制，提供原始表/日志及版本。"},{"name":"分析约束","detail":"冻结过滤/缺失/异常/重复规则、分组与时间粒度、数据许可/预算和可用代码/SQL环境。"},{"name":"独立检查","detail":"准备关键汇总/图的独立复算、合理替代与数据错误控制；保留探索路径和挑选过程。"}],"deliverables":["可重跑查询/程序、数据核对与分析口径","分布/尾部/异常、变量关系和时间图表与底层结果","备选假设/反证、稳健性/选择影响及下一项可检验问题"],"acceptance":[{"criterion":"探索对应同一总体","method":"独立核对粒度、单位/时间、采样、缺失/重复和过滤后的总量，异常不静默删除。","failure":"毫秒/秒混用或重复行制造长尾。"},{"criterion":"发现可复算且不过度挑选","method":"从程序生成关键分位/关系/时间图，用不同合理过滤/分组/时间窗挑战，记录探索选择。","failure":"只展示最显著关系或平滑掉变化。"},{"criterion":"解释守住识别边界","method":"列至少一个竞争解释与可否定检查，分开统计关联、机制假设和有证据因果。","failure":"看到发布后尾延迟增大就断言发布造成。"}],"assessment":{"recommendation":"可整包委派数据诊断、探索与解释，固定分析后直接程序重跑；以结果/反证接收，不用一个漂亮故事或组成分解覆盖全部探索。","conditions":["冻结过滤/缺失/异常/重复规则、分组与时间粒度、数据许可/预算和可用代码/SQL环境。","准备关键汇总/图的独立复算、合理替代与数据错误控制；保留探索路径和挑选过程。"],"evidenceBasis":"NIST EDA定义覆盖结构/异常/假设，专用数据Agent案例提供查询-结果反馈的邻近工作法；采用建议为任务适配，不声称此探索题已实测。","localStatus":"本次核查公开资料与任务契约，未在本站运行目标任务或新增模型试次。","gaps":["取一包含单位错误、重复及真实延迟长尾的数据，冻结问题；Agent交分布/关系/时间探索，独立复算并用替代解释挑战。"],"mode":"bounded-delivery","rationale":"NIST EDA定义覆盖结构/异常/假设，专用数据Agent案例提供查询-结果反馈的邻近工作法；采用建议为任务适配，不声称此探索题已实测。","frontier":"有公开探索统计方法和专用查询Agent案例；目标数据机制/非组成探索尚无本站试验，相关性不能迁移为因果。","nextValidation":"取一包含单位错误、重复及真实延迟长尾的数据，冻结问题；Agent交分布/关系/时间探索，独立复算并用替代解释挑战。","futureSignals":["新窗口/数据定义下主要发现可复算且备选解释经得起反证，可扩大探索。","单位/选择错误或事后挑图把关联写成因果时拒绝结论。"]},"workflow":{"before":"清洗、看图、猜机制和写报告分散，探索选择不易追踪。","now":"Agent可自主选择SQL/统计程序和图表，依据异常/反证返回数据检查；稳定计算不必循环Agent。","retained":["选择问题、可比口径和业务机制，确认需要何种下一步实验；不从故事流畅决定行动。"],"added":["列至少一个竞争解释与可否定检查，分开统计关联、机制假设和有证据因果。"]},"humanResponsibility":"选择问题、可比口径和业务机制，确认需要何种下一步实验；不从故事流畅决定行动。","failureHandling":"不可比/缺关键字段先交数据诊断与影响范围，不凭空补机制；预算内未找到解释也保留已排除假设。","example":{"input":"服务延迟中位数稳定但尾部上升；可能单位混用、重试重复、少数大请求或真实时间变化。","output":"数据核对、分布/尾部分位、大小-延迟关系和时间变化，竞争解释/反证及底层表。","failure":"删异常/平滑后声称一切正常，或将时间关联当发布因果。"},"sourceIds":["WU-A-EDA-METHOD","WU-DATA-AGENT"],"evidenceLinks":[{"sourceId":"WU-A-EDA-METHOD","supports":"定义非组成探索的可接收范围。","notCovered":"统计方法不是AI表现。"},{"sourceId":"WU-DATA-AGENT","supports":"邻近分析案例支持任务组织。","notCovered":"内部案例没有测试本分布/关系/时间探索，不能给成功率。"}],"resourceRefs":[],"career":{"direction":"数据科学与探索分析","evidence":"可重跑查询/程序、数据核对与分析口径","nextStep":"取一包含单位错误、重复及真实延迟长尾的数据，冻结问题；Agent交分布/关系/时间探索，独立复算并用替代解释挑战。"},"team":{"bottleneck":"NIST EDA定义覆盖结构/异常/假设，专用数据Agent案例提供查询-结果反馈的邻近工作法；采用建议为任务适配，不声称此探索题已实测。","allocation":"Agent可自主选择SQL/统计程序和图表，依据异常/反证返回数据检查；稳定计算不必循环Agent。","nextStep":"取一包含单位错误、重复及真实延迟长尾的数据，冻结问题；Agent交分布/关系/时间探索，独立复算并用替代解释挑战。"},"claims":[{"id":"A04-EDA","statement":"NIST EDA方法把分布、结构、异常与假设探索纳入分析，不应由单一组成率分解代表。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-A-EDA-METHOD","locator":"1.1.1 EDA purpose","relation":"method","supports":"定义非组成探索的可接收范围。","limit":"统计方法不是AI表现。"}],"configuration":null,"progressIds":[],"limitation":"统计方法不是AI表现。","evidenceStatus":"active"},{"id":"A04-ADOPT","statement":"可把专用数据Agent查询/结果反馈方法迁移到可重放探索，但目标数据定义、竞争假设和独立复算必须显式交付。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-DATA-AGENT","locator":"Context is everything；query/results","relation":"partial","supports":"邻近分析案例支持任务组织。","limit":"内部案例没有测试本分布/关系/时间探索，不能给成功率。"}],"configuration":null,"progressIds":[],"limitation":"内部案例没有测试本分布/关系/时间探索，不能给成功率。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开运行时可用，内部数据Agent不是可下载同配置产品；本站未对新探索题运行模型。","tools":[{"name":"SQL/Python探索环境","url":"https://www.python.org/","version":"实际Python/统计包/数据快照固定","access":"公开运行时；目标数据/查询与图表依赖另备"}]},"variants":[{"title":"分布异常与关系","inputDifference":"连续/类别变量，存在重尾、异常或混杂。","assessmentChange":"先核对数据生成/单位/过滤，再探索关系。","acceptanceChange":"分位/分布/散点与独立汇总、合理过滤敏感性、竞争解释。"},{"title":"时间变化与采样","inputDifference":"时间窗口、采样/季节/部署或记录方式变化。","assessmentChange":"按可比窗口和采样机制接收，不直接归因。","acceptanceChange":"时区/时间粒度、缺测/延迟、窗口稳健性和变点解释。"}],"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-F06","domainId":"finance","title":"交易凭证到结账包：分录、调整和报表勾稽","goal":"把冻结期间的交易资料变成可追踪、可重算且未决事项明确的候选结账包。","trigger":"期末结账或账务规则变更，需要从凭证追到账表。","scope":"限已给准则、科目和授权账套；合并抵销及特殊会计判断按专门变体处理。","inputs":[{"name":"原始凭证与合同","detail":"期间、主体、币种、单据完整性及来源"},{"name":"科目政策与账套","detail":"会计政策版本、科目映射、调整规则和只读初始账套"}],"deliverables":["凭证—分录—科目明细追踪","调整分录和试算平衡","报表及附注候选与结账差异","会计判断、控制缺陷和未取得证据"],"acceptance":[{"criterion":"交易事实完整","method":"从银行和业务流水反查凭证，含缺票、重复与跨期负例","failure":"平衡分录掩盖漏记收入"},{"criterion":"规则与账表一致","method":"重新执行分录、重算试算表和报表勾稽，复核收入确认等政策适用","failure":"借贷平衡但科目或期间错误"},{"criterion":"受控变更","method":"变更前后权限、日志、调整与抵销单独核对","failure":"静默用平衡项消掉差异"}],"assessment":{"recommendation":"可分阶段委派；把冻结期间的交易资料变成可追踪、可重算且未决事项明确的候选结账包。 接收依据见下方实质检查，不以流畅解释替代结果。","conditions":["限已给准则、科目和授权账套；合并抵销及特殊会计判断按专门变体处理。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"账务校验、官方月结模板和会计Agent产品是三种不同依据；据此给分阶段候选结账建议，不宣称真实过账已验证。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["账务规则工具，不证明原始交易真实、会计政策正确或AI入账成功率。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"staged-delivery","rationale":"工具校验、官方月结模板与已发布会计Agent共同支持分阶段交付候选结账；准则事实、复杂合并和真实过账须单独验收。","frontier":"公开月结模板产物包括计提与滚动表，且模板不直接过账；Intuit发布分类/对账产品。复杂准则、合并和原始凭证真实性仍没有同任务整包结果。","nextValidation":"一份单体月度账套加漏记、重复和跨期三类植入错误，核对调整、报表及错误是否逐一被揭示。","futureSignals":["独立账套跨多个期间的分录、截止、报表和例外均能对回原始证据，可扩大结账与授权过账","借贷平衡掩盖漏记、跨期或不当确认时，按事项退回而非只改平衡分录"]},"workflow":{"before":"凭证读取、科目选择、计提表、结账和报表多次手工交接。","now":"Agent围绕同一期间账套建立凭证追踪，按已给政策形成分录，调用账务校验并重算报表；发现跨期、缺证或政策冲突时回到具体事项修订。","retained":["交易事实完整","规则与账表一致","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现平衡分录掩盖漏记收入等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"期末服务合同含预收款和跨期履约资料。","output":"交付按已给政策形成的分录、调整和账表追踪；缺履约证据单列。","failure":"将全额收款作当期收入，账表仍平衡。"},"sourceIds":["WU-FIN-BEAN-VALIDATION","WU-FIN-CLOSE-TEMPLATE","WU-FIN-INTUIT-AGENT"],"evidenceLinks":[{"sourceId":"WU-FIN-BEAN-VALIDATION","supports":"确定性校验器可检查账户和交易平衡。","notCovered":"账务规则工具，不证明原始交易真实、会计政策正确或AI入账成功率。"},{"sourceId":"WU-FIN-CLOSE-TEMPLATE","supports":"计提、滚动表和差异说明组成月结候选包。","notCovered":"该模板不直接过账；是可配置参考而非可靠性试验，不能当永久人机边界。"},{"sourceId":"WU-FIN-INTUIT-AGENT","supports":"厂商已有会计Agent产品覆盖交易分类和对账，而非仅提示词建议。","notCovered":"产品自报非独立结果；没有本站账套、错误分母或人工介入完整日志，不使用客户节时百分比。"}],"resourceRefs":[],"career":{"direction":"交易凭证到结账包：分录、调整和报表勾稽","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"一份单体月度账套加漏记、重复和跨期三类植入错误，核对调整、报表及错误是否逐一被揭示。"},"team":{"bottleneck":"先区分资料、计算和现实状态/假设哪一项造成返工。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"一份单体月度账套加漏记、重复和跨期三类植入错误，核对调整、报表及错误是否逐一被揭示。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F06-method","statement":"确定性校验器可检查账户和交易平衡。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-BEAN-VALIDATION","locator":"validate_active_accounts、validate_currency_constraints、validate_check_transaction_balances","relation":"method","supports":"确定性校验器可检查账户和交易平衡。","limit":"账务规则工具，不证明原始交易真实、会计政策正确或AI入账成功率。"}],"progressIds":[],"configuration":null,"limitation":"账务规则工具，不证明原始交易真实、会计政策正确或AI入账成功率。","evidenceStatus":"active"},{"id":"W-F06-adoption","statement":"工具校验、官方月结模板与已发布会计Agent共同支持分阶段交付候选结账；准则事实、复杂合并和真实过账须单独验收。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-BEAN-VALIDATION","locator":"validate_active_accounts、validate_currency_constraints、validate_check_transaction_balances","relation":"adjacent","supports":"可检查确定性账务校验实现。","limit":"不把方法存在当作AI已完成本任务。"},{"sourceId":"WU-FIN-CLOSE-TEMPLATE","locator":"What you produce / Guardrails","relation":"partial","supports":"计提、滚动表和差异说明组成月结候选包。","limit":"该模板不直接过账；是可配置参考而非可靠性试验，不能当永久人机边界。"},{"sourceId":"WU-FIN-INTUIT-AGENT","locator":"Accounting Agent、Finance Agent与产品可用性段落","relation":"partial","supports":"厂商介绍分类、对账及财务洞察的产品交付。","limit":"产品自报非独立结果；没有本站账套、错误分母或人工介入完整日志，不使用客户节时百分比。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"},{"id":"W-F06-wu-fin-close-template","statement":"计提、滚动表和差异说明组成月结候选包。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-CLOSE-TEMPLATE","locator":"What you produce / Guardrails","relation":"method","supports":"计提、滚动表和差异说明组成月结候选包。","limit":"该模板不直接过账；是可配置参考而非可靠性试验，不能当永久人机边界。"}],"progressIds":[],"configuration":null,"limitation":"该模板不直接过账；是可配置参考而非可靠性试验，不能当永久人机边界。","evidenceStatus":"active"},{"id":"W-F06-intuit-agent","statement":"厂商已有会计Agent产品覆盖交易分类和对账，而非仅提示词建议。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-INTUIT-AGENT","locator":"Accounting Agent、Finance Agent与产品可用性段落","relation":"partial","supports":"厂商已有会计Agent产品覆盖交易分类和对账，而非仅提示词建议。","limit":"产品自报非独立结果；没有本站账套、错误分母或人工介入完整日志，不使用客户节时百分比。"}],"progressIds":["P-FIN-ACCOUNTING-PRODUCT"],"configuration":"Intuit2025-07-22 Enterprise Suite发布；自报产品，未核机构账套、错误分母和完整人工介入。","limitation":"产品自报非独立结果；没有本站账套、错误分母或人工介入完整日志，不使用客户节时百分比。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"已核对公开代码/文档入口，未安装或重放。Agent模型和数据接口需另接；账户、API和计算费用未测。","tools":[{"name":"Beancount固定代码：账户、币种与交易平衡校验","url":"https://github.com/beancount/beancount/blob/9747213802acb733ce53acc6e963f5961f47d6cd/beancount/ops/validation.py#L194-L410","version":"固定commit见URL","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"},{"name":"官方固定月结Agent模板","url":"https://github.com/anthropics/financial-services/blob/574ed3624aebd0418c7e96cd101262f30210ab26/plugins/agent-plugins/month-end-closer/agents/month-end-closer.md","version":"固定commit见URL","access":"公开参考；本任务未安装运行；需另接模型/内部数据，实际费用未知。"}]},"variants":[{"title":"单体月结","inputDifference":"科目政策已定","assessmentChange":"可分阶段委派候选结账","acceptanceChange":"凭证到报表并反查完整性"},{"title":"合并或特殊交易","inputDifference":"内部往来、控制权及复杂合同增加","assessmentChange":"会计口径先形成有依据的备选，再生成调整","acceptanceChange":"加验合并抵销、重复主体和确认时点"}],"evidenceStatus":"active"},{"id":"W-F07","domainId":"finance","title":"经营驱动到预算决策：预测、偏差和资源方案","goal":"交付可重算的驱动预算、情景比较和实际偏差解释，使资源取舍有可检查依据。","trigger":"年度预算、滚动预测或投资项目需要更新。","scope":"固定主体、期间、业务驱动和决策用途；预测场景不是已发生事实。","inputs":[{"name":"历史实际与业务驱动","detail":"交易/会计口径对照、销量价格人员及约束"},{"name":"预算模板与方案","detail":"联动公式、资本项目、折现口径和允许修改区域"}],"deliverables":["驱动预测与联动预算","实际预算差异桥及反证","资本投入或资源方案比较","预测误差、假设修订和管理报告"],"acceptance":[{"criterion":"模型联动保持","method":"用真实计算引擎重算现金流、利润和资产负债关键桥","failure":"变动收入却未更新应收"},{"criterion":"偏差解释可反驳","method":"重算量价结构、基期及残差，追踪业务原因的独立资料","failure":"构成变化误写提价"},{"criterion":"决策假设可比较","method":"在相同期间和约束重算至少两个方案，列不可识别参数","failure":"把乐观场景当确定收益"}],"assessment":{"recommendation":"可分阶段委派；交付可重算的驱动预算、情景比较和实际偏差解释，使资源取舍有可检查依据。 接收依据见下方实质检查，不以流畅解释替代结果。","conditions":["固定主体、期间、业务驱动和决策用途；预测场景不是已发生事实。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"完整文件评测为模型接收提供相邻实证，Finance Agent产品说明有经营预测入口；预算因果驱动和资源收益另验。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["2026-04-15版本与其中模型配置；非当前全部模型排行、非本站试验。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"staged-delivery","rationale":"相邻工作簿证据支持尝试建模交付；经营驱动、预测和实际资源收益仍需企业输入验证。","frontier":"已有跨表金融工作流与厂商情景预测产品，适合交联动预算和差异桥；尚未证明特定企业驱动可识别、预测稳定或资源方案最终有效。","nextValidation":"冻结一版预算，改变销量、账期和资本支出，再核对联动、约束、情景排序和预测误差。","futureSignals":["量价、账期和资本支出变化能正确传到整表，且预测偏差在新期间被解释，可扩滚动方案","只提高局部准确而漏现金或约束、或不断改写历史预测时，缩回模型候选并重校准"]},"workflow":{"before":"预算表、实际数据、资本项目和偏差说明分开更新，容易出现现金流未同步。","now":"Agent按决策目标联合更新驱动、工作簿、情景和说明，用真实重算反馈定位断链；可比较新驱动但须保留原预测，不能靠改历史假设抹平误差。","retained":["模型联动保持","偏差解释可反驳","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现变动收入却未更新应收等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"收入上调但客户账期延长。","output":"利润预测和营运资金同时更新，列现金缺口与融资情景。","failure":"只上调利润却未更新应收及现金。"},"sourceIds":["WU-FIN-FINCH-V5","WU-FIN-FINROBOT-DCF","WU-FIN-INTUIT-AGENT"],"evidenceLinks":[{"sourceId":"WU-FIN-FINCH-V5","supports":"Finch v5评估完整财务文件工作流。","notCovered":"2026-04-15版本与其中模型配置；非当前全部模型排行、非本站试验。"},{"sourceId":"WU-FIN-FINROBOT-DCF","supports":"仅可检验计算/结构部分的工具入口。","notCovered":"不证明整个任务完成。"},{"sourceId":"WU-FIN-INTUIT-AGENT","supports":"Intuit公开Finance Agent的报告、KPI、情景预测产品入口。","notCovered":"厂商发布，不证明经营驱动、预算准确率或采纳收益。"}],"resourceRefs":[],"career":{"direction":"经营驱动到预算决策：预测、偏差和资源方案","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"冻结一版预算，改变销量、账期和资本支出，再核对联动、约束、情景排序和预测误差。"},"team":{"bottleneck":"先区分资料、计算和现实状态/假设哪一项造成返工。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"冻结一版预算，改变销量、账期和资本支出，再核对联动、约束、情景排序和预测误差。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F07-method","statement":"Finch v5按真实财务文件工作流及整交付评分，提供预算预测类工作簿的相邻任务证据。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-FINCH-V5","locator":"§2.3.1—2.3.2、§3.1、§3.2、Appendix D/F","relation":"partial","supports":"Finch v5按真实财务文件工作流及整交付评分，提供预算预测类工作簿的相邻任务证据。","limit":"2026-04-15版本与其中模型配置；非当前全部模型排行、非本站试验。"}],"progressIds":[],"configuration":"2026-04-15 v5；172个文件工作流，表格与演示产物按原文工作流和整包分母；非本站运行。","limitation":"2026-04-15版本与其中模型配置；非当前全部模型排行、非本站试验。","evidenceStatus":"active"},{"id":"W-F07-adoption","statement":"相邻工作簿证据支持尝试建模交付；经营驱动、预测和实际资源收益仍需企业输入验证。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-FINCH-V5","locator":"§2.3.1—2.3.2、§3.1、§3.2、Appendix D/F","relation":"adjacent","supports":"172个复合工作流的作者报告与整包验收方法。","limit":"不把方法存在当作AI已完成本任务。"},{"sourceId":"WU-FIN-INTUIT-AGENT","locator":"Finance Agent / Payments Agent","relation":"partial","supports":"Intuit公开Finance Agent的报告、KPI、情景预测产品入口。","limit":"厂商发布，不证明经营驱动、预算准确率或采纳收益。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"},{"id":"W-F07-intuit-agent","statement":"Intuit公开Finance Agent的报告、KPI、情景预测产品入口。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-INTUIT-AGENT","locator":"Finance Agent / Payments Agent / Pricing and Availability","relation":"partial","supports":"Intuit公开Finance Agent的报告、KPI、情景预测产品入口。","limit":"厂商发布，不证明经营驱动、预算准确率或采纳收益。"}],"progressIds":["P-FIN-ACCOUNTING-PRODUCT"],"configuration":"2025-07-22美国Enterprise Suite产品发布；合同定价与数据接入；无公开本任务正确率和完整人工日志。","limitation":"厂商发布，不证明经营驱动、预算准确率或采纳收益。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"规范/评测是接收依据，另列代码工具为局部计算入口；Agent模型、授权数据和接入需自行准备，端到端成本未测。","tools":[{"name":"Finch v5：完整财务工作流与产物验收","url":"https://arxiv.org/html/2512.13168v5","version":"页面版本见来源","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"},{"name":"FinRobot固定DCF计算与边界处理","url":"https://github.com/AI4Finance-Foundation/FinRobot/blob/2717499b8e30f242640af08c4ad9afd1113c2d45/finrobot_desktop/finrobot/engine/compute/operators/dcf.py#L48-L177","version":"固定commit见URL","access":"可检查代码，需自行准备依赖；未本站运行。"}]},"variants":[{"title":"滚动预测","inputDifference":"期间和最新实际更新","assessmentChange":"优先委派模型更新与偏差桥","acceptanceChange":"新旧版本勾稽"},{"title":"资本项目","inputDifference":"投入时点和机会成本增加","assessmentChange":"候选方案可比较，最终收益待跟踪","acceptanceChange":"检验折现、残值、融资与不可逆约束"}],"evidenceStatus":"active"},{"id":"W-F08","domainId":"finance","title":"用途明确的估值模型：方法、参数和反证","goal":"交付基准日一致的估值模型、方法选择和可反驳的假设区间。","trigger":"企业、证券、资产或并购交易需要估值或复核。","scope":"计算方法可共享；企业/股权、资产及交易用途不能共享未经说明的价值口径。","inputs":[{"name":"估值定义与资料","detail":"对象、权利、基准日、用途、财务和可比资料"},{"name":"方法与约束","detail":"允许模型、资本结构、折现/增长及可比筛选依据"}],"deliverables":["正常化财务与企业股权桥","DCF或可比模型及敏感性","方法取舍和关键假设证据","估值报告与无法可靠估计的范围"],"acceptance":[{"criterion":"估值对象一致","method":"核对企业/股权价值、债务现金、股份数与日期","failure":"企业价值直接除以股数"},{"criterion":"计算独立复算","method":"用第二实现重算现金流和终值，检查增长与折现边界","failure":"分母接近零仍输出精确合理价格"},{"criterion":"可比与假设受挑战","method":"检查同行差异和相反证据，报告参数区间","failure":"仅因结果符合报价倒推假设"}],"assessment":{"recommendation":"可分阶段委派；交付基准日一致的估值模型、方法选择和可反驳的假设区间。 接收依据见下方实质检查，不以流畅解释替代结果。","conditions":["计算方法可共享；企业/股权、资产及交易用途不能共享未经说明的价值口径。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"固定DCF代码和可下载研究报告证明局部计算与产物可检查；正式估值采用为编辑迁移，不把代码当价值判断。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"staged-delivery","rationale":"资料和计算能整合为候选估值包；可查代码只支持可执行计算，方法适合度和假设仍是明确待验内容。","frontier":"FinRobot把计算实现与研究稿相连，但本次仅抽查计算边界和报告存在；特殊资产、可比选择与估值目的仍无整包AI验证。","nextValidation":"独立复算一个现金流模型，加入净债务符号错、折现率不适用和不可比样本，检查是否识别并改写结论。","futureSignals":["独立模型在不同价值口径和新资料中能保持桥接、方法适用和反证，可扩大估值任务","终值失真、权利对象混淆或可比选择被单一报价牵引时，停止给单一确定估值"]},"workflow":{"before":"收集可比、正常化财务、估值和撰写意见逐次交接。","now":"Agent可并行或往返比较估值方法、筛选可比与假设，用计算器重算现金流和边界；方法不适用或终值失真时应换路线或保留不可估计，而非只润色价格解释。","retained":["估值对象一致","计算独立复算","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现企业价值直接除以股数等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"给出三年预测、净债务与WACC区间。","output":"交付企业与股权价值桥、终值占比和参数敏感性。","failure":"把不可持续增长与低折现拼出单一高价。"},"sourceIds":["WU-FIN-FINROBOT-DCF","WU-FIN-FINROBOT-REPORT"],"evidenceLinks":[{"sourceId":"WU-FIN-FINROBOT-DCF","supports":"固定DCF代码可检查现金流折现及边界。","notCovered":"仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。"},{"sourceId":"WU-FIN-FINROBOT-REPORT","supports":"公开可取的报告产物入口。","notCovered":"只核对产物存在及来源身份，没有复算全报告，不作为质量或收益独立验证。"}],"resourceRefs":[],"career":{"direction":"用途明确的估值模型：方法、参数和反证","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"独立复算一个现金流模型，加入净债务符号错、折现率不适用和不可比样本，检查是否识别并改写结论。"},"team":{"bottleneck":"先区分资料、计算和现实状态/假设哪一项造成返工。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"独立复算一个现金流模型，加入净债务符号错、折现率不适用和不可比样本，检查是否识别并改写结论。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F08-method","statement":"固定DCF代码可检查现金流折现及边界。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-FINROBOT-DCF","locator":"calculate_dcf；terminal growth/WACC与terminal FCF边界","relation":"method","supports":"固定DCF代码可检查现金流折现及边界。","limit":"仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。"}],"progressIds":[],"configuration":null,"limitation":"仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。","evidenceStatus":"active"},{"id":"W-F08-adoption","statement":"资料和计算能整合为候选估值包；可查代码只支持可执行计算，方法适合度和假设仍是明确待验内容。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-FINROBOT-DCF","locator":"calculate_dcf；terminal growth/WACC与terminal FCF边界","relation":"adjacent","supports":"确定性DCF实现可审阅，可与Agent的资料和假设生成分开。","limit":"不把方法存在当作AI已完成本任务。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"},{"id":"W-F08-finrobot-report","statement":"公开可取的报告产物入口。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-FIN-FINROBOT-REPORT","locator":"report/NVDA_report.pdf；同commit README的V0安装入口","relation":"partial","supports":"公开可取的报告产物入口。","limit":"只核对产物存在及来源身份，没有复算全报告，不作为质量或收益独立验证。"}],"progressIds":[],"configuration":null,"limitation":"只核对产物存在及来源身份，没有复算全报告，不作为质量或收益独立验证。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"已核对公开代码/文档入口，未安装或重放。Agent模型和数据接口需另接；账户、API和计算费用未测。","tools":[{"name":"FinRobot固定DCF计算与边界处理","url":"https://github.com/AI4Finance-Foundation/FinRobot/blob/2717499b8e30f242640af08c4ad9afd1113c2d45/finrobot_desktop/finrobot/engine/compute/operators/dcf.py#L48-L177","version":"固定commit见URL","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"}]},"variants":[{"title":"企业现金流估值","inputDifference":"预测及净债务桥","assessmentChange":"分阶段委派模型和假设分析","acceptanceChange":"复算终值、桥及情景"},{"title":"无形资产或特殊用途","inputDifference":"权利与价值口径不同","assessmentChange":"只迁移计算和追溯方法，需专门证据","acceptanceChange":"独立核对资产方法和权利范围"}],"evidenceStatus":"active"},{"id":"W-F09","domainId":"finance","title":"企业资金包：现金缺口、融资、支付与套保候选","goal":"交付账户时点一致的现金预测与资金安排，并能追踪授权内指令及实际状态。","trigger":"支付、融资到期或汇率变化需要资金安排。","scope":"分析、指令准备与执行状态分开；真实支付须有账户接口、权限和可核对回单，不因责任归属排除自动执行研究。","inputs":[{"name":"账户和现金事项","detail":"可用/受限余额、价值日、应收应付和债务到期"},{"name":"资金规则与权限","detail":"币种、账户、融资契约、限额和支付接口权限"}],"deliverables":["现金可用性与缺口预测","融资方案、契约和成本比较","汇率利率敞口及套保情景","支付调拨指令候选、回单状态及异常","权限与对账控制问题"],"acceptance":[{"criterion":"时间与资金守恒","method":"按账户币种和价值日独立复算余额，排除受限资金","failure":"到账前现金重复可用"},{"criterion":"风险和成本一致","method":"复核债务契约、费用、到期和套保敞口","failure":"套保方向增加净风险"},{"criterion":"指令实际状态可确认","method":"在沙盒验证限额、幂等、回单和失败恢复，不以创建指令当到账","failure":"重复付款或仅发出未确认"}],"assessment":{"recommendation":"可分阶段委派；交付账户时点一致的现金预测与资金安排，并能追踪授权内指令及实际状态。 接收依据见下方实质检查，不以流畅解释替代结果。","conditions":["分析、指令准备与执行状态分开；真实支付须有账户接口、权限和可核对回单，不因责任归属排除自动执行研究。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"现金预测可重算；收款Agent只支持应收预测和沟通的相邻路径，ORE仅支持风险计算方法，均不证明付款完成。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"staged-delivery","rationale":"现金和指令数据可结构化，适合候选分析及沙盒状态闭环；风险工具对企业资金仅是局部方法迁移，银行接入尚无本站证据。","frontier":"已有收款提醒/预测产品，资金方案仍须对账户价值日和真实可用现金验收；真实调拨、融资成交和套保闭环需沙盒与银行状态补证。","nextValidation":"冻结账户日历、加入受限款、跨时区入账和重复支付，在沙盒检查缺口及状态闭环。","futureSignals":["价值日、限额和多币种预测能对上实际到账，重复/失败指令可恢复后扩付款自动化","出现未到账先记可用、受限款重复使用或银行状态未知时，阻断相应资金动作"]},"workflow":{"before":"资金人员导出账户、预测、准备指令，再分别追到账和风险。","now":"Agent用同一价值日现金状态组织预测、融资/套保备选与指令，运行额度和幂等校验，按实际回报修订计划；接口断连先读回状态，不盲目重发。","retained":["时间与资金守恒","风险和成本一致","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现到账前现金重复可用等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"账上现金100，受限30，明日支付80。","output":"明确可用70及至少10缺口，候选融资与支付状态。","failure":"按账面100判定充足并提前结案。"},"sourceIds":["WU-FIN-ORE-STRESS","WU-FIN-INTUIT-AGENT"],"evidenceLinks":[{"sourceId":"WU-FIN-ORE-STRESS","supports":"风险情景工具有可检查配置。","notCovered":"示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。"},{"sourceId":"WU-FIN-INTUIT-AGENT","supports":"Intuit的Payments Agent声明涉及收款预测、发票及提醒；不等于执行企业资金付款。","notCovered":"收款沟通与银行调拨/融资套保不同；不迁移节时或到账数字。"}],"resourceRefs":[],"career":{"direction":"企业资金包：现金缺口、融资、支付与套保候选","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"冻结账户日历、加入受限款、跨时区入账和重复支付，在沙盒检查缺口及状态闭环。"},"team":{"bottleneck":"先区分资料、计算和现实状态/假设哪一项造成返工。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"冻结账户日历、加入受限款、跨时区入账和重复支付，在沙盒检查缺口及状态闭环。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F09-method","statement":"风险情景工具有可检查配置。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-ORE-STRESS","locator":"XVA Stress Testing / XVA P&L Explain；run_stress.py与配置文件","relation":"method","supports":"风险情景工具有可检查配置。","limit":"示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。"}],"progressIds":[],"configuration":null,"limitation":"示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。","evidenceStatus":"active"},{"id":"W-F09-adoption","statement":"现金和指令数据可结构化，适合候选分析及沙盒状态闭环；风险工具对企业资金仅是局部方法迁移，银行接入尚无本站证据。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-ORE-STRESS","locator":"XVA Stress Testing / XVA P&L Explain；run_stress.py与配置文件","relation":"adjacent","supports":"可检查的定价风险情景输入和运行入口。","limit":"不把方法存在当作AI已完成本任务。"},{"sourceId":"WU-FIN-INTUIT-AGENT","locator":"Finance Agent / Payments Agent","relation":"adjacent","supports":"Intuit的Payments Agent声明涉及收款预测、发票及提醒；不等于执行企业资金付款。","limit":"收款沟通与银行调拨/融资套保不同；不迁移节时或到账数字。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"},{"id":"W-F09-intuit-agent","statement":"Intuit的Payments Agent声明涉及收款预测、发票及提醒；不等于执行企业资金付款。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-INTUIT-AGENT","locator":"Finance Agent / Payments Agent / Pricing and Availability","relation":"partial","supports":"Intuit的Payments Agent声明涉及收款预测、发票及提醒；不等于执行企业资金付款。","limit":"收款沟通与银行调拨/融资套保不同；不迁移节时或到账数字。"}],"progressIds":["P-FIN-ACCOUNTING-PRODUCT"],"configuration":"2025-07-22美国Enterprise Suite产品发布；合同定价与数据接入；无公开本任务正确率和完整人工日志。","limitation":"收款沟通与银行调拨/融资套保不同；不迁移节时或到账数字。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"已核对公开代码/文档入口，未安装或重放。Agent模型和数据接口需另接；账户、API和计算费用未测。","tools":[{"name":"ORE固定XVA压力与损益归因算例","url":"https://github.com/OpenSourceRisk/Engine/blob/3d75a69087911a7e2cf1e6882da7bc27135c0762/Examples/XvaRisk/Readme.md","version":"固定commit见URL","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"}]},"variants":[{"title":"现金与融资","inputDifference":"价值日、契约和额度","assessmentChange":"可委派候选现金方案","acceptanceChange":"独立核对可用性"},{"title":"支付或套保","inputDifference":"真实指令与市场风险增加","assessmentChange":"先验证接口幂等及授权再扩围","acceptanceChange":"核对银行成交回单、暴露和恢复"}],"evidenceStatus":"active"},{"id":"W-F10","domainId":"finance","title":"信用评估包：还款来源、条款与贷后预警","goal":"交付原始事实可追踪、压力可复算并含反证的授信或续作分析。","trigger":"新授信、额度续作或借款人风险变化。","scope":"覆盖资料、偿债、结构建议与贷后分析；评级和批准权限单列，不把制度签字当技术永久界限。","inputs":[{"name":"借款人证据包","detail":"财报、流水、债务、担保及关联方"},{"name":"信用政策与用途","detail":"评级/限额口径、期限、行业和例外规则"}],"deliverables":["事实核验、缺件与矛盾清单","偿债及压力模型","额度期限担保与条款建议","反证、例外和信用备忘录","契约监测和风险变化解释"],"acceptance":[{"criterion":"还款不只盈利","method":"把利润、现金流、到期与续借依赖对照","failure":"利润增长掩盖到期缺口"},{"criterion":"担保与压力真实","method":"估算处置折扣并明确法律可实现性未核状态，重算现金压力","failure":"以未核担保补足还款"},{"criterion":"政策与预警可追踪","method":"政策版本、例外、观察期和变动证据逐项列示","failure":"旧评级自动批准新额度"}],"assessment":{"recommendation":"可用于产生和挑战候选；交付原始事实可追踪、压力可复算并含反证的授信或续作分析。 接收依据见下方实质检查，不以流畅解释替代结果。","conditions":["覆盖资料、偿债、结构建议与贷后分析；评级和批准权限单列，不把制度签字当技术永久界限。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"FinGAIA公开例只为财报和实体资料检索提供局部参照；实际任务见下方具体主张。授信决定的整包采用另依公开方法、明确条件和独立验收提出，不从角色名推出专业能力。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["只核对产物存在及来源身份，没有复算全报告，不作为质量或收益独立验证。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"candidate-exploration","rationale":"检索和交叉核对可组成信用候选包的资料层；FinGAIA的角色名不能证明授信能力。还款来源、条款、担保和反证须用专属案件独立接收。","frontier":"已有跨网页/附件的财报事实检索评估；该公开信用角色样例没有完成借款人信用判断。信用报告整包采用仍是基于可追踪材料、现金流和明确准则的条件推断。","nextValidation":"同一借款包加入未披露债务、关联回款和担保缺页，核对建议是否改判并留下缺件。","futureSignals":["在独立借款包识别隐藏债务、关联回款和期限压力，且条件建议随反证改判，可扩大分析","对缺债或无效担保继续无条件接受，或把盈利当现金还款时，拒收相关信用建议"]},"workflow":{"before":"借款资料收集、信用模型、条款和报告互相分离。","now":"Agent围绕可验证还款来源搜索反证、重算压力并提出条件，发现关联回款或新债务时回修整个信用结论；贷后新信息复用原判断而不自动沿用评级。","retained":["还款不只盈利","担保与压力真实","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现利润增长掩盖到期缺口等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"企业利润增长但大额债务即将到期。","output":"列真实现金还款与再融资两种路径、限制和拒判条件。","failure":"将利润率提升写成必然可偿债。"},"sourceIds":["WU-FIN-FINROBOT-REPORT","WU-FIN-FINROBOT-DCF","WU-FIN-FINGAIA"],"evidenceLinks":[{"sourceId":"WU-FIN-FINROBOT-REPORT","supports":"已有金融研究报告产物可取得。","notCovered":"只核对产物存在及来源身份，没有复算全报告，不作为质量或收益独立验证。"},{"sourceId":"WU-FIN-FINROBOT-DCF","supports":"确定性DCF实现可审阅，可与Agent的资料和假设生成分开。","notCovered":"仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。"},{"sourceId":"WU-FIN-FINGAIA","supports":"Appendix A Fig6在“Loan Credit Analysis”名下的公开例实际核对年报股东权益、季度利润、股东及审计师姓名。只支持财报和实体资料检索子能力。","notCovered":"借款人还款来源、违约或完整授信决定及本业务端到端可靠性。"}],"resourceRefs":[],"career":{"direction":"信用评估包：还款来源、条款与贷后预警","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"同一借款包加入未披露债务、关联回款和担保缺页，核对建议是否改判并留下缺件。"},"team":{"bottleneck":"先区分资料、计算和现实状态/假设哪一项造成返工。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"同一借款包加入未披露债务、关联回款和担保缺页，核对建议是否改判并留下缺件。"},"reviewedAt":"2026-10-21","claims":[{"id":"W-F10-method","statement":"已有金融研究报告产物可取得。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-FINROBOT-REPORT","locator":"report/NVDA_report.pdf；同commit README的V0安装入口","relation":"method","supports":"已有金融研究报告产物可取得。","limit":"只核对产物存在及来源身份，没有复算全报告，不作为质量或收益独立验证。"}],"progressIds":[],"configuration":null,"limitation":"只核对产物存在及来源身份，没有复算全报告，不作为质量或收益独立验证。","evidenceStatus":"active"},{"id":"W-F10-adoption","statement":"借款资料、政策和计算规则已冻结时，可委派有出处的信用候选包；还款来源、条款及反证按专属案件独立接收。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-FINROBOT-REPORT","locator":"report/NVDA_report.pdf；同commit README的V0安装入口","relation":"adjacent","supports":"公开可取的报告产物入口。","limit":"不把方法存在当作AI已完成本任务。"},{"sourceId":"WU-FIN-FINGAIA","locator":"Table2、Agents/Evaluation Methods、Table3及AppendixB","relation":"adjacent","supports":"财报和实体资料检索可作为交付中取材/计算环节的参照。","limit":"完整采用建议为编辑推断；Appendix A Fig6在“Loan Credit Analysis”名下的公开例实际核对年报股东权益、季度利润、股东及审计师姓名。它不验证借款人还款来源、违约或完整授信决定。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"},{"id":"W-F10-finrobot-dcf","statement":"确定性DCF实现可审阅，可与Agent的资料和假设生成分开。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-FIN-FINROBOT-DCF","locator":"calculate_dcf；terminal growth/WACC与terminal FCF边界","relation":"partial","supports":"确定性DCF实现可审阅，可与Agent的资料和假设生成分开。","limit":"仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。"}],"progressIds":[],"configuration":null,"limitation":"仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。","evidenceStatus":"active"},{"id":"W-F10-fingaia","statement":"Appendix A Fig6在“Loan Credit Analysis”名下的公开例实际核对年报股东权益、季度利润、股东及审计师姓名。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-FINGAIA","locator":"Appendix A Fig6在“Loan Credit Analysis”名下的","relation":"adjacent","supports":"财报和实体资料检索的局部评估参照。","limit":"角色标签不等于实际专业任务；公开例不能证明借款人还款来源、违约或完整授信决定，不把分类分数当本任务成功率。"}],"progressIds":["P-FIN-SCENARIOS"],"configuration":"FinGAIA v1零样本网页/附件原环境；论文角色分类需对照Appendix A实际样例；正文与附表口径有差异，不使用该标签分数为本专业成功率。","limitation":"角色标签不等于实际专业任务；公开例不能证明借款人还款来源、违约或完整授信决定，不把分类分数当本任务成功率。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"可读固定版本代码说明及报告产物；报告PDF不是运行工具，本专业整包任务尚未安装或重放，数据接口、API及人时成本未知。","tools":[{"name":"FinRobot固定项目与安装说明","url":"https://github.com/AI4Finance-Foundation/FinRobot/tree/2717499b8e30f242640af08c4ad9afd1113c2d45","version":"commit 2717499b8e30f242640af08c4ad9afd1113c2d45","access":"开放项目说明；模型、数据、依赖需自行准备，未验证本专业任务可一键执行。"},{"name":"FinRobot已有NVDA示例研究报告","url":"https://github.com/AI4Finance-Foundation/FinRobot/blob/2717499b8e30f242640af08c4ad9afd1113c2d45/finrobot_autogen/report/NVDA_report.pdf","version":"固定commit见URL","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"}]},"variants":[{"title":"标准企业授信","inputDifference":"财报和现金流较齐","assessmentChange":"可尝试整包候选评估","acceptanceChange":"独立核对到期和反证"},{"title":"非标/跨境","inputDifference":"复杂结构及担保生效不同","assessmentChange":"分段完成已证事实，保留未决条件","acceptanceChange":"加验关联方和法律实现材料"}],"evidenceStatus":"active"},{"id":"W-F11","domainId":"finance","title":"风险到审计证据：程序、底稿与发现闭环","goal":"交付主张、程序、原始证据和结论可追踪的审计底稿包。","trigger":"内部审计或外部审计的一项明确领域需实施程序。","scope":"AS1105仅用于美国PCAOB语境的证据方法示例；实际项目先固定标准及期间。AI取证、程序执行与报告签署分开记录。","inputs":[{"name":"审计委托与主张","detail":"范围、重要性、风险、适用标准期间"},{"name":"总体和原始证据","detail":"可验证完整性的数据、样本、外部证据和控制记录"}],"deliverables":["风险与程序映射","总体、样本与执行底稿","控制设计/运行及实质性发现","支持反驳证据及未决项","审计结论草稿、整改和再测试记录"],"acceptance":[{"criterion":"程序回应主张","method":"从风险追到程序与样本，检查完整性和可靠性","failure":"只有公司提供摘要而无原件"},{"criterion":"推断不越样本","method":"核对总体边界、抽样方法、例外及反证","failure":"若干正常样本即断言全部无错"},{"criterion":"发现到整改闭环","method":"复验原始问题与后续期间控制，不以回复代修复","failure":"管理层承诺整改即标完成"}],"assessment":{"recommendation":"可分阶段委派；交付主张、程序、原始证据和结论可追踪的审计底稿包。 接收依据见下方实质检查，不以流畅解释替代结果。","conditions":["AS1105仅用于美国PCAOB语境的证据方法示例；实际项目先固定标准及期间。AI取证、程序执行与报告签署分开记录。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"AS1105只给美国PCAOB语境的证据方法；2024有限访谈有草稿、研究和部分对账使用，不能当2026能力上限。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["职业方法来源，非AI能力证明；具体审计期间与适用修订应按委托另核。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"staged-delivery","rationale":"专业标准支持底稿接收方法；历史使用观察和工具不证明完整审计可委派，适合先交付可逐项检查的程序产物。","frontier":"可见使用观察与确定性核对方法，但未取得整项AI审计的充分适当证据、独立取证与整改复验成绩；当前产物按审计主张逐项接收。","nextValidation":"选一个收入截止或付款控制领域，预置伪证、抽样偏差和冲突回函，验证底稿保留反证及未决主张。","futureSignals":["在新总体、冲突回函和控制失效样本上，程序能改变结论并保留取证链，可扩大程序包","若摘要冒充原件、正常样本被外推总体或整改回复代复验，缩回相应结论"]},"workflow":{"before":"风险表、取证、计算底稿、发现和整改分别产生，难追结论基础。","now":"Agent按主张自主选择可访问程序、数据核查和反证搜索，将真实执行证据和未取得证据分开；矛盾会改变程序或结论，整改后再测试，不把写好底稿视作取证。","retained":["程序回应主张","推断不越样本","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现只有公司提供摘要而无原件等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"收入样本发票日期正确但交付单跨期。","output":"底稿保留冲突并调整截止测试结论。","failure":"只核发票即认为收入完整正确。"},"sourceIds":["WU-FIN-AS1105","WU-FIN-PCAOB-GENAI-2024","WU-FIN-BEAN-VALIDATION"],"evidenceLinks":[{"sourceId":"WU-FIN-AS1105","supports":"AS1105区分证据充分性、适当性与冲突。","notCovered":"职业方法来源，非AI能力证明；具体审计期间与适用修订应按委托另核。"},{"sourceId":"WU-FIN-PCAOB-GENAI-2024","supports":"有限访谈中的资料研究、草稿和部分对账使用观察。","notCovered":"历史有限访谈，非2026能力上限、非正式规则、非审计全流程成绩。"},{"sourceId":"WU-FIN-BEAN-VALIDATION","supports":"仅可检验计算/结构部分的工具入口。","notCovered":"不证明整个任务完成。"}],"resourceRefs":[],"career":{"direction":"风险到审计证据：程序、底稿与发现闭环","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"选一个收入截止或付款控制领域，预置伪证、抽样偏差和冲突回函，验证底稿保留反证及未决主张。"},"team":{"bottleneck":"先区分资料、计算和现实状态/假设哪一项造成返工。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"选一个收入截止或付款控制领域，预置伪证、抽样偏差和冲突回函，验证底稿保留反证及未决主张。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F11-method","statement":"AS1105区分证据充分性、适当性与冲突。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-AS1105","locator":"§.02—.10、.11、.29；美国PCAOB适用审计语境","relation":"method","supports":"AS1105区分证据充分性、适当性与冲突。","limit":"职业方法来源，非AI能力证明；具体审计期间与适用修订应按委托另核。"}],"progressIds":[],"configuration":null,"limitation":"职业方法来源，非AI能力证明；具体审计期间与适用修订应按委托另核。","evidenceStatus":"active"},{"id":"W-F11-adoption","statement":"专业标准支持底稿接收方法；历史使用观察和工具不证明完整审计可委派，适合先交付可逐项检查的程序产物。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-AS1105","locator":"§.02—.10、.11、.29；美国PCAOB适用审计语境","relation":"adjacent","supports":"证据相关可靠性、主张与冲突证据的专业方法。","limit":"不把方法存在当作AI已完成本任务。"},{"sourceId":"WU-FIN-PCAOB-GENAI-2024","locator":"printed pp.4—5：Current Use / Risks；封面非规则声明","relation":"partial","supports":"有限访谈中的资料研究、草稿和部分对账使用观察。","limit":"历史有限访谈，非2026能力上限、非正式规则、非审计全流程成绩。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"},{"id":"W-F11-pcaob-genai-2024","statement":"有限访谈中的资料研究、草稿和部分对账使用观察。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-PCAOB-GENAI-2024","locator":"printed pp.4—5：Current Use / Risks；封面非规则声明","relation":"partial","supports":"有限访谈中的资料研究、草稿和部分对账使用观察。","limit":"历史有限访谈，非2026能力上限、非正式规则、非审计全流程成绩。"}],"progressIds":[],"configuration":"PCAOB 2024年7月对若干审计机构的有限访谈；资料研究、草稿及部分对账的使用观察，无任务准确率分母。","limitation":"历史有限访谈，非2026能力上限、非正式规则、非审计全流程成绩。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"规范/评测是接收依据，另列代码工具为局部计算入口；Agent模型、授权数据和接入需自行准备，端到端成本未测。","tools":[{"name":"PCAOB AS1105职业证据约束","url":"https://pcaobus.org/oversight/standards/auditing-standards/details/AS1105","version":"页面版本见来源","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"},{"name":"Beancount固定代码：账户、币种与交易平衡校验","url":"https://github.com/beancount/beancount/blob/9747213802acb733ce53acc6e963f5961f47d6cd/beancount/ops/validation.py#L194-L410","version":"固定commit见URL","access":"可检查代码，需自行准备依赖；未本站运行。"}]},"variants":[{"title":"内部控制审计","inputDifference":"控制目标和运行样本","assessmentChange":"可分阶段整理与执行可访问程序","acceptanceChange":"分别评估设计与运行"},{"title":"外部财报审计","inputDifference":"独立证据和法定意见用途","assessmentChange":"程序可委派范围按取证权限验证","acceptanceChange":"加验总体、独立性和重要性，不自动签意见"}],"evidenceStatus":"active"},{"id":"W-F12","domainId":"finance","title":"固定法域税年的税务计算与申报候选","goal":"交付事实到税表的规则定位、计算和待决资料，不把模拟器输出冒充申报完成。","trigger":"需要处理固定税年申报、修订或税务问询。","scope":"示例规则仅限美国联邦个人所得税2024年；其他地区、企业税、筹划及争议先取得相应规则，不从示例外推。 TY25研究产物与本页TY24规则示例分属不同年度，不能拼接计算或声称同例已验证。","inputs":[{"name":"主体与交易事实","detail":"税籍、税年、收入扣除凭据及原申报"},{"name":"固定规则及工具","detail":"官方说明版本、表单、参数、允许的模拟或电子申报接口"}],"deliverables":["资料和主体期间核对","条款表单逐项映射","可复算税务候选与账税差异","缺件、歧义、问询回复和修订","期限及规则变化影响清单"],"acceptance":[{"criterion":"法域期间锁定","method":"所有计算标注税年、主体、适用条款及版本","failure":"美国2024参数用于中国2026"},{"criterion":"事实到表单追踪","method":"核对收入/扣除证据、资格条件和各行勾稽","failure":"未知资格自动默认允许扣除"},{"criterion":"模拟和申报区分","method":"独立对照官方样例与表单，核查提交回执及错误","failure":"计算报告冒充已电子申报"}],"assessment":{"recommendation":"可分阶段委派；交付事实到税表的规则定位、计算和待决资料，不把模拟器输出冒充申报完成。 接收依据见下方实质检查，不以流畅解释替代结果。","conditions":["示例规则仅限美国联邦个人所得税2024年；其他地区、企业税、筹划及争议先取得相应规则，不从示例外推。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"TaxCalcBench保存产物是直接AI证据；Tax-Calculator是模型工具；IRS2024只支持本页TY24固定规则例，三者税年不可拼接。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["不是电子报税系统；公开说明无独立Agent成功率，不能覆盖中国税种或企业税务。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"staged-delivery","rationale":"已有复杂税表保存产物与模型使用Agent，可试固定法域年度的整包候选；不能只靠语言算税，须表单、确定性计算和提交状态分别验。","frontier":"固定TY25库已有含州税的PDF任务输出，展示比TY24结构化计算更宽的输入；本页TY24规则例未被这些产物跑过，TY25结果仅支持候选工作法的相邻迁移。","nextValidation":"仅用2024美国个人税固定样例，比较官方表单与模型输出；加入错误税年、缺件及资格矛盾，必须拒绝自动补假设。","futureSignals":["在新税年、资格边界和跨表规则独立样例逐行及整表一致，并有真实提交回执后，才扩授权申报","模型把不同税年参数、未知资格或模拟输出混作已申报时，拒收并回到固定规则核对"]},"workflow":{"before":"凭证整理、规则阅读、录表、计算和问询回复多次重复录入。","now":"Agent先锁法域税年并把事实映射规则与表单，计算交确定性引擎或独立复算，沿跨表错误回修；缺资格条件返回未决，已有正确表格也要单查提交状态。","retained":["法域期间锁定","事实到表单追踪","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现美国2024参数用于中国2026等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"教学性TY24美国个税资料包误混入TY25参数；这是本页编制的失败例，没有运行模型。","output":"隔离错误税年，形成TY24规则映射和待补凭据；另列TY25公开研究产物作为相邻路径，不充当本例通过证据。","failure":"静默套最新参数，结果可运行但年份不符。"},"sourceIds":["WU-FIN-TAXCALC","WU-FIN-IRS1040-2024","WU-FIN-TAXCALCBENCH","WU-FIN-TAX-SAVED-CASE"],"evidenceLinks":[{"sourceId":"WU-FIN-TAXCALC","supports":"固定Tax-Calculator仓库提供模型使用Agent入口。","notCovered":"不是电子报税系统；公开说明无独立Agent成功率，不能覆盖中国税种或企业税务。"},{"sourceId":"WU-FIN-IRS1040-2024","supports":"只作美国联邦个人所得税2024年示范的规则和表单核对入口。","notCovered":"不提供本网站用户当前税率/申报结论，不外推州税、企业税或中国规则。"},{"sourceId":"WU-FIN-TAXCALCBENCH","supports":"固定TY25库保留复杂PDF和州/联邦计算的模型产物，且区分整份严格正确与逐行分。","notCovered":"通常每配置单次保存结果，按最佳thinking设置展示，部分模型覆盖不足；生成成本不等于端到端报税成本。"},{"sourceId":"WU-FIN-TAX-SAVED-CASE","supports":"可以直接打开一份保存成功的逐行税表评估，检查正确字段与实际记录成本。","notCovered":"单个作者自评产物、未本站重放；选中成功例不证明新案例稳定性或提交成功。"}],"resourceRefs":[],"career":{"direction":"固定法域税年的税务计算与申报候选","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"仅用2024美国个人税固定样例，比较官方表单与模型输出；加入错误税年、缺件及资格矛盾，必须拒绝自动补假设。"},"team":{"bottleneck":"先区分资料、计算和现实状态/假设哪一项造成返工。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"仅用2024美国个人税固定样例，比较官方表单与模型输出；加入错误税年、缺件及资格矛盾，必须拒绝自动补假设。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F12-method","statement":"固定Tax-Calculator仓库提供模型使用Agent入口。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-TAXCALC","locator":"Model-Development Agent / Model-Use Agent；同commit README定义美国联邦个税与工资税","relation":"method","supports":"固定Tax-Calculator仓库提供模型使用Agent入口。","limit":"不是电子报税系统；公开说明无独立Agent成功率，不能覆盖中国税种或企业税务。"}],"progressIds":[],"configuration":null,"limitation":"不是电子报税系统；公开说明无独立Agent成功率，不能覆盖中国税种或企业税务。","evidenceStatus":"active"},{"id":"W-F12-adoption","statement":"已有复杂税表保存产物与模型使用Agent，可试固定法域年度的整包候选；不能只靠语言算税，须表单、确定性计算和提交状态分别验。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-TAXCALC","locator":"Model-Development Agent / Model-Use Agent；同commit README定义美国联邦个税与工资税","relation":"adjacent","supports":"固定版本的税务微观模拟与自然语言模型使用路径。","limit":"不把方法存在当作AI已完成本任务。"},{"sourceId":"WU-FIN-TAXCALCBENCH","locator":"TY2025表格、结果口径与Known issues；固定2026-10-19 commit","relation":"partial","supports":"PDF输入、州及联邦计算和保存输出可审阅。","limit":"通常每配置单次保存结果，按最佳thinking设置展示，部分模型覆盖不足；生成成本不等于端到端报税成本。"},{"sourceId":"WU-FIN-TAX-SAVED-CASE","locator":"逐行expected/actual及Strictly correct return、API Usage and Cost","relation":"partial","supports":"一个保存的模型评估文件，可检查逐行比较及成本口径。","limit":"单个作者自评产物、未本站重放；选中成功例不证明新案例稳定性或提交成功。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"},{"id":"W-F12-irs1040-2024","statement":"只作美国联邦个人所得税2024年示范的规则和表单核对入口。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-IRS1040-2024","locator":"封面税年、printed pp.31—37与64起税表","relation":"partial","supports":"只作美国联邦个人所得税2024年示范的规则和表单核对入口。","limit":"不提供本网站用户当前税率/申报结论，不外推州税、企业税或中国规则。"}],"progressIds":[],"configuration":null,"limitation":"不提供本网站用户当前税率/申报结论，不外推州税、企业税或中国规则。","evidenceStatus":"active"},{"id":"W-F12-taxcalcbench","statement":"固定TY25库保留复杂PDF和州/联邦计算的模型产物，且区分整份严格正确与逐行分。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-FIN-TAXCALCBENCH","locator":"TY2025表格、结果口径与Known issues；固定2026-10-19 commit","relation":"partial","supports":"固定TY25库保留复杂PDF和州/联邦计算的模型产物，且区分整份严格正确与逐行分。","limit":"通常每配置单次保存结果，按最佳thinking设置展示，部分模型覆盖不足；生成成本不等于端到端报税成本。"}],"progressIds":["P-FIN-TAX-ARTIFACT"],"configuration":"50个TY25案例，常为每配置1份保存结果、按best-thinking展示；快照不推新案例稳定性。","limitation":"通常每配置单次保存结果，按最佳thinking设置展示，部分模型覆盖不足；生成成本不等于端到端报税成本。","evidenceStatus":"active"},{"id":"W-F12-tax-saved-case","statement":"可以直接打开一份保存成功的逐行税表评估，检查正确字段与实际记录成本。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-FIN-TAX-SAVED-CASE","locator":"逐行expected/actual及Strictly correct return、API Usage and Cost","relation":"partial","supports":"可以直接打开一份保存成功的逐行税表评估，检查正确字段与实际记录成本。","limit":"单个作者自评产物、未本站重放；选中成功例不证明新案例稳定性或提交成功。"}],"progressIds":[],"configuration":"ty25-ca-001，gpt-6.1-sol high无web；报告严格通过、115.41s、$0.093068(litellm_estimate)，仅保存完成次成本，未本站复跑。","limitation":"单个作者自评产物、未本站重放；选中成功例不证明新案例稳定性或提交成功。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"已核对公开代码/文档入口，未安装或重放。Agent模型和数据接口需另接；账户、API和计算费用未测。","tools":[{"name":"Tax-Calculator固定模型与可选Agent入口","url":"https://github.com/PSLmodels/Tax-Calculator/blob/94c7f5f85a6d5a0c5f703179a8c4f2566d27b0cc/agents/README.md","version":"固定commit见URL","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"}]},"variants":[{"title":"固定税年简单个税","inputDifference":"规则和凭据齐","assessmentChange":"可阶段委派核对与计算","acceptanceChange":"逐行比官方表单"},{"title":"跨境/企业/更正","inputDifference":"新增法域、税种或原申报依赖","assessmentChange":"先补官方规则，不能迁移现有成绩","acceptanceChange":"单独专家规则核对与更正差异"},{"title":"TY25公开研究产物（与本页TY24规则例不同）","inputDifference":"固定TY25 PDF与州/联邦任务，需要TY25对应官方规则；不采用本页IRS2024规则","assessmentChange":"保存成功例与整表/逐行统计可检查，仅支持候选工作法；不能证明本站TY24例或用户新案例通过","acceptanceChange":"按TY25原数据和reference逐行核对、重查规则版本；同配置重试与提交回执另验"}],"evidenceStatus":"active"},{"id":"W-F13","domainId":"finance","title":"合规案例与控制：规则、事实、调查和整改","goal":"交付规则版本与事实一致、能复查异常理由的候选案件或控制检查包。","trigger":"新业务审查、客户尽调、异常警报或整改复检。","scope":"须输入法域期间与内部政策；本文不声明当前AML门槛或法律要求。合成数据只用于验收器，不能充当真实犯罪标签。","inputs":[{"name":"固定规则与检查目标","detail":"法域期间、政策版本、控制及升级要求"},{"name":"授权事实材料","detail":"实体/受益人资料、交易、原始文件和历史调查"}],"deliverables":["规则到控制映射","客户档案、实体关联和缺件","异常解释、冲突和反证","调查时间线与候选处理理由","整改复验、沟通或上报候选及状态"],"acceptance":[{"criterion":"规则和事实分开","method":"检查每项结论定位条款与原文件，保留不适用情况","failure":"新闻风险直接变成违法"},{"criterion":"异常有反证","method":"与正常大额/复杂关系负例比较，复核实体误匹配","failure":"同名即关联、复杂交易即洗钱"},{"criterion":"调查和整改闭环","method":"核对证据取得、升级及后续控制测试，回执另列","failure":"生成报告就宣称控制有效"}],"assessment":{"recommendation":"可用于产生和挑战候选；交付规则版本与事实一致、能复查异常理由的候选案件或控制检查包。 接收依据见下方实质检查，不以流畅解释替代结果。","conditions":["须输入法域期间与内部政策；本文不声明当前AML门槛或法律要求。合成数据只用于验收器，不能充当真实犯罪标签。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"AMLSim等公开方法支持生成带标签的交易图与异常场景，以及对规则/调查候选作可检查验证；不是AI调查成功率。完整专业调查仍缺直接AI结果，采用建议依据方法与独立接收。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["合成器不是AI调查器；合成标签不能证明真实洗钱、可疑报告法律结论或现实漏报率。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"candidate-exploration","rationale":"可将事实回链、规则匹配与反证整理交给Agent形成调查候选，依据是可核对的方法与资料组织。完整取证、制度适用和整改有效性必须用专属案件及真实状态独立接收。","frontier":"已有合成交易与规则检查方法，可组织事实、疑点和反证的调查候选。没有直接覆盖完整取证、欺诈确认、制度适用及整改有效性的本任务AI结果；下一步用专业案件和最终状态验证。","nextValidation":"固定一条内部政策，准备同名客户、正常复杂交易、缺受益人和真实异常的正负案例，分别测事实遗漏、误报与理由。","futureSignals":["在同名实体、正常复杂交易和缺受益人新案例中同时控制误漏并保留证据，可扩案例处理","名单版本变化、实体误合并或把疑点当违法事实时，重审相应规则与调查结论"]},"workflow":{"before":"规则文本、客户文件、筛查命中和调查记录彼此分离。","now":"Agent按固定政策组织实体证据、规则结果和竞争解释，调用授权筛查或数据查询；新证据可反转命中，真实调查或整改状态独立记录，合成样例仅校准错误检查。","retained":["规则和事实分开","异常有反证","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现新闻风险直接变成违法等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"同名企业共享收款人，证件编号不同。","output":"保留不同实体并列需要核实的资金关系。","failure":"仅凭名字合并并报告确定洗钱。"},"sourceIds":["WU-FIN-AMLSIM","WU-FIN-KYC-TEMPLATE"],"evidenceLinks":[{"sourceId":"WU-FIN-AMLSIM","supports":"AMLSim提供含已知模式的合成数据。","notCovered":"合成器不是AI调查器；合成标签不能证明真实洗钱、可疑报告法律结论或现实漏报率。"},{"sourceId":"WU-FIN-KYC-TEMPLATE","supports":"客户实体、规则结果、命中和升级包参考。","notCovered":"不涵盖真实交易监测或全部合规；需内部规则、数据许可和筛查接口。"}],"resourceRefs":[],"career":{"direction":"合规案例与控制：规则、事实、调查和整改","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"固定一条内部政策，准备同名客户、正常复杂交易、缺受益人和真实异常的正负案例，分别测事实遗漏、误报与理由。"},"team":{"bottleneck":"先区分资料、计算和现实状态/假设哪一项造成返工。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"固定一条内部政策，准备同名客户、正常复杂交易、缺受益人和真实异常的正负案例，分别测事实遗漏、误报与理由。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F13-method","statement":"AMLSim提供含已知模式的合成数据。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-AMLSIM","locator":"目标、运行和scripts/validation/validate_alerts.py","relation":"method","supports":"AMLSim提供含已知模式的合成数据。","limit":"合成器不是AI调查器；合成标签不能证明真实洗钱、可疑报告法律结论或现实漏报率。"}],"progressIds":[],"configuration":null,"limitation":"合成器不是AI调查器；合成标签不能证明真实洗钱、可疑报告法律结论或现实漏报率。","evidenceStatus":"active"},{"id":"W-F13-adoption","statement":"可依据既定规则、事实回链和可检查交易方法组织调查候选；完整取证、合规裁决及整改有效性需要专门验证。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-AMLSIM","locator":"目标、运行和scripts/validation/validate_alerts.py","relation":"adjacent","supports":"能提供带已知模式的合成数据以设计案例正负例。","limit":"不把方法存在当作AI已完成本任务。"},{"sourceId":"WU-FIN-KYC-TEMPLATE","locator":"What you produce / new-client onboarding not transaction monitoring","relation":"partial","supports":"客户实体、规则结果、命中和升级包参考。","limit":"不涵盖真实交易监测或全部合规；需内部规则、数据许可和筛查接口。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"},{"id":"W-F13-wu-fin-kyc-template","statement":"客户实体、规则结果、命中和升级包参考。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-KYC-TEMPLATE","locator":"What you produce / new-client onboarding not transaction monitoring","relation":"method","supports":"客户实体、规则结果、命中和升级包参考。","limit":"不涵盖真实交易监测或全部合规；需内部规则、数据许可和筛查接口。"}],"progressIds":[],"configuration":null,"limitation":"不涵盖真实交易监测或全部合规；需内部规则、数据许可和筛查接口。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"已核对公开代码/文档入口，未安装或重放。Agent模型和数据接口需另接；账户、API和计算费用未测。","tools":[{"name":"IBM AMLSim固定合成交易与已知模式","url":"https://github.com/IBM/AMLSim/blob/7338a4bcb1af9bcfea2201ad7daccfe2a4d569ca/README.md","version":"固定commit见URL","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"},{"name":"官方固定KYC案例包模板","url":"https://github.com/anthropics/financial-services/blob/574ed3624aebd0418c7e96cd101262f30210ab26/plugins/agent-plugins/kyc-screener/agents/kyc-screener.md","version":"固定commit见URL","access":"公开参考；本任务未安装运行；需另接模型/内部数据，实际费用未知。"}]},"variants":[{"title":"KYC资料检查","inputDifference":"实体证件和受益人链","assessmentChange":"可尝试资料与条件匹配候选","acceptanceChange":"核实实体同一性和缺页"},{"title":"异常交易/合规调查","inputDifference":"时序与关系证据","assessmentChange":"输出竞争解释与调查计划，追踪真实取证","acceptanceChange":"不能把合成标签当现实犯罪结论"}],"evidenceStatus":"active"},{"id":"W-F14","domainId":"finance","title":"交易执行资料包：尽调、披露和交割条件","goal":"使一项融资或并购交易的分析、文件版本、问题与条件可追踪至可验证交割状态。","trigger":"交易由初筛进入尽调、材料准备或交割协调。","scope":"估值可引用F08；本任务接收是资料及条件闭环，法律意见、签约授权和实际交割状态分别记录。","inputs":[{"name":"交易目标和结构","detail":"融资/并购类型、参与方、条款、期限与约束"},{"name":"资料室与条件表","detail":"版本化证据、尽调问题、披露及签署交割清单"}],"deliverables":["交易结构与选项说明","尽调问题、红旗及缺件","模型到披露/演示材料一致更新","条件责任时间表与争议","授权范围内交割状态及未完成事项"],"acceptance":[{"criterion":"资料版本一致","method":"从披露数字和条款反查模型与原件日期","failure":"旧版债务数留在路演材料"},{"criterion":"尽调反证保留","method":"核对风险问题是否有关闭证据及相反资料","failure":"口头答复当已解决"},{"criterion":"交割条件真实闭环","method":"逐条件检查签署原件、资金到账及状态","failure":"勾清单代替实际交割"}],"assessment":{"recommendation":"可分阶段委派；使一项融资或并购交易的分析、文件版本、问题与条件可追踪至可验证交割状态。 接收依据见下方实质检查，不以流畅解释替代结果。","conditions":["估值可引用F08；本任务接收是资料及条件闭环，法律意见、签约授权和实际交割状态分别记录。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"交易模型代码可支撑经济计算，资料/条件闭环是编辑迁移；尚无本交易完整尽调、披露和交割AI成绩。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"staged-delivery","rationale":"公开金融工具与文档处理支持候选资料整合；多方真实承诺、法律结论和交割接口尚需专项验证，不据此否认可自动协调。","frontier":"能检查模型及文件生成路径，尚未取得多方真实承诺、法律文件和资金交割的整包可重放AI产物；因此重点接收版本、问题及条件而非宣布交易完成。","nextValidation":"用脱敏交易资料植入旧文件、未满足条件和冲突签署日期，核对材料同步、红旗及阻断交割的理由。","futureSignals":["新交易包中模型、披露和条件同步正确，签署到账均能独立确认，可扩至授权协调","旧版材料、红旗被口头关闭或资金未到却标交割时，阻断该交易阶段"]},"workflow":{"before":"资料室、模型、推介文件和交割表各自更新，多方反复传背景。","now":"Agent围绕交易目标管理版本、问题与条件，把模型更新传到材料，按原始证据关闭问题；尚未签署或未到账就保留未决，内部组织可动态改计划。","retained":["资料版本一致","尽调反证保留","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现旧版债务数留在路演材料等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"融资模型已更新债务，但条款附件仍用旧数。","output":"同步材料并将尚未签署的条件列为未关闭。","failure":"文件齐全就宣布融资交割成功。"},"sourceIds":["WU-FIN-FINROBOT-DCF"],"evidenceLinks":[{"sourceId":"WU-FIN-FINROBOT-DCF","supports":"可审阅计算可支撑交易模型部分。","notCovered":"仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。"}],"resourceRefs":[],"career":{"direction":"交易执行资料包：尽调、披露和交割条件","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"用脱敏交易资料植入旧文件、未满足条件和冲突签署日期，核对材料同步、红旗及阻断交割的理由。"},"team":{"bottleneck":"先区分资料、计算和现实状态/假设哪一项造成返工。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"用脱敏交易资料植入旧文件、未满足条件和冲突签署日期，核对材料同步、红旗及阻断交割的理由。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F14-method","statement":"可审阅计算可支撑交易模型部分。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-FINROBOT-DCF","locator":"calculate_dcf；terminal growth/WACC与terminal FCF边界","relation":"method","supports":"可审阅计算可支撑交易模型部分。","limit":"仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。"}],"progressIds":[],"configuration":null,"limitation":"仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。","evidenceStatus":"active"},{"id":"W-F14-adoption","statement":"公开金融工具与文档处理支持候选资料整合；多方真实承诺、法律结论和交割接口尚需专项验证，不据此否认可自动协调。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-FINROBOT-DCF","locator":"calculate_dcf；terminal growth/WACC与terminal FCF边界","relation":"adjacent","supports":"确定性DCF实现可审阅，可与Agent的资料和假设生成分开。","limit":"不把方法存在当作AI已完成本任务。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"已核对公开代码/文档入口，未安装或重放。Agent模型和数据接口需另接；账户、API和计算费用未测。","tools":[{"name":"FinRobot固定DCF计算与边界处理","url":"https://github.com/AI4Finance-Foundation/FinRobot/blob/2717499b8e30f242640af08c4ad9afd1113c2d45/finrobot_desktop/finrobot/engine/compute/operators/dcf.py#L48-L177","version":"固定commit见URL","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"}]},"variants":[{"title":"推介或初筛","inputDifference":"目标名单和公开资料","assessmentChange":"可整包候选材料","acceptanceChange":"核对可比和引用"},{"title":"执行与交割","inputDifference":"非公开证据和多方承诺","assessmentChange":"在权限内分阶段追踪并核验状态","acceptanceChange":"法律、资金和签署条件各有证据"}],"evidenceStatus":"active"},{"id":"W-F15","domainId":"finance","title":"量化研究：时间一致数据到可反驳回测","goal":"交付固定信息集、成本约束和独立留出下可重放的策略或信号研究。","trigger":"要检验金融信号、更新策略或解释失效。","scope":"研究回测、模拟执行与真实收益区分；可动态探索，但独立检验集和预算须留存。","inputs":[{"name":"时点数据与假设","detail":"版本、可得时间、退市和复权处理"},{"name":"研究约束","detail":"策略定义、样本划分、基线、成本、市场限制和搜索预算"}],"deliverables":["时间一致数据及诊断","策略代码与全部实验记录","成本前后、基线和样本外结果","漂移、失败和停止条件","可交给执行验证的研究包"],"acceptance":[{"criterion":"无未来信息","method":"按as-of时间重建样本，做标签和特征时间审计","failure":"用晚发布财报预测早期价格"},{"criterion":"样本外不被反复消耗","method":"锁定留出并保存试验次数和选择过程","failure":"反复调参后仍叫独立样本"},{"criterion":"收益满足约束","method":"独立重算持仓现金、费用和成交限制，比较基线","failure":"零费用/无成交量限制虚增收益"}],"assessment":{"recommendation":"可用于产生和挑战候选；交付固定信息集、成本约束和独立留出下可重放的策略或信号研究。 接收依据见下方实质检查，不以流畅解释替代结果。","conditions":["研究回测、模拟执行与真实收益区分；可动态探索，但独立检验集和预算须留存。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"InvestorBench提供直接Agent历史决策回放，Qlib提供可配置重算；单资产历史表现不等于当前策略成功或实盘。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["方法和示例；不证明Agent策略有效，回测不等于实盘成交。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"candidate-exploration","rationale":"直接Agent历史决策回放与可重放工具支持候选策略研究；旧时期、模型知识和单资产范围不同，需严格新留出、成本及市场约束后再扩大。","frontier":"已有旧模型/历史窗的Agent结果及基线比较；当期数据、模型知识泄漏、多重搜索和可成交成本尚需新留出验证，不能直接继承收益。","nextValidation":"先重放固定基线，再故意移入未来财报、遗漏退市及漏算费用，要求诊断；只在新留出段比较候选。","futureSignals":["冻结新留出、登记全部搜索后仍优于合法基线并满足成本容量，才扩研究到受控执行","性能只在已用历史段存在，或考虑费用/退市即失效时，保留失败而不改测试口径"]},"workflow":{"before":"想法、取数、回测、挑选和报告逐次手动迭代。","now":"Agent可自主生成候选、重建时点数据和调用回测器，利用反例及成本反馈修订；独立留出与全部尝试日志隔离保留，不能对测试集无限搜索后报一次成功。","retained":["无未来信息","样本外不被反复消耗","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现用晚发布财报预测早期价格等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"基本面信号采用财报期末时间而非披露时间。","output":"重建信息可得时点并重报样本外及成本后结果。","failure":"回测很高就称具可部署alpha。"},"sourceIds":["WU-FIN-QLIB-BACKTEST","WU-FIN-INVESTORBENCH"],"evidenceLinks":[{"sourceId":"WU-FIN-QLIB-BACKTEST","supports":"固定Qlib文档可分别输出成本前后回测。","notCovered":"方法和示例；不证明Agent策略有效，回测不等于实盘成交。"},{"sourceId":"WU-FIN-INVESTORBENCH","supports":"已有Agent在单资产历史市场中完成研究决策回放和基线比较，不能外推当前可交易alpha。","notCovered":"单资产历史模拟，早期模型与有限市场时段；不支持组合、真实成交、当前模型alpha或利润保证。"}],"resourceRefs":[],"career":{"direction":"量化研究：时间一致数据到可反驳回测","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"先重放固定基线，再故意移入未来财报、遗漏退市及漏算费用，要求诊断；只在新留出段比较候选。"},"team":{"bottleneck":"先区分资料、计算和现实状态/假设哪一项造成返工。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"先重放固定基线，再故意移入未来财报、遗漏退市及漏算费用，要求诊断；只在新留出段比较候选。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F15-method","statement":"固定Qlib文档可分别输出成本前后回测。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-QLIB-BACKTEST","locator":"Running backtest；exchange_kwargs与cost分别输出","relation":"method","supports":"固定Qlib文档可分别输出成本前后回测。","limit":"方法和示例；不证明Agent策略有效，回测不等于实盘成交。"}],"progressIds":[],"configuration":null,"limitation":"方法和示例；不证明Agent策略有效，回测不等于实盘成交。","evidenceStatus":"active"},{"id":"W-F15-adoption","statement":"直接Agent历史决策回放与可重放工具支持候选策略研究；旧时期、模型知识和单资产范围不同，需严格新留出、成本及市场约束后再扩大。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-QLIB-BACKTEST","locator":"Running backtest；exchange_kwargs与cost分别输出","relation":"adjacent","supports":"可检查回测配置、仓位和成本分析路径。","limit":"不把方法存在当作AI已完成本任务。"},{"sourceId":"WU-FIN-INVESTORBENCH","locator":"§3—4、AppendixF/G与Limitations","relation":"partial","supports":"在固定历史市场环境比较Agent决策与基线。","limit":"单资产历史模拟，早期模型与有限市场时段；不支持组合、真实成交、当前模型alpha或利润保证。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"},{"id":"W-F15-investorbench","statement":"已有Agent在单资产历史市场中完成研究决策回放和基线比较，不能外推当前可交易alpha。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-INVESTORBENCH","locator":"§3—4、AppendixF/G与Limitations","relation":"partial","supports":"已有Agent在单资产历史市场中完成研究决策回放和基线比较，不能外推当前可交易alpha。","limit":"单资产历史模拟，早期模型与有限市场时段；不支持组合、真实成交、当前模型alpha或利润保证。"}],"progressIds":[],"configuration":"论文13模型；股票/加密/ETF不同固定历史窗，warmup与test分开；不是现代交易所撮合或新留出获利。","limitation":"单资产历史模拟，早期模型与有限市场时段；不支持组合、真实成交、当前模型alpha或利润保证。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"已核对公开代码/文档入口，未安装或重放。Agent模型和数据接口需另接；账户、API和计算费用未测。","tools":[{"name":"Qlib固定文档：成本前后回测","url":"https://github.com/microsoft/qlib/blob/54355232463878d2eebb91fe0ee5fa7fa1f5976c/docs/component/strategy.rst#L137-L260","version":"固定commit见URL","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"}]},"variants":[{"title":"低频信号","inputDifference":"日级数据和固定策略","assessmentChange":"可研究候选并复算","acceptanceChange":"成本、复权和退市审计"},{"title":"高频或执行策略","inputDifference":"分钟以下数据、队列和冲击","assessmentChange":"增加仿真与真实撮合差异验证","acceptanceChange":"不能把低频回测直接外推实盘"}],"evidenceStatus":"active"},{"id":"W-F16","domainId":"finance","title":"定价与风险场景：合约、校准到可复算敞口","goal":"交付日期约定一致的估值、敏感性、压力和归因，并标出模型失效条件。","trigger":"新合约、市场变动、风险限额或资金套保需要重新估值。","scope":"工具主要覆盖市场和交易对手模型；信用组合、流动性和操作风险需专用模型和证据变体。","inputs":[{"name":"合约及市场快照","detail":"现金流、日历、计息、曲线、波动、币种和基准日"},{"name":"模型与情景","detail":"校准方法、风险因子、参数、容差、组合和压力定义"}],"deliverables":["可重放定价和校准残差","风险敞口、敏感性和压力结果","损益归因、模型差异与限制","限额异常与缓释候选"],"acceptance":[{"criterion":"合约语义不变","method":"核对日历、计息、方向、币种和行权，手算简化特例","failure":"日期错而算法数值稳定"},{"criterion":"数值与模型各验","method":"独立基准、网格收敛和替代模型对比，保留校准残差","failure":"校准好就当未来分布真值"},{"criterion":"组合压力可解释","method":"因子冲击、对冲、非线性和归因残差重算","failure":"所有敞口直接相加漏关联"}],"assessment":{"recommendation":"可分阶段委派；交付日期约定一致的估值、敏感性、压力和归因，并标出模型失效条件。 接收依据见下方实质检查，不以流畅解释替代结果。","conditions":["工具主要覆盖市场和交易对手模型；信用组合、流动性和操作风险需专用模型和证据变体。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"FinGAIA公开例只为历史回撤的资料获取与数值计算提供局部参照；实际任务见下方具体主张。完整定价与风险控制的整包采用另依公开方法、明确条件和独立验收提出，不从角色名推出专业能力。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["不是任何模型均正确的证明；市场、合约与日期约定需冻结。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"staged-delivery","rationale":"历史统计、合约计算与模型校准分别接收。已有数值方法可支持委派可复算风险场景；FinGAIA只提供局部历史回撤参照，模型选择与现实覆盖需独立挑战。","frontier":"Agent历史回撤计算已有局部评估；QuantLib/ORE等有公开定价与风险运行方法。尚不能由该历史统计任务推出完整定价模型、组合尾部或限额行动可靠。","nextValidation":"取一份固定合约与市场数据，比较两个引擎，再改变日历、方向和曲线，检查误差与解释是否一致。","futureSignals":["新合约与压力中，语义、独立引擎及误差控制都一致，并能识别模型失效后扩围","日期方向错误、校准好却外推错误分布，或遗漏关联敞口时，拒收相应估值和风险结论"]},"workflow":{"before":"合约解释、曲线、数值求解和风险报表在不同表格间传递。","now":"Agent围绕同一合约与快照构造校准和压力，调用专业引擎并比较独立方法；价格差异驱动回查日期、方向、参数和模型，不只增加推理文本。","retained":["合约语义不变","数值与模型各验","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现日期错而算法数值稳定等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"利率互换的收付方向被颠倒。","output":"独立现金流核对揭示NPV和DV01符号变化。","failure":"价格绝对值相同就判正确。"},"sourceIds":["WU-FIN-QUANTLIB-PRICING","WU-FIN-ORE-STRESS","WU-FIN-FINGAIA"],"evidenceLinks":[{"sourceId":"WU-FIN-QUANTLIB-PRICING","supports":"固定QuantLib算例可对照多种定价引擎。","notCovered":"不是任何模型均正确的证明；市场、合约与日期约定需冻结。"},{"sourceId":"WU-FIN-ORE-STRESS","supports":"可检查的定价风险情景输入和运行入口。","notCovered":"示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。"},{"sourceId":"WU-FIN-FINGAIA","supports":"Fig8对三个指定指数计算2014-03-01至2024-03-01历史最大回撤，只支持历史数据与数值计算。","notCovered":"衍生品定价、尾部风险校准或机构风险控制及本业务端到端可靠性。"}],"resourceRefs":[],"career":{"direction":"定价与风险场景：合约、校准到可复算敞口","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"取一份固定合约与市场数据，比较两个引擎，再改变日历、方向和曲线，检查误差与解释是否一致。"},"team":{"bottleneck":"先区分资料、计算和现实状态/假设哪一项造成返工。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"取一份固定合约与市场数据，比较两个引擎，再改变日历、方向和曲线，检查误差与解释是否一致。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F16-method","statement":"固定QuantLib算例可对照多种定价引擎。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-QUANTLIB-PRICING","locator":"模型校准及tree/fdm NPV输出","relation":"method","supports":"固定QuantLib算例可对照多种定价引擎。","limit":"不是任何模型均正确的证明；市场、合约与日期约定需冻结。"}],"progressIds":[],"configuration":null,"limitation":"不是任何模型均正确的证明；市场、合约与日期约定需冻结。","evidenceStatus":"active"},{"id":"W-F16-adoption","statement":"可把合约和市场输入交给代理组织可复算定价/风险场景；模型、校准和最终限额状态分别验收，不由历史回撤题推定完成。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-QUANTLIB-PRICING","locator":"模型校准及tree/fdm NPV输出","relation":"adjacent","supports":"可检查数值定价与不同求解路线对照。","limit":"不把方法存在当作AI已完成本任务。"},{"sourceId":"WU-FIN-FINGAIA","locator":"Table2、Agents/Evaluation Methods、Table3及AppendixB","relation":"adjacent","supports":"历史回撤的资料获取与数值计算可作为交付中取材/计算环节的参照。","limit":"完整采用建议为编辑推断；Appendix A Fig8公开例计算沪深300、恒生及S&P500三指数的历史最大回撤。它不验证衍生品定价、尾部风险校准或机构风险控制。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"},{"id":"W-F16-ore-stress","statement":"可检查的定价风险情景输入和运行入口。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-FIN-ORE-STRESS","locator":"XVA Stress Testing / XVA P&L Explain；run_stress.py与配置文件","relation":"partial","supports":"可检查的定价风险情景输入和运行入口。","limit":"示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。"}],"progressIds":[],"configuration":null,"limitation":"示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。","evidenceStatus":"active"},{"id":"W-F16-fingaia","statement":"Appendix A Fig8公开例计算沪深300、恒生及S&P500三指数2014-03-01至2024-03-01最大回撤。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-FINGAIA","locator":"Appendix A Fig8","relation":"partial","supports":"历史回撤的资料获取与数值计算的局部评估参照。","limit":"角色标签不等于实际专业任务；公开例不能证明衍生品定价、尾部风险校准或机构风险控制，不把分类分数当本任务成功率。"}],"progressIds":["P-FIN-SCENARIOS"],"configuration":"FinGAIA v1零样本网页/附件原环境；论文角色分类需对照Appendix A实际样例；正文与附表口径有差异，不使用该标签分数为本专业成功率。","limitation":"角色标签不等于实际专业任务；公开例不能证明衍生品定价、尾部风险校准或机构风险控制，不把分类分数当本任务成功率。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"已核对公开代码/文档入口，未安装或重放。Agent模型和数据接口需另接；账户、API和计算费用未测。","tools":[{"name":"QuantLib固定BermudanSwaption示例","url":"https://github.com/lballabio/QuantLib/blob/eaeac565149ced7c1e0cd515c0fa3ed0b89f2b39/Examples/BermudanSwaption/BermudanSwaption.cpp#L157-L290","version":"固定commit见URL","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"}]},"variants":[{"title":"单合约估值","inputDifference":"结构和市场可冻结","assessmentChange":"可分阶段委派计算包","acceptanceChange":"独立价格与敏感性"},{"title":"组合/尾部风险","inputDifference":"相关、流动性及极端分布","assessmentChange":"补专门模型和样本后扩围","acceptanceChange":"压力传播、关联和缺失敞口另验"}],"evidenceStatus":"active"},{"id":"W-F17","domainId":"finance","title":"独立模型挑战：概念、实现、表现与使用范围","goal":"交付能独立重现并挑战开发结论的验证报告及整改证据。","trigger":"模型初次准入、重要变更或定期独立验证。","scope":"逻辑/数值正确只占一层；用途、统计表现、治理和独立性都属于结果。","inputs":[{"name":"模型完整包","detail":"用途、理论、代码、数据、参数和开发报告"},{"name":"验证要求","detail":"独立基准、样本、容差、重要性与使用限制"}],"deliverables":["概念和用途挑战","独立实现/数据差异","样本外、基准和压力结果","治理依赖、覆盖及变更检查","发现等级、使用条件及整改复验"],"acceptance":[{"criterion":"独立性可核对","method":"不同实现或依据从原始数据复现，避免照抄开发输出","failure":"与开发共用错误函数当独立验证"},{"criterion":"有效性不是算术","method":"检验概念、校准、性能、适用分布和限制","failure":"回测通过即允许全部用途"},{"criterion":"问题改变使用","method":"将严重发现映射限制、补救与复验","failure":"报告有发现但仍无条件批准"}],"assessment":{"recommendation":"可用于产生和挑战候选；交付能独立重现并挑战开发结论的验证报告及整改证据。 接收依据见下方实质检查，不以流畅解释替代结果。","conditions":["逻辑/数值正确只占一层；用途、统计表现、治理和独立性都属于结果。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"独立验证任务需要与开发区别的依据；风险和回测工具只支持部分重算，尚无整包金融模型AI验证实测。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"candidate-exploration","rationale":"开源计算和基准让独立挑战可落地；尚无这类整包AI独立验证成绩，适合可追踪候选报告和专项复算。","frontier":"可以构造独立计算与反例接收，但未证实AI能同时发现概念、数据、表现和治理缺陷；所以本包要求实质发现改变使用边界。","nextValidation":"给模型包植入符号错、训练泄漏、超用途和变更未登记，独立检查是否分别发现而非只重复报告。","futureSignals":["在预先植入的泄漏、用途和变更缺陷及未见模型中持续发现问题，可扩大独立挑战","与开发共用错误、只复述报告或发现不影响批准时，不能算完成验证"]},"workflow":{"before":"开发报告审读、重算和治理清单分开做，易复制开发假设。","now":"Agent从用途和原始数据重建独立基准，按差异选择反例、压力和留出检查；即使数值一致，也继续挑战标签、用途和变更，结论映射实际使用限制。","retained":["独立性可核对","有效性不是算术","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现与开发共用错误函数当独立验证等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"模型实现一致但验证集被训练使用。","output":"数值层通过、统计层失败，明确限制与补测。","failure":"重算一致便建议无条件使用。"},"sourceIds":["WU-FIN-ORE-STRESS","WU-FIN-QLIB-BACKTEST"],"evidenceLinks":[{"sourceId":"WU-FIN-ORE-STRESS","supports":"公开风险算例可作为独立对照材料。","notCovered":"示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。"},{"sourceId":"WU-FIN-QLIB-BACKTEST","supports":"可检查回测配置、仓位和成本分析路径。","notCovered":"方法和示例；不证明Agent策略有效，回测不等于实盘成交。"}],"resourceRefs":[],"career":{"direction":"独立模型挑战：概念、实现、表现与使用范围","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"给模型包植入符号错、训练泄漏、超用途和变更未登记，独立检查是否分别发现而非只重复报告。"},"team":{"bottleneck":"先区分资料、计算和现实状态/假设哪一项造成返工。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"给模型包植入符号错、训练泄漏、超用途和变更未登记，独立检查是否分别发现而非只重复报告。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F17-method","statement":"公开风险算例可作为独立对照材料。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-ORE-STRESS","locator":"XVA Stress Testing / XVA P&L Explain；run_stress.py与配置文件","relation":"method","supports":"公开风险算例可作为独立对照材料。","limit":"示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。"}],"progressIds":[],"configuration":null,"limitation":"示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。","evidenceStatus":"active"},{"id":"W-F17-adoption","statement":"开源计算和基准让独立挑战可落地；尚无这类整包AI独立验证成绩，适合可追踪候选报告和专项复算。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-ORE-STRESS","locator":"XVA Stress Testing / XVA P&L Explain；run_stress.py与配置文件","relation":"adjacent","supports":"可检查的定价风险情景输入和运行入口。","limit":"不把方法存在当作AI已完成本任务。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"},{"id":"W-F17-qlib-backtest","statement":"可检查回测配置、仓位和成本分析路径。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-FIN-QLIB-BACKTEST","locator":"Running backtest；exchange_kwargs与cost分别输出","relation":"partial","supports":"可检查回测配置、仓位和成本分析路径。","limit":"方法和示例；不证明Agent策略有效，回测不等于实盘成交。"}],"progressIds":[],"configuration":null,"limitation":"方法和示例；不证明Agent策略有效，回测不等于实盘成交。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"已核对公开代码/文档入口，未安装或重放。Agent模型和数据接口需另接；账户、API和计算费用未测。","tools":[{"name":"ORE固定XVA压力与损益归因算例","url":"https://github.com/OpenSourceRisk/Engine/blob/3d75a69087911a7e2cf1e6882da7bc27135c0762/Examples/XvaRisk/Readme.md","version":"固定commit见URL","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"}]},"variants":[{"title":"定价模型","inputDifference":"合约、曲线和数值引擎","assessmentChange":"可委派独立复算与差异诊断","acceptanceChange":"替代引擎和容差"},{"title":"统计信用或风险模型","inputDifference":"数据分布、标签和用途","assessmentChange":"增加有效性挑战及独立留出","acceptanceChange":"稳定性、校准、漂移和治理均需结论"}],"evidenceStatus":"active"},{"id":"W-F18","domainId":"finance","title":"授权内交易执行：订单、头寸与异常恢复","goal":"交付满足限额和市场约束的执行记录、状态和成本解释。","trigger":"已有交易意图需要执行、改撤单或处理市场异常。","scope":"执行方案、沙盒模拟和真实券商状态分开；有授权与连接时研究可扩至真实操作，不将签字制度当技术永久边界。","inputs":[{"name":"执行意图和限额","detail":"方向、规模、期限、基准、账户及风险"},{"name":"市场和接口","detail":"行情时点、成交规则、订单状态、权限和故障恢复"}],"deliverables":["执行计划和可检验约束","订单/成交/撤单事件账","现金、头寸和限额状态","滑点、费用、冲击和未执行原因","异常暂停与恢复证据"],"acceptance":[{"criterion":"状态与头寸守恒","method":"按事件重放部分成交、撤单失败和重试，对照券商回报","failure":"重复订单或把撤单请求当成功"},{"criterion":"成本与市场约束真实","method":"检查成交量、价差、限价和基准时点","failure":"使用事后最优价"},{"criterion":"异常恢复受控","method":"沙盒断连、乱序回报和超限负例必须暂停并对账","failure":"不知状态继续追单"}],"assessment":{"recommendation":"可分阶段委派；交付满足限额和市场约束的执行记录、状态和成本解释。 接收依据见下方实质检查，不以流畅解释替代结果。","conditions":["执行方案、沙盒模拟和真实券商状态分开；有授权与连接时研究可扩至真实操作，不将签字制度当技术永久边界。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"Qlib执行仿真有状态和市场约束方法；它不是实盘接口，也没有本任务AI订单可靠性证据。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["不是交易所或券商真实接口，成交和市场冲击模型须单独核对。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"staged-delivery","rationale":"仿真和事件日志适合验证执行候选；尚无真实接口AI执行可靠性证据，先沙盒校准再按授权扩围。","frontier":"可交可重放执行计划、事件账和成本分析；真实成交/撤单、超限与断连恢复仍需券商沙盒，再按授权扩大。","nextValidation":"重放部分成交、重复回报、断连和限额变化，独立核对订单、头寸和成本，不以回测报表代实盘回报。","futureSignals":["在部分成交、乱序回报和未知状态下头寸守恒且可恢复，才扩到真实账户限定执行","以回测价格当成交、重复下单或未知撤单状态仍继续交易时，暂停相关执行"]},"workflow":{"before":"交易计划、订单、成交、风险和执行分析分开记录。","now":"Agent围绕授权执行目标动态处理回报和市场条件，按状态变化改撤并重算头寸；遇未知状态、超限或断连先阻断再核对，仿真和真实执行各自留下证据。","retained":["状态与头寸守恒","成本与市场约束真实","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现重复订单或把撤单请求当成功等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"下单100股只成交40，撤单回报尚未到。","output":"记录40头寸及60未知待核，停止重复全量下单。","failure":"把100计成交并再买60。"},"sourceIds":["WU-FIN-QLIB-EXECUTION"],"evidenceLinks":[{"sourceId":"WU-FIN-QLIB-EXECUTION","supports":"Qlib公开执行仿真可配置市场约束。","notCovered":"不是交易所或券商真实接口，成交和市场冲击模型须单独核对。"}],"resourceRefs":[],"career":{"direction":"授权内交易执行：订单、头寸与异常恢复","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"重放部分成交、重复回报、断连和限额变化，独立核对订单、头寸和成本，不以回测报表代实盘回报。"},"team":{"bottleneck":"先区分资料、计算和现实状态/假设哪一项造成返工。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"重放部分成交、重复回报、断连和限额变化，独立核对订单、头寸和成本，不以回测报表代实盘回报。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F18-method","statement":"Qlib公开执行仿真可配置市场约束。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-QLIB-EXECUTION","locator":"backtesting配置、exchange、volume_threshold与simulator","relation":"method","supports":"Qlib公开执行仿真可配置市场约束。","limit":"不是交易所或券商真实接口，成交和市场冲击模型须单独核对。"}],"progressIds":[],"configuration":null,"limitation":"不是交易所或券商真实接口，成交和市场冲击模型须单独核对。","evidenceStatus":"active"},{"id":"W-F18-adoption","statement":"仿真和事件日志适合验证执行候选；尚无真实接口AI执行可靠性证据，先沙盒校准再按授权扩围。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-QLIB-EXECUTION","locator":"backtesting配置、exchange、volume_threshold与simulator","relation":"adjacent","supports":"可配置执行仿真环境。","limit":"不把方法存在当作AI已完成本任务。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"已核对公开代码/文档入口，未安装或重放。Agent模型和数据接口需另接；账户、API和计算费用未测。","tools":[{"name":"Qlib固定执行仿真配置","url":"https://github.com/microsoft/qlib/blob/54355232463878d2eebb91fe0ee5fa7fa1f5976c/docs/component/rl/quickstart.rst","version":"固定commit见URL","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"}]},"variants":[{"title":"模拟执行","inputDifference":"冻结市场和订单流","assessmentChange":"可委派生成并检验执行候选","acceptanceChange":"已知事件账及成本"},{"title":"真实交易","inputDifference":"实时接口和不可逆订单","assessmentChange":"先验证授权、幂等、限额和恢复","acceptanceChange":"以真实回报核对，不继承模拟成交保证"}],"evidenceStatus":"active"},{"id":"W-F19","domainId":"finance","title":"金融交易后运营：确认、交收和生命周期闭环","goal":"让交易、现金证券和生命周期事件在同一截止口径下可核对，异常明确未完成。","trigger":"交易成交后需要确认交收、对账或处理到期公司行动。","scope":"确认和账表匹配不等于真实交收；实际指令和状态只能按授权接口与回执接收。","inputs":[{"name":"交易和外部快照","detail":"成交条款、托管/银行账、账户和截止时间"},{"name":"生命周期与操作规则","detail":"结算日、事件权益、净额、费用和异常权限"}],"deliverables":["确认与差异清单","现金证券应收应付及实际状态","分红利息到期等权益计算","交收失败、补件、升级和操作留痕","修复与交接记录"],"acceptance":[{"criterion":"生命周期不重复不遗漏","method":"逐事件重放成交、分红、到期、撤销和调整","failure":"股息重复记账"},{"criterion":"账面和现实状态区分","method":"对照清算托管回执及现金证券两腿","failure":"只有账面匹配便标交收"},{"criterion":"异常恢复闭环","method":"重复指令、迟到回报和失败交收必须保留未决及修复证据","failure":"强制平账或提前结案"}],"assessment":{"recommendation":"可分阶段委派；让交易、现金证券和生命周期事件在同一截止口径下可核对，异常明确未完成。 接收依据见下方实质检查，不以流畅解释替代结果。","conditions":["确认和账表匹配不等于真实交收；实际指令和状态只能按授权接口与回执接收。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"账务校验和对账支持内部一致，金融生命周期及交收状态为专门编辑应用；不宣称AI已跑通清算系统。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["账务规则工具，不证明原始交易真实、会计政策正确或AI入账成功率。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"staged-delivery","rationale":"对账方法能用于异常发现；证券生命周期及实际交收需要专门状态模型，账务校验不是清算系统性能证据。","frontier":"可以把确认、两腿交收和公司行动组织成状态可验的候选；暂无可检查的本输入实际自动交收或复杂例外恢复结果。","nextValidation":"构造一份含部分交收、公司行动和迟到回报的交易包，独立核对现金证券两腿及未决项。","futureSignals":["不同产品的权益、现金证券回报和失败修复都经独立重放后，可扩大运营闭环","只核交易号或账面总额、隐去单腿失败或重复权益时，拒收该交收完成判断"]},"workflow":{"before":"运营逐表匹配成交、资金、证券和公司行动，再追例外。","now":"Agent围绕交易生命周期组织确认、权益、应付与回报，稳定计算直接用程序，按例外证据修复状态；实际未交收不因账面一致或报告完整被隐藏。","retained":["生命周期不重复不遗漏","账面和现实状态区分","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现股息重复记账等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"证券已记入但现金腿失败。","output":"保留部分交收和失败原因，形成修复清单。","failure":"交易号匹配即标全部交收。"},"sourceIds":["WU-FIN-BEAN-VALIDATION"],"evidenceLinks":[{"sourceId":"WU-FIN-BEAN-VALIDATION","supports":"账务校验可支持内部平衡核查。","notCovered":"账务规则工具，不证明原始交易真实、会计政策正确或AI入账成功率。"}],"resourceRefs":[],"career":{"direction":"金融交易后运营：确认、交收和生命周期闭环","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"构造一份含部分交收、公司行动和迟到回报的交易包，独立核对现金证券两腿及未决项。"},"team":{"bottleneck":"先区分资料、计算和现实状态/假设哪一项造成返工。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"构造一份含部分交收、公司行动和迟到回报的交易包，独立核对现金证券两腿及未决项。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F19-method","statement":"账务校验可支持内部平衡核查。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-BEAN-VALIDATION","locator":"validate_active_accounts、validate_currency_constraints、validate_check_transaction_balances","relation":"method","supports":"账务校验可支持内部平衡核查。","limit":"账务规则工具，不证明原始交易真实、会计政策正确或AI入账成功率。"}],"progressIds":[],"configuration":null,"limitation":"账务规则工具，不证明原始交易真实、会计政策正确或AI入账成功率。","evidenceStatus":"active"},{"id":"W-F19-adoption","statement":"对账方法能用于异常发现；证券生命周期及实际交收需要专门状态模型，账务校验不是清算系统性能证据。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-BEAN-VALIDATION","locator":"validate_active_accounts、validate_currency_constraints、validate_check_transaction_balances","relation":"adjacent","supports":"可检查确定性账务校验实现。","limit":"不把方法存在当作AI已完成本任务。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"已核对公开代码/文档入口，未安装或重放。Agent模型和数据接口需另接；账户、API和计算费用未测。","tools":[{"name":"Beancount固定代码：账户、币种与交易平衡校验","url":"https://github.com/beancount/beancount/blob/9747213802acb733ce53acc6e963f5961f47d6cd/beancount/ops/validation.py#L194-L410","version":"固定commit见URL","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"}]},"variants":[{"title":"普通现金证券","inputDifference":"条款和交收回报标准","assessmentChange":"候选匹配和异常可委派","acceptanceChange":"两腿状态及手续费"},{"title":"衍生品/公司行动","inputDifference":"多事件和条件权益","assessmentChange":"增加生命周期状态与专业规则验证","acceptanceChange":"不能只检查初始成交"}],"evidenceStatus":"active"},{"id":"W-F20","domainId":"finance","title":"保险负债估计：赔款准备金或寿险现金流可复算交付","goal":"交付固定估计日的保险负债候选、假设、不确定性及变动桥。","trigger":"估计日更新、负债评估或重大经验变化。","scope":"产险赔款发展三角与寿险长期现金流分别建模；此单元不覆盖新产品费率、资本或再保方案，也不等于个案准备金账务状态。","inputs":[{"name":"估计日负债基础","detail":"产险累计或增量赔款三角、尾部及通胀；或寿险逐期保单与生存退保费用资料"},{"name":"估计用途与上期版本","detail":"给定会计或内部估计口径、折现、上期假设与变动；制度适用需单独核对"}],"deliverables":["估计日数据与假设口径","产险准备金或寿险负债及独立复算","不确定性、尾部情景与前期估计变动桥"],"acceptance":[{"criterion":"数据与暴露对齐","method":"核对三角累计/增量、赔付/已报、承保期间和口径","failure":"累计数据当增量"},{"criterion":"计算和假设分层","method":"独立模型或已知样例复算，再检验尾部、死亡退保和费用假设","failure":"IBNR可运行就称准备金充足"},{"criterion":"用途与长期反馈区分","method":"按用途检查现金流、情景与经验偏差，不把短期拟合当长期校准","failure":"用单一历史均值覆盖新产品尾部"}],"assessment":{"recommendation":"可委派固定估计日的负债候选及变动解释，按产险三角或寿险逐期现金流分别独立复算。","conditions":["产险发展三角和寿险现金流有不同模型；工具仅支持局部计算，不能外推个体承保或各地法定口径。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"INS-ActBench是精算LLM直接研究，chainladder/lifelib是确定性方法；前者按题和答案格评分，不能由其均分推整模型交付。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["不证明AI会选对方法、尾部假设或真实准备金充分性。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"staged-delivery","rationale":"可接收固定估计日的可复算负债候选；Mack只覆盖产险三角，寿险用逐期模型。INS-ActBench展示专业实作，未展示本机构尾部或真实准备金充分。","frontier":"已有表格/R实作而非只有考试知识；真实保单结构、尾部假设、产寿险模型选择和多轮经验回顾仍需新样本，当前建议整包候选加独立复算。","nextValidation":"产险固定一组三角独立复算IBNR并注入累计增量错标；寿险固定一组保单核对逐期现金流、折现与退保压力；两者都解释上期到本期变化。","futureSignals":["新三角和不同尾部的历期回看同时通过，才扩大产险准备金采用","寿险新产品现金流与假设压力通过，才扩大长期负债；不能由Mack结果代替"]},"workflow":{"before":"经验数据、假设选择、产寿险模型、压力和报告分处工具。","now":"Agent按负债估计目的选择三角或逐期现金流模型，核对估计日数据，运行候选和尾部敏感性，并解释前期到本期变动；经验基础与定价由独立交付承接。","retained":["数据与暴露对齐","计算和假设分层","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现累计数据当增量等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"给定同一赔款三角但误标为增量。","output":"先辨识口径，再复算准备金、误差与尾部情景。","failure":"算法有输出就作为足额准备金。"},"sourceIds":["WU-FIN-CHAIN-MACK","WU-FIN-LIFELIB","WU-FIN-INS-ACTBENCH"],"evidenceLinks":[{"sourceId":"WU-FIN-CHAIN-MACK","supports":"公开Mack实现给出准备金及预测误差。","notCovered":"不证明AI会选对方法、尾部假设或真实准备金充分性。"},{"sourceId":"WU-FIN-LIFELIB","supports":"寿险现金流、定价和模型测试的公开实现入口。","notCovered":"README用途声明与代码入口，不是AI实测或任何法域监管适用性结论。"},{"sourceId":"WU-FIN-INS-ACTBENCH","supports":"精算已有直接LLM表格/R代码和案例评估；答案格均分与真实整包保险模型可靠性不同。","notCovered":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。"}],"resourceRefs":[],"career":{"direction":"保险负债估计：赔款准备金或寿险现金流可复算交付","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"产险固定一组三角独立复算IBNR并注入累计增量错标；寿险固定一组保单核对逐期现金流、折现与退保压力；两者都解释上期到本期变化。"},"team":{"bottleneck":"数值计算通过后，尾部方法和长期假设仍可能失配，需独立压力和历期回看。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"产险固定一组三角独立复算IBNR并注入累计增量错标；寿险固定一组保单核对逐期现金流、折现与退保压力；两者都解释上期到本期变化。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F20-method","statement":"公开Mack实现给出准备金及预测误差。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-CHAIN-MACK","locator":"MackChainladder、ukmotor示例、ibnr_及预测误差","relation":"method","supports":"公开Mack实现给出准备金及预测误差。","limit":"不证明AI会选对方法、尾部假设或真实准备金充分性。"}],"progressIds":[],"configuration":null,"limitation":"不证明AI会选对方法、尾部假设或真实准备金充分性。","evidenceStatus":"active"},{"id":"W-F20-adoption","statement":"可接收固定估计日的可复算负债候选；Mack只覆盖产险三角，寿险用逐期模型。INS-ActBench展示专业实作，未展示本机构尾部或真实准备金充分。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-CHAIN-MACK","locator":"MackChainladder、ukmotor示例、ibnr_及预测误差","relation":"adjacent","supports":"产险赔款发展三角和准备金方法可检查。","limit":"不把方法存在当作AI已完成本任务。"},{"sourceId":"WU-FIN-INS-ACTBENCH","locator":"§3.2—3.3、§4.1 Evaluation、Table3、§4.4","relation":"partial","supports":"专业精算材料改编的直接LLM任务评估。","limit":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。"},{"sourceId":"WU-FIN-LIFELIB","locator":"What is lifelib / What for；basiclife模型库","relation":"partial","supports":"寿险现金流、定价和模型测试的公开实现入口。","limit":"README用途声明与代码入口，不是AI实测或任何法域监管适用性结论。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"},{"id":"W-F20-lifelib","statement":"lifelib提供寿险逐期现金流模型入口，支持长期负债候选的局部方法。","basis":"inspectable-artifact","sourceLinks":[{"sourceId":"WU-FIN-LIFELIB","locator":"What is lifelib / What for；basiclife模型库","relation":"partial","supports":"lifelib提供寿险逐期现金流模型入口，支持长期负债候选的局部方法。","limit":"README用途声明与代码入口，不是AI实测或任何法域监管适用性结论。"}],"progressIds":[],"configuration":null,"limitation":"README用途声明与代码入口，不是AI实测或任何法域监管适用性结论。","evidenceStatus":"active"},{"id":"W-F20-ins-actbench","statement":"INS-ActBench Table3中GPT-5.5表格均分67.7%、R输出评分71.1%；表格按答案格平均，不是整表或完整精算模型通过率。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-INS-ACTBENCH","locator":"§3.2—3.3、§4.1 Evaluation、Table3、§4.4","relation":"partial","supports":"精算已有直接LLM表格/R代码和案例评估；答案格均分与真实整包保险模型可靠性不同。","limit":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。"}],"progressIds":["P-FIN-ACT-PRACTICE"],"configuration":"9模型；Practice零样本，表格636题/R516题，执行输出与0.2%相对误差；表格按答案格平均，案例改编多选。","limitation":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"已核对公开代码/文档入口，未安装或重放。Agent模型和数据接口需另接；账户、API和计算费用未测。","tools":[{"name":"chainladder固定Mack准备金实现","url":"https://github.com/casact/chainladder-python/blob/a090275cfc74deb93ece455cc576f5d20235adcf/chainladder/methods/mack.py#L8-L162","version":"固定commit见URL","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"}]},"variants":[{"title":"产险赔款准备金","inputDifference":"赔款发展三角、尾部和通胀","assessmentChange":"Mack提供可复算候选与预测误差，不负责方法适合度","acceptanceChange":"累计增量对账、独立复算、尾部及历期回看"},{"title":"寿险长期负债","inputDifference":"逐期保单生存、退保、费用和折现","assessmentChange":"改用逐期现金流，不迁移Mack结果","acceptanceChange":"逐期给付与费用、概率、折现及假设压力"}],"evidenceStatus":"active"},{"id":"W-F21","domainId":"finance","title":"客户目标到财富计划：情景、费用与持续调整","goal":"交付与客户目标约束一致的候选财务计划、实施事项和调整依据。","trigger":"客户需要退休、教育、保障或资产配置计划更新。","scope":"规划计算、产品适合度、实施与长期目标效果分开；不提供本网站用户个人买卖建议。","inputs":[{"name":"客户事实与目标","detail":"资产负债、收入支出、期限、流动性和损失承受"},{"name":"产品及规划假设","detail":"现金流、费用、风险、税务版本和实施授权"}],"deliverables":["事实缺口与目标冲突","退休等现金流情景和保障需求","产品/配置费用风险比较","实施清单、沟通解释及授权状态","再平衡和目标达成监测"],"acceptance":[{"criterion":"计划不偷换目标","method":"核对期限、流动性、负债和可承受损失，展示冲突","failure":"靠增加风险弥补必需现金"},{"criterion":"现金流和费用可复算","method":"独立算通胀、费用、税情景和不利路径","failure":"只展示均值收益"},{"criterion":"解释与持续服务有效","method":"客户能说明关键取舍，变化后更新计划并跟踪实施状态","failure":"生成报告即称客户目标实现"}],"assessment":{"recommendation":"可用于产生和挑战候选；交付与客户目标约束一致的候选财务计划、实施事项和调整依据。 接收依据见下方实质检查，不以流畅解释替代结果。","conditions":["规划计算、产品适合度、实施与长期目标效果分开；不提供本网站用户个人买卖建议。","需可访问、版本固定的资料和能独立重放的计算/状态记录"],"evidenceBasis":"FinGAIA公开例只为基金排名及持仓检索提供局部参照；实际任务见下方具体主张。财富配置与客户目标的整包采用另依公开方法、明确条件和独立验收提出，不从角色名推出专业能力。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。","尚需本任务授权样本和错误注入的独立接收记录"],"mode":"candidate-exploration","rationale":"对给定目标形成可复算财富计划是方法驱动的条件建议。基金排序/持仓检索只支持资料层；客户约束、适当性、配置取舍与持续调整不能从角色名称继承。","frontier":"公开Agent评估已包含基金排序和持仓查找，但这个配置角色例不是投资组合优化。可复算费用、情景和现金流方法可组织财富计划候选，客户目标和执行效果仍须单独验证。","nextValidation":"用脱敏家庭资料植入流动性需求、费用和收入中断，核对方案冲突、负面情景及客户理解；随后跟踪实际实施。","futureSignals":["新客户约束、费用和收入中断都能被正确处理且客户理解关键取舍，再扩大规划与实施","计划用高期望收益掩盖必需流动性、产品条款过期或客户误解时，退回方案修订"]},"workflow":{"before":"客户资料、目标、产品表、现金流计算和持续回顾各做一次。","now":"Agent在同一目标约束下搜索计划和不利情景，用计算反馈暴露期限费用冲突，再生成可解释方案和跟踪事项；客户目标或条款变化时回修，不把高期望收益当默认最优。","retained":["计划不偷换目标","现金流和费用可复算","实际委托的授权边界及最终使用责任"],"added":["保留输入版本、实际工具配置、计算/操作记录与未解决项","区分模型/候选正确性和现实决策效果"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现靠增加风险弥补必需现金等实质错误应拒收相应部分；无法满足假设、取证或状态条件时交付明确未决及所需资料，不补造成功。","example":{"input":"两年内购房资金却被长期锁定。","output":"识别目标冲突，列可用流动性与备选情景。","failure":"只比较预期收益而忽略不能赎回。"},"sourceIds":["WU-FIN-FINROBOT-DCF","WU-FIN-RETIREMENT-REFERENCE","WU-FIN-FINGAIA"],"evidenceLinks":[{"sourceId":"WU-FIN-FINROBOT-DCF","supports":"确定性金融计算可作为局部方法底座。","notCovered":"仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。"},{"sourceId":"WU-FIN-RETIREMENT-REFERENCE","supports":"退休情景模拟有可检查历史实现。","notCovered":"仓库明确只是旧参考，当前同名服务闭源；未复放，不将旧代码与当前服务成绩互换。"},{"sourceId":"WU-FIN-FINGAIA","supports":"Appendix A Fig9在“Portfolio Fund Allocation”名下的公开例实际数图中动物种类，再据指定数量按回撤排序基金并查持仓代码。只支持基金排名及持仓检索子能力。","notCovered":"客户适当性、资产配置优化或长期财富计划及本业务端到端可靠性。"}],"resourceRefs":[],"career":{"direction":"客户目标到财富计划：情景、费用与持续调整","evidence":"提交脱敏输入、完整产物和一项实质失败的发现/修订记录。","nextStep":"用脱敏家庭资料植入流动性需求、费用和收入中断，核对方案冲突、负面情景及客户理解；随后跟踪实际实施。"},"team":{"bottleneck":"先区分资料、计算和现实状态/假设哪一项造成返工。","allocation":"Agent围绕交付执行并保存证据；接收者按关键假设、独立计算和实际状态验收。","nextStep":"用脱敏家庭资料植入流动性需求、费用和收入中断，核对方案冲突、负面情景及客户理解；随后跟踪实际实施。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F21-method","statement":"确定性金融计算可作为局部方法底座。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-FINROBOT-DCF","locator":"calculate_dcf；terminal growth/WACC与terminal FCF边界","relation":"method","supports":"确定性金融计算可作为局部方法底座。","limit":"仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。"}],"progressIds":[],"configuration":null,"limitation":"仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。","evidenceStatus":"active"},{"id":"W-F21-adoption","statement":"客户目标和约束明确时，可委派现金流、费用及配置情景候选；适当性、权衡和持续执行效果须独立接收。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-RETIREMENT-REFERENCE","locator":"README历史参考声明；同commit js/cFIREsimOpen.js","relation":"adjacent","supports":"历史退休计算工具提供局部方法，结合固定目标和独立现金流验收形成采用假设。","limit":"仓库明确只是旧参考，当前同名服务闭源；未复放，不将旧代码与当前服务成绩互换。"},{"sourceId":"WU-FIN-FINGAIA","locator":"Table2、Agents/Evaluation Methods、Table3及AppendixB","relation":"adjacent","supports":"基金排名及持仓检索可作为交付中取材/计算环节的参照。","limit":"完整采用建议为编辑推断；Appendix A Fig9在“Portfolio Fund Allocation”名下的公开例实际数图中动物种类，再据指定数量按回撤排序基金并查持仓代码。它不验证客户适当性、资产配置优化或长期财富计划。"}],"progressIds":[],"configuration":null,"limitation":"未有本站试次；本任务整包可靠性、重试及人工成本未知，最小验证见本页的下一项验证。","evidenceStatus":"active"},{"id":"W-F21-wu-fin-retirement-reference","statement":"退休情景模拟有可检查历史实现。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-RETIREMENT-REFERENCE","locator":"README历史参考声明；同commit js/cFIREsimOpen.js","relation":"method","supports":"退休情景模拟有可检查历史实现。","limit":"仓库明确只是旧参考，当前同名服务闭源；未复放，不将旧代码与当前服务成绩互换。"}],"progressIds":[],"configuration":null,"limitation":"仓库明确只是旧参考，当前同名服务闭源；未复放，不将旧代码与当前服务成绩互换。","evidenceStatus":"active"},{"id":"W-F21-fingaia","statement":"Appendix A Fig9在“Portfolio Fund Allocation”名下的公开例实际数图中动物种类，再据指定数量按回撤排序基金并查持仓代码。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-FINGAIA","locator":"Appendix A Fig9在“Portfolio Fund Allocation”名下的","relation":"adjacent","supports":"基金排名及持仓检索的局部评估参照。","limit":"角色标签不等于实际专业任务；公开例不能证明客户适当性、资产配置优化或长期财富计划，不把分类分数当本任务成功率。"}],"progressIds":["P-FIN-SCENARIOS"],"configuration":"FinGAIA v1零样本网页/附件原环境；论文角色分类需对照Appendix A实际样例；正文与附表口径有差异，不使用该标签分数为本专业成功率。","limitation":"角色标签不等于实际专业任务；公开例不能证明客户适当性、资产配置优化或长期财富计划，不把分类分数当本任务成功率。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"已核对公开代码/文档入口，未安装或重放。Agent模型和数据接口需另接；账户、API和计算费用未测。","tools":[{"name":"FinRobot固定DCF计算与边界处理","url":"https://github.com/AI4Finance-Foundation/FinRobot/blob/2717499b8e30f242640af08c4ad9afd1113c2d45/finrobot_desktop/finrobot/engine/compute/operators/dcf.py#L48-L177","version":"固定commit见URL","access":"公开读取；运行需自行准备依赖、授权数据和Agent模型；规范文档本身不是执行工具"},{"name":"cFIREsim旧开源退休模拟参考","url":"https://github.com/boknows/cFIREsim-open/blob/c7c4ba441830c1da457b663d2e602567306e8459/README.md","version":"固定commit见URL","access":"公开参考；本任务未安装运行；需另接模型/内部数据，实际费用未知。"}]},"variants":[{"title":"现金流目标规划","inputDifference":"期限和事实明确","assessmentChange":"可尝试可复算候选计划","acceptanceChange":"现金流、费用和不利情景"},{"title":"具体产品配置","inputDifference":"适合度与实施权限增加","assessmentChange":"补产品法域和客户证据再扩围","acceptanceChange":"风险理解、费用、流动性和实施状态"}],"evidenceStatus":"active"},{"id":"W-F22","domainId":"finance","title":"承保决定到保单状态：绑定、拒绝与再保转介","goal":"在固定准则、条件和权限下完成可追踪承保处理，使决定、保单或再保状态与证据一致。","trigger":"核保资料与条件已完成评估，需接受、拒绝、条件承保或转移风险。","scope":"保留实际承保决定、报价绑定/出单、拒绝、再保安排和例外处置；先在沙盒验证状态操作。公开方法不证明Agent已会正确做全部决定。","inputs":[{"name":"决定依据与条件","detail":"申请、准则版本、已核风险因素、费率限额免赔、需要的再保结构及反证"},{"name":"权限与业务状态","detail":"账户、操作角色、批准额度、阻断事项、保单/再保系统版本和可访问接口；真实操作依委托授权"}],"deliverables":["接受、拒绝或条件承保的理由和条件快照","报价到绑定/出单、拒绝或待批准的实际状态及版本","再保条款、分入方确认/批准和净自留风险记录","失败、重试、状态冲突与需转介的问题闭环"],"acceptance":[{"criterion":"决定符合固定目标而非只操作成功","method":"从申请风险、准则及反证重做接受/拒绝/条件判断；保险范围与条件不能在系统操作中被改写","failure":"为完成绑定删除不利条件或把缺失资料默认合格"},{"criterion":"授权与状态都正确","method":"独立只读查询绑定/拒绝状态、条件、保额和主体；测试未批准、过期报价、并发变更与重复调用","failure":"返回200或已有报价即称承保已生效"},{"criterion":"再保不是一句转介","method":"核对自留/分出限额、条约或临分适用、确认条款及批准状态；需要外部确认的保留未生效","failure":"发出再保请求就称风险已转移"},{"criterion":"失败保持一致且可恢复","method":"故意断连、响应丢失或条件变更，先查询既有状态再恢复；核对无重复保单和错误拒保","failure":"不知前次结果重复绑定，或越权绕过阻断"}],"assessment":{"recommendation":"可分阶段委派这一完整结果；先交可重放沙盒包，验证状态与判断后按已有权限扩围。","conditions":["业务口径、权限和版本明确","可访问原始证据及系统只读核对，必要操作授权已具备"],"evidenceBasis":"Guidewire产品声明说明保单变更Agent受Early Access限制；业务API只证明状态操作方法，最终核保判断和再保执行为待检迁移。","localStatus":"未本站运行；真实接口、端到端成功率、重试、API成本、等待与人工时间未知。","gaps":["尚需授权实例验证数据、权限、幂等、状态和失败恢复","再保真实确认与复杂承保判断尚缺直接证据"],"mode":"staged-delivery","rationale":"职业结果可以包括授权内真实承保处理。现有Agent产品与业务API支持先做沙盒的决定—状态闭环候选；准则判断、再保确认和真实系统失败恢复尚需独立样本，不因签字制度推永久人类独占。","frontier":"已有专业工作流框架和受限保单变更助手，可检查绑定API；没有公开整包接受/拒绝风险、再保确认与失败恢复的AI成绩，不能等同全自动核保。","nextValidation":"沙盒准备标准接受、拒绝、超权、并发报价和再保未确认五类案件，独立核对理由、条件及最终状态；记录重试与介入。","futureSignals":["在独立标准/例外申请中决定、条件、权限和保单/再保状态一致，重复与并发可恢复后按授权扩围","AI为成功绑定删除条件、错读待批准状态或再保未确认便生效时，阻断实际承保动作"]},"workflow":{"before":"候选判断、专业估计、业务记账和实际状态由多次交接分开，后续难核哪一步真的完成。","now":"Agent围绕承保目标核对申请、准则、报价和反证，按权限提交绑定/拒绝或等待批准；再保需独立确认条款和风险转移，状态读回驱动修订，不能因API可调用跳过未满足条件。","retained":["固定保障/估计口径、未决事实与授权责任","原始证据、状态与专业判断分别接收"],"added":["版本化决策和状态差异","断连重试、重复事件、待批准及未付状态的独立检查"]},"humanResponsibility":"确定委托权限、产品规则和责任归属，保证实际操作符合授权；可由AI承担哪种判断或行动应由任务验证决定，不从责任制度推断永久技术上限。","failureHandling":"准则/估计存在未决、权限不足或状态不明时保留明确待处理；禁止删除条件、伪造批准、把请求成功当业务完成；响应丢失先只读确认状态再决定恢复。","example":{"input":"商业险申请条件已核，报价尚有一项未批准的再保条件。","output":"交付依据一致的承保处理和再保状态；批准不足时保持待处理，批准齐全后按授权执行并核对保单条件。","failure":"以系统能bind为由提前让保单生效，未确认风险转移。"},"sourceIds":["WU-FIN-GW-AGENT","WU-FIN-GW-BIND"],"evidenceLinks":[{"sourceId":"WU-FIN-GW-AGENT","supports":"厂商提供接入保单/理赔/账单工作流的框架和专门Agent。","notCovered":"Policy Change为Early Access，Claim Summary/FNOL为Restricted Availability；不是公众无条件可用、完整核保决定或真实承保AI成绩。"},{"sourceId":"WU-FIN-GW-BIND","supports":"产品业务API为绑定与出单的状态操作提供接口。","notCovered":"方法文档，不证明Agent已选对承保条件；需商业实例、权限与实际产品配置。"}],"resourceRefs":[],"career":{"direction":"核保业务处理","evidence":"提供脱敏输入、专业判断、两轮状态变化及失败恢复证据。","nextStep":"沙盒准备标准接受、拒绝、超权、并发报价和再保未确认五类案件，独立核对理由、条件及最终状态；记录重试与介入。"},"team":{"bottleneck":"判断/估计、授权、数据还是接口状态阻塞处理？","allocation":"按真实权限分配判断与操作，独立接收者检查状态与金额，不预设永久人机比例。","nextStep":"沙盒准备标准接受、拒绝、超权、并发报价和再保未确认五类案件，独立核对理由、条件及最终状态；记录重试与介入。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F22-method","statement":"公开业务API可将绑定/出单作为可读回核对的状态，而不只生成建议。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-GW-BIND","locator":"POST bind-and-issue及前置引用；Palisades2026.03","relation":"method","supports":"公开业务API可将绑定/出单作为可读回核对的状态，而不只生成建议。","limit":"方法文档，不证明Agent已选对承保条件；需商业实例、权限与实际产品配置。"}],"progressIds":[],"configuration":null,"limitation":"见来源限定；不将公开方法或厂商自报当成本站整包能力验证。","evidenceStatus":"active"},{"id":"W-F22-vendor","statement":"Guidewire发布保单变更Agent及核心流程接入，但该功能Early Access，未公布自动核保决定可靠性。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-GW-AGENT","locator":"Policy Change for ProNavigator及*、** availability footnotes","relation":"partial","supports":"Guidewire发布保单变更Agent及核心流程接入，但该功能Early Access，未公布自动核保决定可靠性。","limit":"Policy Change为Early Access，Claim Summary/FNOL为Restricted Availability；不是公众无条件可用、完整核保决定或真实承保AI成绩。"}],"progressIds":["P-FIN-INSURANCE-AGENT"],"configuration":"厂商产品/部署自报；具体模型、预算和错误分母未公开；不是本站或独立复现。","limitation":"见来源限定；不将公开方法或厂商自报当成本站整包能力验证。","evidenceStatus":"active"},{"id":"W-F22-adoption","statement":"职业结果可以包括授权内真实承保处理。现有Agent产品与业务API支持先做沙盒的决定—状态闭环候选；准则判断、再保确认和真实系统失败恢复尚需独立样本，不因签字制度推永久人类独占。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-GW-BIND","locator":"POST bind-and-issue及前置引用；Palisades2026.03","relation":"partial","supports":"职业结果可以包括授权内真实承保处理。现有Agent产品与业务API支持先做沙盒的决定—状态闭环候选；准则判断、再保确认和真实系统失败恢复尚需独立样本，不因签字制度推永久人类独占。","limit":"方法文档，不证明Agent已选对承保条件；需商业实例、权限与实际产品配置。"},{"sourceId":"WU-FIN-GW-AGENT","locator":"Policy Change for ProNavigator及*、** availability footnotes","relation":"partial","supports":"厂商提供接入保单/理赔/账单工作流的框架和专门Agent。","limit":"Policy Change为Early Access，Claim Summary/FNOL为Restricted Availability；不是公众无条件可用、完整核保决定或真实承保AI成绩。"}],"progressIds":[],"configuration":null,"limitation":"见来源限定；不将公开方法或厂商自报当成本站整包能力验证。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开API方法和产品/部署说明可检查；商业业务实例与模型配置未本站获得。Qusar专门Agent存在Early/Restricted Access限制；方法文档不是模型服务。","tools":[{"name":"Guidewire Qusar Agent发布与准入范围","url":"https://www.guidewire.com/en/about/press-center/press-releases/20260803/guidewire-introduces-qusar-release-to-help-insurers-build-and-control-ai-agents","version":"来源所列产品/发布日期","access":"公开说明可读，真实系统需商业实例、权限和数据；未取得本站运行环境，费用未知。"},{"name":"PolicyCenter 2026.03承保绑定状态方法","url":"https://docs.guidewire.com/cloud/is/202603/cloudapibf/cloudAPI/PolicyCenter/job-types/submissions/c_binding_and_issuing_a_submission_at_the_same_time.html","version":"2026.03业务方法","access":"公开说明可读，真实系统需商业实例、权限和数据；未取得本站运行环境，费用未知。"}]},"variants":[{"title":"规则内接受或拒绝","inputDifference":"准则、资料、报价及执行权限完整","assessmentChange":"先校准决定与状态双重验收，稳定后按授权扩大实际处理","acceptanceChange":"对照业务条件及只读状态，不以HTTP成功当业务成功"},{"title":"再保或例外承保","inputDifference":"限额超额、外部再保确认、非标条件或需批准","assessmentChange":"保留完整条件与转介/确认闭环，不编造已批准","acceptanceChange":"复核风险转移、自留和外部确认，未满足条件阻断生效"}],"evidenceStatus":"active"},{"id":"W-F23","domainId":"finance","title":"赔案持续处理：准备金、和解付款与结案状态","goal":"随新证据持续更新赔案估计和准备金账务，在授权内完成和解付款及可核查结案，未决事项不被隐藏。","trigger":"损失新信息、估计修订、和解达成或付款状态变化，需要更新赔案及账务。","scope":"这是赔案生命周期任务，区别于一次理算金额和群体精算IBNR。证据取得、估计合理性、准备金批准/记账、支付指令、实际付讫与结案分别接收。","inputs":[{"name":"赔案与估计依据","detail":"原始合同事实、保障和损失证据、估计总损失/费用、已付、待付、追偿和原准备金版本"},{"name":"持续业务记录与权限","detail":"新证据、准备金线/币种、和解条件、受款人、权限额度、批准与支付系统回报；已给机构口径和规则"}],"deliverables":["旧新估计及证据、未决事实和改变原因","按准备金线/币种区分建议、待批准、已记账的变动及差异","授权和解条件、实际达成状态与付款对象/指令","付款批准、已提交、失败、付讫等真实回报和对账","结案条件检查、关闭或保持未决状态及重开记录"],"acceptance":[{"criterion":"持续估计与准备金账务区分","method":"独立按新旧证据复做估计；在给定口径下计算调整量，核对已付/待付/未决和准备金线，待批准不记作已记账","failure":"把新的总估计直接当追加额，或用一次理算候选代持续准备金"},{"criterion":"和解与付款真实且不越权","method":"核对和解证据、授权、受款人、额度、币种及支付验证状态；重复/失败回报先查后重试","failure":"发出和解建议当已达成，支付请求创建即称实际已付"},{"criterion":"关闭依赖结果而非材料齐全","method":"按实例规则核查exposure、未决支付/争议/追偿和准备金处置，独立读回closed状态；支付失败或新证据必要时重开","failure":"报告已写完便结案，掩盖银行拒付与未解决责任"},{"criterion":"循环更新不丢历史","method":"回放至少两轮新证据和一次已关闭后更正，保持估计版本、权限和状态差异，实际工时重试分别记","failure":"覆盖旧估计，使后续无法解释准备金变化"}],"assessment":{"recommendation":"可分阶段委派这一完整结果；先交可重放沙盒包，验证状态与判断后按已有权限扩围。","conditions":["业务口径、权限和版本明确","可访问原始证据及系统只读核对，必要操作授权已具备"],"evidenceBasis":"Allianz部署说明是低复杂Agent流程的厂商证据；ClaimCenter方法区分估计、批准、准备金、支付对象及结案，不证明所有环节AI可靠。","localStatus":"未本站运行；真实接口、端到端成功率、重试、API成本、等待与人工时间未知。","gaps":["尚需授权实例验证数据、权限、幂等、状态和失败恢复","个案持续估计准确性、长期结果和跨险种迁移尚缺直接证据"],"mode":"staged-delivery","rationale":"实际理赔已有Agent嵌入低复杂业务的采用材料，公开API可独立核对持续准备金和支付/结案状态。可先委派沙盒的估计—状态闭环；真实自动支付范围要按具体产品、权限和失败验证决定，不能从厂商案例外推。","frontier":"低额食品变质理赔已有Agent嵌入，案例仍保留最终人工支付决定；德国宠物自动化属另一系统。公开准备金/支付/关闭接口支持研究真实生命周期，但未证连续估计和跨险种自动付款。","nextValidation":"用两轮损失估计、部分已付、超权准备金、支付失败和新证据重开组成脱敏沙盒，独立重算估计/调整并核对批准、支付及关闭状态。","futureSignals":["两轮新证据中估计及调整准确、授权状态与回执一致，支付失败/新争议可重开后，扩大限定生命周期","新的总估计被当追加额、未批准被记账、请求当付讫或带未决结案时，拒收并保留赔案状态"]},"workflow":{"before":"候选判断、专业估计、业务记账和实际状态由多次交接分开，后续难核哪一步真的完成。","now":"Agent随新损失证据修订赔案估计和准备金调整，调用受权业务系统核对批准/记账；和解证据、支付回报与未决exposure决定关闭或重开，稳定金额用程序复算，不把一次理算当持续维护。","retained":["固定保障/估计口径、未决事实与授权责任","原始证据、状态与专业判断分别接收"],"added":["版本化决策和状态差异","断连重试、重复事件、待批准及未付状态的独立检查"]},"humanResponsibility":"确定委托权限、产品规则和责任归属，保证实际操作符合授权；可由AI承担哪种判断或行动应由任务验证决定，不从责任制度推断永久技术上限。","failureHandling":"准则/估计存在未决、权限不足或状态不明时保留明确待处理；禁止删除条件、伪造批准、把请求成功当业务完成；响应丢失先只读确认状态再决定恢复。","example":{"input":"机构定义未决准备金为估计总损失减已付；旧总估计100、已付40、未决60，新证据使总估计130。","output":"未决候选90、调整候选+30；待批准与已记账分开。付款30若失败仍保留待处理，满足真实条件后才更新结案。","failure":"追加130或把支付对象创建当付讫，把赔案直接关闭。"},"sourceIds":["WU-FIN-GW-RESERVE","WU-FIN-GW-PAYMENT","WU-FIN-GW-CLOSE","WU-FIN-ALLIANZ-NEMO","WU-FIN-ALLIANZ-2026"],"evidenceLinks":[{"sourceId":"WU-FIN-GW-RESERVE","supports":"准备金交易、批准及变动而非仅计算报告的公开方法。","notCovered":"估计合理性及持续维护需专业证据；未批准金额不进入准备金线，方法本身非AI试验。"},{"sourceId":"WU-FIN-GW-PAYMENT","supports":"付款请求的前置验证与状态对象可检查。","notCovered":"创建支票/支付对象不等于银行实际付讫，需外部回报及实例规则；不是AI成功证据。"},{"sourceId":"WU-FIN-GW-CLOSE","supports":"状态关闭受业务校验，存在未关闭exposure等阻断。","notCovered":"基础配置示例，不能替代实际机构结案标准或债务履行检查。"},{"sourceId":"WU-FIN-ALLIANZ-NEMO","supports":"自报Agent把覆盖天气反欺诈及金额形成可交最终支付决策的处理包。","notCovered":"自报部署，保险公司保留最终支付决定；未公开模型、预算、逐案正确率或可重放产物，不扩大到人身或争议案件。"},{"sourceId":"WU-FIN-ALLIANZ-2026","supports":"厂商分别描述自动宠物理赔与Agent食品理赔，不宜统称同系统。","notCovered":"自动宠物处理数字不证明LLM Agent；国家险种、拒赔升级与人工环节不同，无独立产物重放。"}],"resourceRefs":[],"career":{"direction":"赔案估计、支付与结案","evidence":"提供脱敏输入、专业判断、两轮状态变化及失败恢复证据。","nextStep":"用两轮损失估计、部分已付、超权准备金、支付失败和新证据重开组成脱敏沙盒，独立重算估计/调整并核对批准、支付及关闭状态。"},"team":{"bottleneck":"判断/估计、授权、数据还是接口状态阻塞处理？","allocation":"按真实权限分配判断与操作，独立接收者检查状态与金额，不预设永久人机比例。","nextStep":"用两轮损失估计、部分已付、超权准备金、支付失败和新证据重开组成脱敏沙盒，独立重算估计/调整并核对批准、支付及关闭状态。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F23-method","statement":"ClaimCenter文档明确准备金批准、付款对象和结案阻断是不同业务状态。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-GW-RESERVE","locator":"Approval status、Example of adding to reserves与币种","relation":"method","supports":"ClaimCenter文档明确准备金批准、付款对象和结案阻断是不同业务状态。","limit":"估计合理性及持续维护需专业证据；未批准金额不进入准备金线，方法本身非AI试验。"},{"sourceId":"WU-FIN-GW-PAYMENT","locator":"Validation levels、payment types与POST check-sets","relation":"method","supports":"付款对象、AbilityToPay及check-set创建状态；创建不等于银行已付。","limit":"创建支票/支付对象不等于银行实际付讫，需外部回报及实例规则；不是AI成功证据。"},{"sourceId":"WU-FIN-GW-CLOSE","locator":"Closing claims / Reopening a closed claim","relation":"method","supports":"关闭、重开及未关闭exposure阻断状态。","limit":"基础配置示例，不能替代实际机构结案标准或债务履行检查。"}],"progressIds":[],"configuration":null,"limitation":"见来源限定；不将公开方法或厂商自报当成本站整包能力验证。","evidenceStatus":"active"},{"id":"W-F23-vendor","statement":"Allianz自报澳洲低额食品变质Agent处理到最终支付决策前，仍有人工决定；德国宠物自动化是另一系统，不等于同一Agent自动付款。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-ALLIANZ-NEMO","locator":"Nemo works like this / Human-in-the-loop；澳洲低额食品变质险","relation":"partial","supports":"Allianz自报澳洲低额食品变质Agent处理到最终支付决策前，仍有人工决定；德国宠物自动化是另一系统，不等于同一Agent自动付款。","limit":"自报部署，保险公司保留最终支付决定；未公开模型、预算、逐案正确率或可重放产物，不扩大到人身或争议案件。"},{"sourceId":"WU-FIN-ALLIANZ-2026","locator":"Responsible AI in action：German pet insurance / Australia food spoilage","relation":"partial","supports":"后续Allianz材料描述德国宠物另一系统，不能与澳洲Nemo合并推自动付款。","limit":"自动宠物处理数字不证明LLM Agent；国家险种、拒赔升级与人工环节不同，无独立产物重放。"}],"progressIds":["P-FIN-CLAIMS-LIFECYCLE"],"configuration":"厂商产品/部署自报；具体模型、预算和错误分母未公开；不是本站或独立复现。","limitation":"见来源限定；不将公开方法或厂商自报当成本站整包能力验证。","evidenceStatus":"active"},{"id":"W-F23-adoption","statement":"实际理赔已有Agent嵌入低复杂业务的采用材料，公开API可独立核对持续准备金和支付/结案状态。可先委派沙盒的估计—状态闭环；真实自动支付范围要按具体产品、权限和失败验证决定，不能从厂商案例外推。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-GW-RESERVE","locator":"Approval status、Example of adding to reserves与币种","relation":"partial","supports":"实际理赔已有Agent嵌入低复杂业务的采用材料，公开API可独立核对持续准备金和支付/结案状态。可先委派沙盒的估计—状态闭环；真实自动支付范围要按具体产品、权限和失败验证决定，不能从厂商案例外推。","limit":"估计合理性及持续维护需专业证据；未批准金额不进入准备金线，方法本身非AI试验。"},{"sourceId":"WU-FIN-ALLIANZ-NEMO","locator":"Nemo works like this / Human-in-the-loop；澳洲低额食品变质险","relation":"partial","supports":"自报Agent把覆盖天气反欺诈及金额形成可交最终支付决策的处理包。","limit":"自报部署，保险公司保留最终支付决定；未公开模型、预算、逐案正确率或可重放产物，不扩大到人身或争议案件。"},{"sourceId":"WU-FIN-GW-PAYMENT","locator":"Validation levels、payment types与POST check-sets","relation":"method","supports":"付款对象、AbilityToPay及check-set创建状态；创建不等于银行已付。","limit":"创建支票/支付对象不等于银行实际付讫，需外部回报及实例规则；不是AI成功证据。"},{"sourceId":"WU-FIN-GW-CLOSE","locator":"Closing claims / Reopening a closed claim","relation":"method","supports":"关闭、重开及未关闭exposure阻断状态。","limit":"基础配置示例，不能替代实际机构结案标准或债务履行检查。"},{"sourceId":"WU-FIN-ALLIANZ-2026","locator":"Responsible AI in action：German pet insurance / Australia food spoilage","relation":"partial","supports":"后续Allianz材料描述德国宠物另一系统，不能与澳洲Nemo合并推自动付款。","limit":"自动宠物处理数字不证明LLM Agent；国家险种、拒赔升级与人工环节不同，无独立产物重放。"}],"progressIds":[],"configuration":null,"limitation":"见来源限定；不将公开方法或厂商自报当成本站整包能力验证。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"商业部署与业务接口已公开说明，未公开可复制的完整生产系统；需要保险机构实例、授权与实际支付回报。","tools":[{"name":"ClaimCenter2026.03准备金创建、追加和批准","url":"https://docs.guidewire.com/cloud/is/202603/cloudapibf/cloudAPI/ClaimCenter/financials/reserves/c_creating-reserves.html","version":"2026.03业务方法","access":"公开说明可读，真实系统需商业实例、权限和数据；未取得本站运行环境，费用未知。"},{"name":"ClaimCenter2026.03付款状态创建方法","url":"https://docs.guidewire.com/cloud/is/202603/cloudapibf/cloudAPI/ClaimCenter/financials/check-creating/c_POSTing-checks.html","version":"2026.03业务方法","access":"公开说明可读，真实系统需商业实例、权限和数据；未取得本站运行环境，费用未知。"},{"name":"ClaimCenter2026.03关闭与重开方法","url":"https://docs.guidewire.com/cloud/is/202603/cloudapibf/cloudAPI/ClaimCenter/fnol/claims/c_closing_claims.html","version":"2026.03业务方法","access":"公开说明可读，真实系统需商业实例、权限和数据；未取得本站运行环境，费用未知。"},{"name":"Allianz实际Agent理赔采用说明","url":"https://www.allianz.com/en/mediacenter/news/articles/251103-when-the-storm-clears-so-should-the-claim-queue.html","version":"来源所列产品/发布日期","access":"公开说明可读，真实系统需商业实例、权限和数据；未取得本站运行环境，费用未知。"},{"name":"Allianz 2026采用回顾：区分自动化与Agent","url":"https://www.allianz.com/en/mediacenter/news/articles/260318-responsible-use-of-ai-at-allianz.html","version":"来源所列产品/发布日期","access":"公开说明可读，真实系统需商业实例、权限和数据；未取得本站运行环境，费用未知。"}]},"variants":[{"title":"准备金持续维护","inputDifference":"至少两轮新证据、已付及原准备金，给定估计和账务口径","assessmentChange":"可先交版本化估计与调整，并在验证环境跟踪批准/记账；估计适合度单独验证","acceptanceChange":"核对调整量、准备金线和币种，未批准金额不列已记账"},{"title":"授权和解支付与结案","inputDifference":"和解证据、批准、支付接口与外部回报完整","assessmentChange":"按授权验证真实操作范围，回报未知就保持未决；不默认永久只能起草","acceptanceChange":"查受款人及支付验证、实际状态、未决exposure/争议，失败阻断结案"},{"title":"低额标准赔案的生产案例迁移","inputDifference":"澳洲食品变质等规则明确、低复杂场景；不是寿险或重伤争议","assessmentChange":"厂商自报支持研究整包候选和实际嵌入路径，其他险种重新检验","acceptanceChange":"保持厂商案例的人介入及险种限制，不能以处理速度替代正确率"}],"evidenceStatus":"active"},{"id":"W-F24","domainId":"finance","title":"保险经验与假设包：暴露、变量、可信度和实际预期回顾","goal":"交付可追溯的经验数据集、假设建议及冻结预测后的实际预期差。","trigger":"续保或精算模型需要更新经验假设，或新一轮实际经验到来。","scope":"经验分析是下游定价和负债模型的输入；赔案文本变量提取只覆盖局部，不能把抽取置信度作为保险事实。","inputs":[{"name":"观察与暴露","detail":"逐期保单、赔案或生存经验、观察窗、删失及口径"},{"name":"既有假设与来源","detail":"冻结预测、文本资料、变量定义和可信度约定"}],"deliverables":["暴露与事件可对账的经验表和变量来源","分层经验、可信度及假设修订理由","冻结基线的实际预期差与未观察尾部"],"acceptance":[{"criterion":"观察口径正确","method":"核对期间、暴露、删失和样本选择，逐项对账","failure":"已结案样本冒充全部风险"},{"criterion":"变量有事实依据","method":"文本变量回链原句和时间；缺失保留未知，冲突分别记录","failure":"模型置信度替代事实"},{"criterion":"回顾不泄露未来","method":"冻结旧假设后比较新经验，按分层及可信度解释偏差","failure":"用修订后参数重写旧预测"}],"assessment":{"recommendation":"交付可追溯的经验数据集、假设建议及冻结预测后的实际预期差。 先按下列口径、数值和事实证据接收候选，再决定业务采用。","conditions":["经验分析是下游定价和负债模型的输入；赔案文本变量提取只覆盖局部，不能把抽取置信度作为保险事实。","输入版本与独立比较基准可用"],"evidenceBasis":"已有赔案文本到精算变量的直接LLM概念验证；经验可信度、删失和长期校准仍需本机构数据检验。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["经验分析是下游定价和负债模型的输入；赔案文本变量提取只覆盖局部，不能把抽取置信度作为保险事实。","本站未重放本任务，整包可靠性及人工成本未知"],"mode":"staged-delivery","rationale":"可把可回链变量与经验汇总交给Agent成包交付；文本原型展示的是提取和分层接口，不能替代样本代表性及长期假设检验。","frontier":"直接研究展示合成赔案文本变量提取与下游分层模型连接，尚未展示跨险种、真实多年经验的稳定假设维护。","nextValidation":"固定一批去标识赔案和保单暴露，注入缺失、时间错位与未结案，双盲核对变量后用冻结假设回看。","futureSignals":["真实未见赔案的变量与时间标签通过独立核对，才扩大文本输入","连续实际预期回顾中分层偏差持续存在，须修订假设而非只润色报告"]},"workflow":{"before":"经验表、文本和旧假设由不同团队整理，回顾时容易改变基线。","now":"Agent对齐观察窗、抽取带时间的候选变量，再用确定性汇总和实际预期回顾暴露差异；先接收经验与假设包，再决定是否进入定价或准备金。","retained":["观察口径正确","变量有事实依据","回顾不泄露未来"],"added":["保存冻结输入、工具配置、候选修订和未决状态","对照独立数值或原文接收，不以输出完整度代替正确性"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现已结案样本冒充全部风险时拒收受影响部分；保留原输入与失败，补齐资料或改模型后重新独立核对。","example":{"input":"同一年度有未结案和迟报赔案，旧假设已冻结。","output":"给出暴露一致经验表、未知变量和分层偏差；不补造最终损失。","failure":"只选已结案样本得到虚假的低损失率。"},"sourceIds":["WU-FIN-CLAIM-VARIABLES","WU-FIN-INS-ACTBENCH","WU-FIN-CLAIMS-IMPLEMENTATION"],"evidenceLinks":[{"sourceId":"WU-FIN-CLAIM-VARIABLES","supports":"合成赔案文本到结构化变量与分层分析的概念验证；两位临床评阅者按量表审阅核心变量。","notCovered":"合成数据和专家量表不是实赔准确率；分层准备金示范不证明真实准备金充分或寿险适用。"},{"sourceId":"WU-FIN-INS-ACTBENCH","supports":"专业精算材料改编的直接LLM任务评估。","notCovered":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。"},{"sourceId":"WU-FIN-CLAIMS-IMPLEMENTATION","supports":"作者公开了四阶段分析脚本、requirements与配置入口。","notCovered":"只读代码与文档；未验证模型调用、提取质量或真实赔案效果。"}],"resourceRefs":[],"career":{"direction":"保险经验与假设包：暴露、变量、可信度和实际预期回顾","evidence":"提交输入口径、可复算产物及一项实质失败的修正记录。","nextStep":"固定一批去标识赔案和保单暴露，注入缺失、时间错位与未结案，双盲核对变量后用冻结假设回看。"},"team":{"bottleneck":"用修订后参数重写旧预测","allocation":"Agent交付暴露与事件可对账的经验表和变量来源；独立接收者按数值与事实证据决定采用。","nextStep":"固定一批去标识赔案和保单暴露，注入缺失、时间错位与未结案，双盲核对变量后用冻结假设回看。"},"reviewedAt":"2026-10-21","claims":[{"id":"W-F24-result","statement":"合成赔案文本到结构化变量与分层分析的概念验证；两位临床评阅者按量表审阅核心变量。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-CLAIM-VARIABLES","locator":"Executive Summary、§2、Appendix B/D","relation":"partial","supports":"合成赔案文本到结构化变量与分层分析的概念验证；两位临床评阅者按量表审阅核心变量。","limit":"合成数据和专家量表不是实赔准确率；分层准备金示范不证明真实准备金充分或寿险适用。"}],"progressIds":["P-FIN-EXPERIENCE"],"configuration":"GPT-4o-mini原型；合成FHIR及文本；核心变量由两位临床评阅者按1–5量表核对，非实赔准确率。","limitation":"合成数据和专家量表不是实赔准确率；分层准备金示范不证明真实准备金充分或寿险适用。","evidenceStatus":"active"},{"id":"W-F24-practice","statement":"专业精算材料改编的直接LLM任务评估。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-INS-ACTBENCH","locator":"§3.2—3.3、§4.1 Evaluation、Table3、§4.4","relation":"adjacent","supports":"专业精算材料改编的直接LLM任务评估。","limit":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。"}],"progressIds":[],"configuration":"9模型；表格636题/R516题，零样本，答案格与执行结果评分；非本产品整包测试。","limitation":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。","evidenceStatus":"active"},{"id":"W-F24-adoption","statement":"可把可回链变量与经验汇总交给Agent成包交付；文本原型展示的是提取和分层接口，不能替代样本代表性及长期假设检验。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-CLAIM-VARIABLES","locator":"Executive Summary、§2、Appendix B/D","relation":"partial","supports":"已有赔案文本到精算变量的直接LLM概念验证；经验可信度、删失和长期校准仍需本机构数据检验。","limit":"合成数据和专家量表不是实赔准确率；分层准备金示范不证明真实准备金充分或寿险适用。"},{"sourceId":"WU-FIN-INS-ACTBENCH","locator":"§3.2—3.3、§4.1 Evaluation、Table3、§4.4","relation":"adjacent","supports":"专业精算表格与R实作提供相邻执行依据；该研究不直接验证赔案文本变量提取。","limit":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。"}],"progressIds":[],"configuration":null,"limitation":"采用建议是编辑推断；最低接收试验为：固定一批去标识赔案和保单暴露，注入缺失、时间错位与未结案，双盲核对变量后用冻结假设回看。","evidenceStatus":"active"},{"id":"W-F24-implementation","statement":"论文作者公开了赔案文本分析脚本与配置；本次只核查代码和运行说明，未执行模型。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-CLAIMS-IMPLEMENTATION","locator":"README Quick Start；requirements；4_analysis-analyze_claims.py及config/4_analysis-analyze_claims.yaml","relation":"method","supports":"提供可取得的公开实现与运行入口。","limit":"代码可读不证明模型输出正确、跨险种适用或真实业务收益。"}],"progressIds":[],"configuration":null,"limitation":"未安装或重放；目标机构数据、隐私授权、模型API和独立变量核对需另行准备。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"论文附录公开了作者的分析代码。已核固定commit的README、依赖、分析脚本与配置；未安装、调用模型或重放。需要模型API、相应资料和运行环境，不是即用托管服务。","tools":[{"name":"作者公开赔案文本分析实现","url":"https://github.com/mdsight/llm-claims-analysis/tree/a699ccab010698fa0fdb0018981a3ac9e53f40af","version":"commit a699ccab010698fa0fdb0018981a3ac9e53f40af","access":"开放实现及Quick Start；需模型API和数据，未在本站运行。"}]},"variants":[{"title":"结构化经验与假设回顾","inputDifference":"保单暴露、生存或损失经验","assessmentChange":"可确定性汇总后比较候选假设","acceptanceChange":"检查删失、可信度及冻结基线"},{"title":"赔案文本变量提取","inputDifference":"病历或理赔笔记及日期","assessmentChange":"增加回链与冲突处理；仅局部研究直接支持","acceptanceChange":"原文盲审，不能用置信度作为真值"}],"evidenceStatus":"active"},{"id":"W-F25","domainId":"finance","title":"保险产品定价包：保障现金流、保费与利润敏感性","goal":"交付固定产品保障和费用下的保费候选、现金流及利润压力结果。","trigger":"产品新设、费率调整或保障条款变更。","scope":"寿险逐期现金流与产险频度强度模型分别验收；既有准备金实现不支持产品定价结论，法定费率适用另需固定法域核对。","inputs":[{"name":"产品和目标","detail":"保障责任、给付触发、费用、利润或风险目标"},{"name":"定价基础","detail":"寿险死亡退保及折现，或产险暴露、频度强度和通胀"}],"deliverables":["保障与费用映射及基准保费模型","逐期或分风险现金流和利润结果","关键假设敏感性、保障变化与交叉补贴说明"],"acceptance":[{"criterion":"保障可追溯","method":"逐条把给付触发及费用映射到计算，测试边界","failure":"漏掉等待期或保障触发"},{"criterion":"保费可独立复算","method":"寿险按逐期概率折现复算，产险按频度强度及费用复算","failure":"三角准备金被当作新业务价格"},{"criterion":"利润风险暴露","method":"改变退保、费用或大赔尾部，检查结果方向和分层","failure":"基准均值盈利即称所有群体适价"}],"assessment":{"recommendation":"交付固定产品保障和费用下的保费候选、现金流及利润压力结果。 先按下列口径、数值和事实证据接收候选，再决定业务采用。","conditions":["寿险逐期现金流与产险频度强度模型分别验收；既有准备金实现不支持产品定价结论，法定费率适用另需固定法域核对。","输入版本与独立比较基准可用"],"evidenceBasis":"lifelib提供寿险模型方法入口；INS-ActBench表格/R实作为专业执行的相邻证据，未核得本产品全链定价试次。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["寿险逐期现金流与产险频度强度模型分别验收；既有准备金实现不支持产品定价结论，法定费率适用另需固定法域核对。","本站未重放本任务，整包可靠性及人工成本未知"],"mode":"staged-delivery","rationale":"定价计算和保障映射可构成完整候选交付，但扩大到发布费率前需本产品回放和新经验校准；不把通用精算题成绩或准备金方法改名为定价成功。","frontier":"精算LLM已经过表格和R实作评估，公开寿险模型可供重放；真实产品保障映射、长期利润与新产品尾部仍无此处直接成绩。","nextValidation":"选一固定产品由独立模型复算基准保费，注入漏保障、退保变化及费用冲击，并检查分层利润。","futureSignals":["新产品保障边界和独立费率复算均通过，才扩大产品种类","计算通过但费率对新经验持续偏离，须缩小定价采用范围"]},"workflow":{"before":"产品条款、假设和费率模型分别流转，评审难定位利润变化。","now":"Agent把保障触发和假设绑定到版本化模型，运行定价与利润压力，产出可独立复算的费率候选和敏感性；业务目标与风险承受约束一起检查。","retained":["保障可追溯","保费可独立复算","利润风险暴露"],"added":["保存冻结输入、工具配置、候选修订和未决状态","对照独立数值或原文接收，不以输出完整度代替正确性"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现漏掉等待期或保障触发时拒收受影响部分；保留原输入与失败，补齐资料或改模型后重新独立核对。","example":{"input":"固定期限寿险，给付及费用不变，只改变退保假设。","output":"给出逐期现金流、保费及利润敏感性并解释方向。","failure":"把现有赔款准备金算法输出当作保费。"},"sourceIds":["WU-FIN-LIFELIB","WU-FIN-INS-ACTBENCH"],"evidenceLinks":[{"sourceId":"WU-FIN-LIFELIB","supports":"寿险现金流、定价和模型测试的公开实现入口。","notCovered":"README用途声明与代码入口，不是AI实测或任何法域监管适用性结论。"},{"sourceId":"WU-FIN-INS-ACTBENCH","supports":"专业精算材料改编的直接LLM任务评估。","notCovered":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。"}],"resourceRefs":[],"career":{"direction":"保险产品定价包：保障现金流、保费与利润敏感性","evidence":"提交输入口径、可复算产物及一项实质失败的修正记录。","nextStep":"选一固定产品由独立模型复算基准保费，注入漏保障、退保变化及费用冲击，并检查分层利润。"},"team":{"bottleneck":"基准均值盈利即称所有群体适价","allocation":"Agent交付保障与费用映射及基准保费模型；独立接收者按数值与事实证据决定采用。","nextStep":"选一固定产品由独立模型复算基准保费，注入漏保障、退保变化及费用冲击，并检查分层利润。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F25-method-1","statement":"寿险现金流、定价和模型测试的公开实现入口。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-LIFELIB","locator":"What is lifelib / What for；basiclife模型库","relation":"method","supports":"寿险现金流、定价和模型测试的公开实现入口。","limit":"README用途声明与代码入口，不是AI实测或任何法域监管适用性结论。"}],"progressIds":[],"configuration":null,"limitation":"README用途声明与代码入口，不是AI实测或任何法域监管适用性结论。","evidenceStatus":"active"},{"id":"W-F25-practice","statement":"专业精算材料改编的直接LLM任务评估。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-INS-ACTBENCH","locator":"§3.2—3.3、§4.1 Evaluation、Table3、§4.4","relation":"adjacent","supports":"专业精算材料改编的直接LLM任务评估。","limit":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。"}],"progressIds":[],"configuration":"9模型；表格636题/R516题，零样本，答案格与执行结果评分；非本产品整包测试。","limitation":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。","evidenceStatus":"active"},{"id":"W-F25-adoption","statement":"定价计算和保障映射可构成完整候选交付，但扩大到发布费率前需本产品回放和新经验校准；不把通用精算题成绩或准备金方法改名为定价成功。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-LIFELIB","locator":"What is lifelib / What for；basiclife模型库","relation":"partial","supports":"lifelib提供寿险模型方法入口；INS-ActBench表格/R实作为专业执行的相邻证据，未核得本产品全链定价试次。","limit":"README用途声明与代码入口，不是AI实测或任何法域监管适用性结论。"},{"sourceId":"WU-FIN-INS-ACTBENCH","locator":"§3.2—3.3、§4.1 Evaluation、Table3、§4.4","relation":"adjacent","supports":"lifelib提供寿险模型方法入口；INS-ActBench表格/R实作为专业执行的相邻证据，未核得本产品全链定价试次。","limit":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。"}],"progressIds":[],"configuration":null,"limitation":"采用建议是编辑推断；最低接收试验为：选一固定产品由独立模型复算基准保费，注入漏保障、退保变化及费用冲击，并检查分层利润。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"公开研究或模型入口可读，未在本站安装或重放；真实数据接入、API、算力及人工接收成本未知。","tools":[{"name":"lifelib固定寿险模型入口","url":"https://github.com/lifelib-dev/lifelib/blob/b706d8bbab530cc9ef94e06600e9c9d1468d8cb1/README.rst#L16-L39","version":"固定来源版本见链接；研究原型非托管产品","access":"公开阅读；运行需取得代码依赖、模型及授权数据，论文不是即用服务"}]},"variants":[{"title":"寿险定价与利润测试","inputDifference":"死亡退保、保额、逐期费用及折现","assessmentChange":"运行逐期模型并交付利润压力","acceptanceChange":"逐期概率、给付与费用独立复算"},{"title":"产险频度强度定价","inputDifference":"暴露、频度强度及大赔通胀","assessmentChange":"需要另建适配费率模型，寿险代码仅邻近","acceptanceChange":"分层损失成本、费用与尾部压力"}],"evidenceStatus":"active"},{"id":"W-F26","domainId":"finance","title":"保险资本与再保方案：净风险、聚合压力和风险转移","goal":"交付保险风险聚合及备选再保结构的净现金流、压力与资本影响。","trigger":"资本规划、产品结构或再保安排需要比较备选。","scope":"内部经济情景与真实风险转移分开验收；公开压力工具不支持任何法域监管资本合格，条款生效及再保接受需实际状态证据。","inputs":[{"name":"风险及估计","detail":"按险种的毛损失或现金流分布、相关假设、负债与资产"},{"name":"再保和资本约束","detail":"自留、限额、恢复条款、对手、费用、目标及固定制度口径"}],"deliverables":["毛到净损失及再保分摊可复算桥","关联和尾部情景下的资本影响与经济取舍","未覆盖风险、条款未决及方案实际状态"],"acceptance":[{"criterion":"条款改变净风险","method":"手工边界例核对起赔、自留、限额和恢复，匹配责任期间","failure":"再保限额被重复使用"},{"criterion":"聚合可反驳","method":"独立改变相关和尾部，比较分散效应及遗漏风险","failure":"独立假设隐去集中风险"},{"criterion":"计算与生效区分","method":"核对报价、确认及合同状态；制度口径逐项核对来源","failure":"候选净现金流冒充已生效转移或法定合格"}],"assessment":{"recommendation":"交付保险风险聚合及备选再保结构的净现金流、压力与资本影响。 先按下列口径、数值和事实证据接收候选，再决定业务采用。","conditions":["内部经济情景与真实风险转移分开验收；公开压力工具不支持任何法域监管资本合格，条款生效及再保接受需实际状态证据。","输入版本与独立比较基准可用"],"evidenceBasis":"保险模型及金融压力工具可用于局部计算，INS-ActBench是专业实作相邻证据；未找到本任务跨条款、全聚合和真实生效的独立AI结果。","localStatus":"本站未运行本任务模型试次；成功率、重试次数、实际墙钟、API费用和人工复核时间未知。开源许可不等于零运行成本。","gaps":["内部经济情景与真实风险转移分开验收；公开压力工具不支持任何法域监管资本合格，条款生效及再保接受需实际状态证据。","本站未重放本任务，整包可靠性及人工成本未知"],"mode":"staged-delivery","rationale":"可先接收同口径可复算的备选方案；有真实条款和确认后再检验授权内执行，不能从方法存在推资本合格或永久排除Agent操作。","frontier":"公开工具能运行风险与现金流情景，专业LLM有实作评估；真实条款映射、依赖尾部和再保生效的完整AI交付仍是具体缺口。","nextValidation":"用两层限额、恢复条款与相关冲击做沙盒，独立复算净赔付和资本变化，再核对一个拒绝或未确认的再保状态。","futureSignals":["多层条款和相关冲击的新例独立复算通过，才扩大结构复杂度","若真实确认、毛净桥或制度口径不一致，方案不得作为已生效资本减免"]},"workflow":{"before":"毛风险、再保条款与资本情景分散，方案比较常缺少同口径桥。","now":"Agent把条款参数接到毛净风险和压力计算，解释资本与经济取舍，并追踪实际再保接受和生效状态；状态缺口进入未决清单。","retained":["条款改变净风险","聚合可反驳","计算与生效区分"],"added":["保存冻结输入、工具配置、候选修订和未决状态","对照独立数值或原文接收，不以输出完整度代替正确性"]},"humanResponsibility":"明确本任务用途、可用资料与操作权限，接收重大假设和使用后果。可委派计算、分析或授权内操作的范围由验证决定，责任归属不构成永久技术上限。","failureHandling":"出现再保限额被重复使用时拒收受影响部分；保留原输入与失败，补齐资料或改模型后重新独立核对。","example":{"input":"同一毛风险比较比例与超赔两方案，有一再保报价尚未接受。","output":"给出可复算毛净桥、尾部压力和明确未生效状态。","failure":"将报价作为已转移风险并减少正式资本。"},"sourceIds":["WU-FIN-LIFELIB","WU-FIN-ORE-STRESS","WU-FIN-INS-ACTBENCH"],"evidenceLinks":[{"sourceId":"WU-FIN-LIFELIB","supports":"寿险现金流、定价和模型测试的公开实现入口。","notCovered":"README用途声明与代码入口，不是AI实测或任何法域监管适用性结论。"},{"sourceId":"WU-FIN-ORE-STRESS","supports":"可检查的定价风险情景输入和运行入口。","notCovered":"示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。"},{"sourceId":"WU-FIN-INS-ACTBENCH","supports":"专业精算材料改编的直接LLM任务评估。","notCovered":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。"}],"resourceRefs":[],"career":{"direction":"保险资本与再保方案：净风险、聚合压力和风险转移","evidence":"提交输入口径、可复算产物及一项实质失败的修正记录。","nextStep":"用两层限额、恢复条款与相关冲击做沙盒，独立复算净赔付和资本变化，再核对一个拒绝或未确认的再保状态。"},"team":{"bottleneck":"候选净现金流冒充已生效转移或法定合格","allocation":"Agent交付毛到净损失及再保分摊可复算桥；独立接收者按数值与事实证据决定采用。","nextStep":"用两层限额、恢复条款与相关冲击做沙盒，独立复算净赔付和资本变化，再核对一个拒绝或未确认的再保状态。"},"reviewedAt":"2026-10-20","claims":[{"id":"W-F26-method-1","statement":"寿险现金流、定价和模型测试的公开实现入口。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-LIFELIB","locator":"What is lifelib / What for；basiclife模型库","relation":"method","supports":"寿险现金流、定价和模型测试的公开实现入口。","limit":"README用途声明与代码入口，不是AI实测或任何法域监管适用性结论。"}],"progressIds":[],"configuration":null,"limitation":"README用途声明与代码入口，不是AI实测或任何法域监管适用性结论。","evidenceStatus":"active"},{"id":"W-F26-method-2","statement":"可检查的定价风险情景输入和运行入口。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-FIN-ORE-STRESS","locator":"XVA Stress Testing / XVA P&L Explain；run_stress.py与配置文件","relation":"method","supports":"可检查的定价风险情景输入和运行入口。","limit":"示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。"}],"progressIds":[],"configuration":null,"limitation":"示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。","evidenceStatus":"active"},{"id":"W-F26-practice","statement":"专业精算材料改编的直接LLM任务评估。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-FIN-INS-ACTBENCH","locator":"§3.2—3.3、§4.1 Evaluation、Table3、§4.4","relation":"adjacent","supports":"专业精算材料改编的直接LLM任务评估。","limit":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。"}],"progressIds":[],"configuration":"9模型；表格636题/R516题，零样本，答案格与执行结果评分；非本产品整包测试。","limitation":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。","evidenceStatus":"active"},{"id":"W-F26-adoption","statement":"可先接收同口径可复算的备选方案；有真实条款和确认后再检验授权内执行，不能从方法存在推资本合格或永久排除Agent操作。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-FIN-LIFELIB","locator":"What is lifelib / What for；basiclife模型库","relation":"partial","supports":"保险模型及金融压力工具可用于局部计算，INS-ActBench是专业实作相邻证据；未找到本任务跨条款、全聚合和真实生效的独立AI结果。","limit":"README用途声明与代码入口，不是AI实测或任何法域监管适用性结论。"},{"sourceId":"WU-FIN-ORE-STRESS","locator":"XVA Stress Testing / XVA P&L Explain；run_stress.py与配置文件","relation":"partial","supports":"保险模型及金融压力工具可用于局部计算，INS-ActBench是专业实作相邻证据；未找到本任务跨条款、全聚合和真实生效的独立AI结果。","limit":"示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。"},{"sourceId":"WU-FIN-INS-ACTBENCH","locator":"§3.2—3.3、§4.1 Evaluation、Table3、§4.4","relation":"adjacent","supports":"保险模型及金融压力工具可用于局部计算，INS-ActBench是专业实作相邻证据；未找到本任务跨条款、全聚合和真实生效的独立AI结果。","limit":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。"}],"progressIds":[],"configuration":null,"limitation":"采用建议是编辑推断；最低接收试验为：用两层限额、恢复条款与相关冲击做沙盒，独立复算净赔付和资本变化，再核对一个拒绝或未确认的再保状态。","evidenceStatus":"active"}],"availability":{"status":"open-implementation","detail":"公开研究或模型入口可读，未在本站安装或重放；真实数据接入、API、算力及人工接收成本未知。","tools":[{"name":"lifelib固定寿险模型入口","url":"https://github.com/lifelib-dev/lifelib/blob/b706d8bbab530cc9ef94e06600e9c9d1468d8cb1/README.rst#L16-L39","version":"固定来源版本见链接；研究原型非托管产品","access":"公开阅读；运行需取得代码依赖、模型及授权数据，论文不是即用服务"}]},"variants":[{"title":"风险资本压力方案","inputDifference":"聚合暴露、相关与尾部分布","assessmentChange":"候选经济情景；监管口径须另核对","acceptanceChange":"压力复算、遗漏风险与口径对照"},{"title":"产品再保险与风险转移","inputDifference":"保障、自留限额、恢复和对手确认","assessmentChange":"增加条款映射与实际生效证据","acceptanceChange":"边界赔款分摊及接受/生效状态"}],"evidenceStatus":"active"},{"id":"W-M05","domainId":"math","title":"命题形式化：交付自然语言与形式声明的语义对照","goal":"将明确的数学定义和命题转为可编译、语义忠实的形式接口，供证明和复用；独立接收声明而非只查编译。","trigger":"自然语言研究/课程命题需要进入形式库，或已有形式化被怀疑弱化原题。","scope":"定义、量词、对象域、边界和库映射同包；证明搜索可帮助暴露歧义，最终证明是否完成另按W-M01接收。","inputs":[{"name":"原文与允许解释","detail":"可信原命题、定义、领域约定、非平凡/边界实例；哪些歧义需要显式选择。"},{"name":"目标形式项目","detail":"Lean/Mathlib工具链、可复用库、名称约定及可信原文边界；允许引入的定义/接口及既有批准基础，不能以新增未批准公理补缺库。"}],"deliverables":["自然语言—形式定义/声明对应表、可编译Prop接口与锁定依赖；例如def Target : Prop := ...表达待证目标，不把命题定义当定理证明","边界/反例/等价改写检查及缺库项；必要shadow/双向蕴涵检查","题意歧义、已选解释和仍未解接口；不把占位证明当成果"],"acceptance":[{"criterion":"对象与量词忠实","method":"独立逐项核对隐含域、空集/零/极限、量词顺序和前提；对合法等价表达接受双向论证，不要求同字符串。","failure":"把“所有x存在y”改“存在y所有x”；漏去非零条件使原命题变弱或无意义。"},{"criterion":"编译与语义分别通过","method":"干净项目检查定义/Prop接口可编译，不靠sorry/admit或新增未批准公理伪装接口；用原文边界实例和独立反例尝试区分原陈述/生成陈述，必要时用shadow检查。机器通过仍需明确语义对照的覆盖范围，完整证明状态单独记录。","failure":"能类型检查但把结论加为假设；语义judge单独认定正确。"},{"criterion":"后续证明接口可用","method":"固定最终声明、依赖和定义；用小引理或已有定理检验接口组合，缺库明确交付而非伪定义。","failure":"为了可证偷偷改定义；未告知使用者改动原目标。"}],"assessment":{"recommendation":"可委派完整形式接口候选与语义检查材料；先接收题意对应，再接收证明。","conditions":["可信原文及定义齐备","目标库与工具链可用","有能区分错题意的边界或语义检查"],"evidenceBasis":"自动形式化及ShadowBench专属评估显示可编译与语义对应是不同指标，不继承数学证明分数。","localStatus":"本单元AI运行、成功率及人时均为null；未执行上述连续函数例。","gaps":["新领域抽象定义需逐项映射","自动语义评价本身可能误判"],"mode":"staged-delivery","rationale":"已有专属评估使候选接口值得委派；语义评价仍不完备，故以可核对原文和可区分反例接收，而非编译率推成正确率。","frontier":"178题上的Opus4.8+Numina-Lean-Agent编译61.8%而SA-Pass11.2%，具体显示了当前该配置“能编译”与“题意对齐”的差距；这不是本形式化任务或职业普遍率。","nextValidation":"以正确等价声明、弱化/换量词/假设结论三种控制校准语义验收，再试未见领域声明。","futureSignals":["扩大：同义改写和新库上下文下语义对照稳定、独立专家一致；可验证检查覆盖提高。","收缩：只优化compile rate，无法识别循环假设或弱化；缩到接口候选。"]},"workflow":{"before":"数学作者写自然语言定义/命题，形式化者再选库对象、补隐含条件并来回确认含义；常在证明失败后才发现声明翻译失真。","now":"代理以原文和目标项目为边界，组织定义/声明映射、库检索和边界实例；编译反馈解决接口问题，等价表达、区分性反例或试证帮助暴露语义歧义。声明语义接收后即可交给W-M01，毋须先完成主命题证明。","retained":["可信原文、定义和允许解释","形式接口与使用目的的接收"],"added":["区分可编译与语义对齐","保留量词/对象域/边界的对应证据","把歧义与缺库接口作为可接续产物"]},"humanResponsibility":"接收者确认原文约定、歧义解释和形式接口的使用范围；代理可完成库映射、声明草拟和语义检查材料。领域接收与自动区分检查互补，不把人或模型赞同当完备语义算法，也不要求人工逐行指定翻译。","failureHandling":"编译失败时交未映射的定义、缺库接口/依赖及可重现错误；语义未定时交候选解释、它们的形式差异和区分性实例/反例。可编译且语义对应的声明可独立接收，即使主证明未完成；明确标“声明已接收/证明未完成”，转W-M01。不得新增未批准公理消除接口难点，不把占位proof当已证成果。","example":{"input":"设计例：实数连续函数在闭区间[a,b]达到最大值；区间非空、端点关系和连续性域需对照原文。","output":"固定项目的定义/Prop目标接口（如def Target : Prop := ...）、自然语言对应及空区间/端点相等检查；声明可独立接收，目标theorem证明后续委派。没有本例运行结果。","failure":"把闭区间换成开区间或只写有上界；形式声明编译仍不能接收。"},"evidenceLinks":[{"sourceId":"WU-M-AUTOFORM","supports":"给出自动形式化与语义评估的研究证据。","notCovered":"不将论文最佳分数套给本任务或2026模型。"},{"sourceId":"WU-M-SHADOW","supports":"支持形式陈述的shadow与专家一致性检查。；用区分性检查拒绝看似可编译的错题意。","notCovered":"shadow不是对任意命题的完备语义算法。；本连续函数例仅任务设计，尚未验证。"}],"resourceRefs":[],"career":{"direction":"数学形式化、形式接口维护、研究库建设","evidence":"可信原文与可编译形式接口、定义/量词对照、边界与等价表达检查；展示一次可编译却语义失配的声明及其修订。","nextStep":"选择目标数学领域与形式库，提交一组原文—接口对照和缺库/歧义处理记录；完整证明单独标状态，必要时接入W-M01。"},"team":{"bottleneck":"瓶颈在原文歧义、库对象不匹配、接口编译还是语义区分检查？","allocation":"代理组织定义/声明、库检索和语义检查；接收者集中处理原文解释与接口使用取舍。声明接收与后续证明搜索分别计进展。","nextStep":"试点含边界、同义改写和新定义的命题；分别记录接口可编译、语义接收、缺库与歧义成本，不用证明找到与否评价翻译是否忠实。"},"sourceIds":["WU-M-AUTOFORM","WU-M-SHADOW"],"reviewedAt":"2026-10-20","claims":[{"id":"M05-AUTOFORM","statement":"专属研究评估自然语言到形式声明，区分研究级上下文与本科题；不是完成证明分数。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-M-AUTOFORM","locator":"Abstract；评测设置/BEq+","relation":"direct","supports":"给出自动形式化与语义评估的研究证据。","limit":"不将论文最佳分数套给本任务或2026模型。"}],"progressIds":[],"configuration":"EMNLP2025；ProofNet#/RLM25，BEq+与人工判断，原论文模型/采样条件限定；本文不转抄最高分为本站概率。","limitation":"不将论文最佳分数套给本任务或2026模型。","evidenceStatus":"active"},{"id":"M05-SEMANTICS","statement":"ShadowBench在178题上，Claude Code(Opus4.8)+Numina-Lean-Agent编译率61.8%，SA-Pass为11.2%；同一配置可编译远多于语义检查通过。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-M-SHADOW","locator":"§3、§5.2、§6","relation":"direct","supports":"支持形式陈述的shadow与专家一致性检查。","limit":"shadow不是对任意命题的完备语义算法。"}],"progressIds":["P-M-SHADOW"],"configuration":"arXiv v1；178研究生/研究级题；Claude Code+Opus4.8，Medium thinking，加Numina-Lean-Agent搜索/编译；每题单次运行；Table3指标与Table10配置。隐藏shadow checker不可自部署；非Agent采样参数不混入此行。","limitation":"shadow不是对任意命题的完备语义算法。","evidenceStatus":"active"},{"id":"M05-ADOPT","statement":"按题意映射、边界实例和独立语义检查接收接口，之后再接收证明。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-M-SHADOW","locator":"§6及case studies","relation":"method","supports":"用区分性检查拒绝看似可编译的错题意。","limit":"本连续函数例仅任务设计，尚未验证。"}],"progressIds":[],"configuration":null,"limitation":"本连续函数例仅任务设计，尚未验证。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"Lean与部分数据/生成设置可获得，但ShadowBench隐藏checker不是公开可自部署验收器。公开原文/生成数据不等同完整复现同一评分；本任务可自建语义控制并独立审查。","tools":[{"name":"自动形式化评估资料","url":"https://aclanthology.org/2025.emnlp-main.907/","version":"EMNLP2025论文/配套资料","access":"可读评估入口；不能当已安装服务。"},{"name":"ShadowBench","url":"https://arxiv.org/html/2608.29270v1","version":"v1 2026-08-29","access":"公开论文、任务/提交说明及部分设置；隐藏checker未公开，本轮未提交或重放，不能称完整开源评估。"}]},"variants":[{"title":"研究新定义","inputDifference":"抽象对象库中尚不存在。","assessmentChange":"先接收定义接口和语义实例，缺库可独立交付。","acceptanceChange":"定义一致性、非平凡实例及与原文映射，不能为了可证换定义。"},{"title":"固定现成声明","inputDifference":"可信形式声明已经给定。","assessmentChange":"直接用W-M01，不重复自动形式化。","acceptanceChange":"不允许改声明，只验完整证明。"},{"title":"既有定义的自然语言命题","inputDifference":"原文对象在目标库已有表示，但量词、隐含前提和表达可能不同。","assessmentChange":"可整包委派声明及语义对照，主证明仍另收。","acceptanceChange":"核库对象/量词/边界，接受合法等价表达；有定义不代表翻译自动正确。"}],"evidenceStatus":"active"},{"id":"W-M06","domainId":"math","title":"科学仿真交付：模型、数值误差与性能一起可复查","goal":"把给定科学模型变成可复现计算/仿真包，证明误差、稳定性和资源开销满足约定，并明确与现实的距离。","trigger":"需要从方程和数据交付可信模拟，或升级数值实现同时保住精度。","scope":"建模假设、方法选择、求解实现、误差/不确定性和科学软件回归为同一交付；不以收敛代替物理模型正确。","inputs":[{"name":"方程与物理对象","detail":"方程、初/边值、单位、参数来源、守恒关系和观测对照；允许简化与非目标。"},{"name":"精度和资源","detail":"可解析/制造解基准，误差范数/容差、网格/步长序列、随机种子/重复数、运行硬件和性能预算。"},{"name":"可复现项目","detail":"固定源/依赖/编译器/BLAS/求解器、数据hash；目标机器CPU/GPU/HPC入口及权限。"}],"deliverables":["方程/离散化/参数来源与可运行项目、锁定环境","基准与网格/步长收敛、残差/守恒量/稳定性/不确定性记录","相同误差容限下的时间/内存比较、关键路径profile与回归","情景结果、可读说明及模型适用/未验证物理假设"],"acceptance":[{"criterion":"离散问题和科学模型一致","method":"独立推导单位、边界和离散算子；以制造解或解析极限验证，不沿用作者同一段代码生成期望。","failure":"边界索引反了但残差很小；单位秒/毫秒错配。"},{"criterion":"误差/稳定性有证据","method":"比较多级网格或步长的误差阶；守恒/正性等物理不变量和病态参数；随机模拟给Monte Carlo误差与种子集。","failure":"只看一个网格的求解器success；不稳定解用平滑图掩盖。"},{"criterion":"优化后科学结果未变","method":"固定硬件、线程、误差目标和输入，比较精度与耗时/内存；测试依赖升级、不同求解路径；性能改善不得靠减精度。","failure":"GPU更快却换算法容差；只比较一次热缓存时间。"},{"criterion":"现实解释可接收","method":"与独立测量/公认算例核对；误差拆成模型、离散、舍入、参数/观测来源，无法校准则只接收数值实现。","failure":"数值收敛直接声称能预测现场；把拟合数据当独立验证。"}],"assessment":{"recommendation":"可委派完整仿真候选及误差/性能报告；解析/高精度对照和独立物理校验决定接收层级。","conditions":["模型与边界可解释","有独立数值基准","目标运行环境及容差明确"],"evidenceBasis":"科学编码已有研究型任务集；题意/金标准/容差缺陷会误拒正确数值代码。","localStatus":"本站本任务AI运行数、成功率、实际成本及人工介入量均为null；代表例为交付设计，未执行。","gaps":["未覆盖全部物理模型认证","目标HPC/观测数据仍需提供"],"mode":"bounded-delivery","rationale":"科学代码研究展示了可执行产物，也暴露基准规格/误差容差的缺陷。因此可整包实现和迭代，但先校准验收，不用旧低分当上限或修订后分数当现场可靠性。","frontier":"SciCode-Verified作者的匹配重评及公开例外表明，修订科学规格/金标准/容差可大幅改变科学编码测量；两条修订版DeepSeek V4 Preview行采用较早harness与三次均值，不能把12行无例外称为同一harness。它仍不代替现实模型或任意HPC验证。","nextValidation":"在1D热方程制造解下验证误差阶和边界，再给代理未见参数/性能修改；加入错边界、过粗网格及缩容差假提速负控制。","futureSignals":["扩大：未见算例、病态参数和依赖变化下误差与性能双达标。","收缩：只在参考数据接近，网格收敛/守恒失败或物理外推无证据。"]},"workflow":{"before":"理论推导、求解器编码、调参、性能优化和论文图表分开交付。","now":"代理围绕模型与误差目标选方法和实验；确定性求解直接运行，数值反馈引导换网格/算法，结果和解释来自同一版本。","retained":["科学假设与观测来源","可接受误差与资源投入"],"added":["数值验收器也需验证","同时保存精度/性能曲线","库版本与硬件语义"]},"humanResponsibility":"应用领域接收者确认模型与实验对照，数值负责人确认误差证据；代理可完成实现、实验、profile和报告。","failureHandling":"不收敛时交参数/残差/稳定性轨迹与最小复现；区分模型不适、离散错误和环境阻塞，不凭好看曲线补结论。","example":{"input":"设计例：1D热方程u_t=αu_xx，固定区间/零边界和u(x,0)=sin(πx/L)；解析解可对照；再改变α、网格和步长。","output":"可重放求解器、L2/L∞误差及收敛、与零Dirichlet边界通量一致的热量平衡/耗散检查、同精度性能比较；不要求总热量恒定，未生成本例产物。","failure":"显式格式越过稳定条件仍报成功；只降低精度后声称加速。"},"claims":[{"id":"M06-SCICODE","statement":"SciCode-Verified作者的匹配重评报告12模型快照的主问题准确区间约由9–27%到69–92%；多数行沿同一评测设置，修订版两条DeepSeek V4 Preview行另有较早harness/三次均值例外。题意/金标准/容差影响科学编码测量。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-M-SCICODE-VERIFIED","locator":"§3 matched reevaluation；§4缺陷；§6","relation":"partial","supports":"提供科学代码能力测量及验收缺陷的专属证据。","limit":"独立审计者自报；不能把修订分数当本仿真成功率。"}],"progressIds":["P-M-SCICODE-AUDIT"],"configuration":"SciCode-Verified arXiv v1 Table1/§3；64可判定主问题/287子题，with-background，pass@1设计，两套固定科学Python环境OR判分。多数行是一次运行；修订版DeepSeek V4 Pro Preview与DeepSeek V4 Flash Preview两行是较早harness revision的三次运行均值，不是所有12行都仅改数据的一次同harness对照。本站未重跑。","limitation":"作者自报的独立审计；区间跨模型，不是本仿真或职业成功率。两条修订版Preview行的harness/重复例外须保留，不能把每行差值全部归于只换数据；修订测量也不是模型升级。","evidenceStatus":"active"},{"id":"M06-ADOPT","statement":"以解析/制造解、收敛阶、物理不变量和同精度资源比较接收科学软件候选。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-M-SCICODE","locator":"Introduction / Dataset Creation / Evaluation","relation":"adjacent","supports":"科学问题到数值实现的公开任务参考。","limit":"热方程验收及HPC迁移是编辑设计，未测。"}],"progressIds":[],"configuration":null,"limitation":"热方程验收及HPC迁移是编辑设计，未测。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开资料与工具入口已核对；下列论文配置、开源实现和本任务实际运行分别记录。本轮不安装、不调用付费服务，也不继承外部成绩。","tools":[{"name":"SciCode-Verified","url":"https://github.com/flyingwagner/scicode-verified/tree/ddab4a92f8d80a7113ab946628e994b52354d838","version":"ddab4a92…；论文v1，数据release需hash核对","access":"源码/任务/评价；与原SciCode分开保存，需API/科学Python。"}]},"variants":[{"title":"随机模拟与不确定性","inputDifference":"随机输入/蒙特卡洛而非确定PDE。","assessmentChange":"可委派可复算模拟与方差控制，不能只接一条随机轨迹。","acceptanceChange":"独立种子重复、Monte Carlo误差、分布/极限对照；不只同seed像素一致。"},{"title":"HPC性能改造","inputDifference":"并行/GPU或不同编译器/浮点模式。","assessmentChange":"先固定科学误差合同，再委派profile与实现。","acceptanceChange":"不同规模加速/通信/内存与误差，硬件一致；未授权集群不运行。"},{"title":"科学软件依赖升级","inputDifference":"求解器、BLAS、编译器或库API变化而科学模型不变。","assessmentChange":"可委派兼容修改及数值回归，先锁原误差合同和合法数值等价。","acceptanceChange":"干净环境重建、基准/收敛与性能重跑；差异分解为算法/舍入/默认值，不以旧缓存通过。"},{"title":"确定性模型与数值实现","inputDifference":"给定确定性方程/初边值和参数，需要选择适用数值方法或实现求解器，不以旧软件升级为前提。","assessmentChange":"可交完整实现、误差和资源报告；模型现实有效性另核。","acceptanceChange":"解析/制造解、网格或步长误差、残差/稳定性与适用平衡规律；容许合法数值等价实现。"},{"title":"候选数值方法的性质与对照验证","inputDifference":"输入是新算法/离散格式及待证性质，而非只为已定模型选成熟solver。","assessmentChange":"可委派实现、对照与失效搜索；收敛/稳定性/复杂度的全称保证另按M04接受论证或反例，实验不代替证明。","acceptanceChange":"锁算法/参数与假设，独立检查推导和可复现实验；基线公平、网格/精度/资源可比，给出失败区域；自然论证、精确反例、适用形式证书各自合法。"}],"sourceIds":["WU-M-SCICODE-VERIFIED","WU-M-SCICODE"],"evidenceLinks":[{"sourceId":"WU-M-SCICODE-VERIFIED","supports":"提供科学代码能力测量及验收缺陷的专属证据。","notCovered":"独立审计者自报；不能把修订分数当本仿真成功率。"},{"sourceId":"WU-M-SCICODE","supports":"科学问题到数值实现的公开任务参考。","notCovered":"热方程验收及HPC迁移是编辑设计，未测。"}],"resourceRefs":[],"career":{"direction":"科学计算、数值方法、研究软件","evidence":"方程/基准、求解器、收敛与性能证据","nextStep":"按领域补模型理论、硬件与测量背景。"},"team":{"bottleneck":"模型误差、数值误差或算力哪项限制交付？","allocation":"代理组织代码和实验，领域人员核对物理假设，独立算例校准接收。","nextStep":"比较未见算例的精度、资源和总接收成本。"},"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-M07","domainId":"math","title":"研究设计包：估计目标、抽样/随机化与分析计划","goal":"把研究问题变成可执行设计、样本量依据和预先分析方案，明确可识别结论与假设。","trigger":"研究尚未看结果，需要设计调查/实验/临床研究并冻结主要分析。","scope":"包含目标总体/估计目标、设计选择、功效与敏感性、缺失策略和分析计划；不把文档清单分数当有效设计，监管/伦理决策另由相应主体作出。","inputs":[{"name":"研究目标","detail":"研究问题、目标总体、终点/时间点/估计量、最小有意义差异、可接受风险与资源。"},{"name":"设计资料","detail":"抽样框/随机化单位、聚类/分层、先验方差/事件率来源、失访与执行限制；临床给锁定方案和intercurrent events。"},{"name":"冻结与接收","detail":"主要/次要/探索分析、multiplicity、中期查看/盲态边界、缺失假设与替代情景；估计目标变化必须有版本。"}],"deliverables":["设计与估计目标对照、抽样/随机化方案和实施要求","可重跑样本量/功效计算及参数来源/敏感性曲线","统计分析计划：分析集、模型、缺失/偏离、中期与多重性、表图模板","可识别性/选择偏差风险、未决设计选择及不同选择的影响"],"acceptance":[{"criterion":"设计回答原问题","method":"独立列总体、变量、比较、时点与估计目标；区分随机/观察设计，确认干扰、选择和依从性假设。","failure":"要治疗策略效应却只分析完成者；观察相关性写成随机试验效应。"},{"criterion":"样本量和实施一致","method":"另一计算路径复算功效/样本量；模拟拟定设计的I类错误/覆盖/功效，考虑分配比、聚类、失访和效应大小不确定。","failure":"把总样本当每组；忽略ICC、双侧/单侧或多重检验。"},{"criterion":"缺失和分析计划可执行","method":"从空结果表和合成边界数据推演分析；临床敏感性须同估计目标，补救治疗/死亡等中间事件策略不得静默变。","failure":"填“使用适当方法”无算法；换目标却说只是敏感性。"},{"criterion":"版本及选择不追结果","method":"对照锁定方案与输出，重大统计选择逐项标来源/新建议/未定；中期方案检验alpha/访问隔离，未授权结果不能用于设计优化。","failure":"看过显著结果后回写“预设”；幻造临床方案未规定的敏感性分析。"}],"assessment":{"recommendation":"可委派完整设计候选和可重算功效/SAP；先接收估计目标与实施假设，再冻结分析，不只审核文案。","conditions":["问题及重要性可定义","设计参数有来源/情景","能独立检验操作特征"],"evidenceBasis":"SAPAI在9份临床方案上比较3模型的27份SAP，由统计师双评分；描述项与统计推理项表现不同。","localStatus":"本站本任务AI运行数、成功率、实际成本及人工介入量均为null；代表例为交付设计，未执行。","gaps":["直接AI证据主要为方法选择/临床SAP，调查设计属迁移","伦理与监管裁决不在生成任务内"],"mode":"staged-delivery","rationale":"SAPAI直接支持方案到计划草稿，同时显示统计推理项仍有漏项；StatQA支持先查方法适用性。因此采用到可计算、可推演的设计候选，不把格式完整当方法正确。","frontier":"已出现临床SAP专属验证；样本和模型受限，临床试验注册/合规和长期结果不由AI文稿证明。","nextValidation":"在公开或授权方案上隐藏参考SAP，让代理输出估计目标、功效代码和计划；先用错误单侧/忽略聚类/改变估计目标负控制校准，再独立模拟。","futureSignals":["扩大：未见设计的操作特征独立复现、条款与终点一致，统计选择可追溯。","收缩：敏感性幻造、参数无来源或I类错误失控；缩到方案对照与计算候选。"]},"workflow":{"before":"设计专家口述、文档作者转写、程序员算样本量，往返澄清。","now":"给研究目标和约束，代理自选计算、设计比较和计划组织；模拟反馈能修正设计，关键选择有依据，不固定一条统计流程。","retained":["研究问题/伦理与实施可行性","估计目标与可接受风险"],"added":["设计到代码操作特征核对","方案—SAP差异可追溯"]},"humanResponsibility":"研究团队决定目标、资源和可实施取舍；统计接收者独立检验设计假设/操作特征，代理可完成草案、计算和边界推演。","failureHandling":"缺参数时交情景范围和决策影响，不造方差/事件率；问题不可识别则交所需设计修改；预算不足区分探索建议与冻结计划。","example":{"input":"设计例：两组12周连续终点试验，给定差异δ、SD和失访范围；明确所有随机受试者的目标及补救治疗处理；不填无来源n。","output":"估计目标、分配/分析集、可重算n与功效曲线、缺失敏感性及表图模板；尚无本例运行。","failure":"只排除失访者并宣称原估计目标不变；用总体人数代替每组样本量。"},"claims":[{"id":"M07-SAPAI","statement":"SAPAI用9方案/3模型生成27份SAP，统计师清单评价描述项准确81–83%、统计推理项67–72%；文稿填全不能代表设计推理同样可靠。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-M-SAPAI","locator":"Abstract Methods/Results","relation":"partial","supports":"临床计划生成的原始领域评估及漏项/幻造风险。","limit":"预印本，清单项分不是整计划通过概率；未验证自主设计或合规。"}],"progressIds":["P-M-SAPAI"],"configuration":"v1；GPT-5、Claude Sonnet4、Gemini2.5Pro；逐节提示，46项0–3统计师双评分；模型/采样成本不用于本站估计。","limitation":"预印本，清单项分不是整计划通过概率；未验证自主设计或合规。 百分比为原研究清单项准确性，不是27份SAP全部通过率。","evidenceStatus":"active"},{"id":"M07-ESTIMAND","statement":"临床主分析与敏感性需要围绕同一估计目标对齐，设计不能靠生成文案决定。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-M-ICH-E9","locator":"Page contents / R1 Current version","relation":"method","supports":"估计目标/敏感性工作约束。","limit":"指南不证明AI设计正确。"}],"progressIds":[],"configuration":null,"limitation":"指南不证明AI设计正确。","evidenceStatus":"active"},{"id":"M07-ADOPT","statement":"研究设计可按估计目标、独立功效/操作特征和可执行计划接收候选，而非仅检查文稿。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-M-STATQA","locator":"Overview / applicability errors","relation":"adjacent","supports":"方法适用性错误要求显式核对前提。","limit":"抽样/因果设计是迁移任务，未有本站直接模型成绩。"}],"progressIds":[],"configuration":null,"limitation":"抽样/因果设计是迁移任务，未有本站直接模型成绩。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开资料与工具入口已核对；下列论文配置、开源实现和本任务实际运行分别记录。本轮不安装、不调用付费服务，也不继承外部成绩。","tools":[{"name":"StatQA","url":"https://github.com/HKUSTDial/StatQA/tree/3d5f6b5b6926600dc952f39d79dfdef82999aeeb","version":"3d5f6b5…；NeurIPS2024","access":"公开数据/脚本，不能当研究设计成品。"},{"name":"SAPAI公开研究","url":"https://www.medrxiv.org/content/10.64898/2026.03.19.26348626v1","version":"v1 2026-03-19","access":"可读论文；所链github本轮未核实可下载，不冒称即用产品。"}]},"variants":[{"title":"抽样调查","inputDifference":"分层/整群/不等概率及不应答。","assessmentChange":"可交抽样/加权设计和偏差情景，不能套两独立样本公式。","acceptanceChange":"纳入概率、设计效应/权重、覆盖误差和不应答敏感性。"},{"title":"临床试验计划","inputDifference":"包含终点、估计目标、中间事件及中期访问。","assessmentChange":"分阶段接收设计/SAP与盲态计算；不自动改变医学或监管目标。","acceptanceChange":"协议—估计目标—分析集—模型—敏感性一致；中期alpha和访问边界。"},{"title":"观察性因果设计","inputDifference":"没有随机分配，混杂/选择假设未知。","assessmentChange":"先交可识别性与偏差分析，不能自动承诺因果结论。","acceptanceChange":"明确时间零点、混杂、positivity、敏感性；不以p值接收因果。"},{"title":"非临床随机或配对实验设计","inputDifference":"非临床研究拟采用独立/聚类随机、配对或重复测量实验，未必涉及患者或临床终点。","assessmentChange":"可交实验设计、功效和分析计划候选；临床SAP成绩不直接迁移成该设计已测。","acceptanceChange":"独立核实验单位、分配/配对/聚类、测量时点和假设；模拟操作特征及效应/失访情景，不套不匹配的两独立样本公式。"},{"title":"实验室研究设计","inputDifference":"输入包括实验处理、供体/培养物等生物实验单位、技术重复、批次/板位、检测方法与资源。","assessmentChange":"可委派设计和可重算功效/分析候选；BixBench3执行给定方法不证明代理会自主选对实验设计。","acceptanceChange":"定义生物重复与技术重复、随机化/平衡批次、对比及多重性；避免伪重复，缺少方差/批次信息时交参数情景而非编n。"}],"sourceIds":["WU-M-SAPAI","WU-M-ICH-E9","WU-M-STATQA"],"evidenceLinks":[{"sourceId":"WU-M-SAPAI","supports":"临床计划生成的原始领域评估及漏项/幻造风险。","notCovered":"预印本，清单项分不是整计划通过概率；未验证自主设计或合规。"},{"sourceId":"WU-M-ICH-E9","supports":"估计目标/敏感性工作约束。","notCovered":"指南不证明AI设计正确。"},{"sourceId":"WU-M-STATQA","supports":"方法适用性错误要求显式核对前提。","notCovered":"抽样/因果设计是迁移任务，未有本站直接模型成绩。"}],"resourceRefs":[],"career":{"direction":"统计设计、试验/调查统计","evidence":"设计、参数出处、功效程序及计划差异","nextStep":"补目标领域和设计理论，展示一个错误设计如何被拒。"},"team":{"bottleneck":"设计不明确还是重复算数耗时？","allocation":"代理比较设计与计算，接收者裁决科学目标并独立检查操作特征。","nextStep":"用未见方案比较完整设计错误与接收人时。"},"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-M08","domainId":"math","title":"统计复现交付：分析集、推断与敏感性同源可重算","goal":"从锁定数据和设计产出可追溯的分析集、统计结果和解释，保留不确定性、偏离与适用范围。","trigger":"已接受设计/方案或需复现研究，等待数据转换、建模诊断和表图报告。","scope":"包含数据质量、分析集转换、模型拟合/诊断、缺失与敏感性、结果报告；预测生产系统或临床诊断不在本包。","inputs":[{"name":"方案和估计对象","detail":"锁定分析计划、终点/比较、总体/样本、模型、缺失/偏离策略及探索标识。"},{"name":"数据和字典","detail":"原始数据hash、采集/随机化单位、重复测量/删失/编码、时间零点、可追踪转换和盲态/访问权限。"},{"name":"复现标准","detail":"固定R/Python及包版本/随机种子、关键输出容差、独立控制和参考方法；允许数值等价结果而非截图像素一致。"},{"name":"实验室/组学元数据（适用时）","detail":"原始文件到样本/处理条件/生物重复/技术重复/批次的映射；参考基因组/注释与特征ID、预定质控/归一化/对比和多重性规则，未规定的方法选择列未决。"}],"deliverables":["只读原始数据到分析集的转换代码/规则、计数桥接、排除原因与质量问题","全流程项目/锁定环境和运行入口，估计值、区间/检验、模型诊断","预设主分析和敏感性/探索表图，单位/分母/删失与不确定性可追踪","偏离清单、可支持结论与未解决假设，不以显著性替代效应/临床意义"],"acceptance":[{"criterion":"分析集可追溯且适用","method":"独立核对原始到分析集的主键、计数、时间/处理组/删失，手工参考切片和边界记录逐条比对。","failure":"重复测量当独立样本；随机化组被按实际治疗覆盖；删失时点使用未来信息。"},{"criterion":"推断目标与方法相符","method":"独立核查抽样/聚类、模型/检验条件、基准水平、对比方向及多重性；参考实现重算关键估计/区间。","failure":"模型能拟合但方差结构错误；训练后挑最小p值冒充预设。"},{"criterion":"诊断和敏感性实际运行","method":"生成残差/拟合与影响诊断；按计划运行缺失、删失、模型假设情景并报告失效；模拟数据校准覆盖/偏差，记录Monte Carlo误差。","failure":"删去影响结论的样本不披露；只在文字写“敏感性一致”而无代码。"},{"criterion":"报告可重建且结论守边界","method":"干净环境从数据生成最终表图/文字数字，核对估计量、区间、比较、临床/实际含义及探索标签。","failure":"表图和正文不同分母；“无显著差异”写成等效；相关性写因果。"}],"assessment":{"recommendation":"方案明确且数据可用时，可整包委派分析执行与候选报告；以分析集追溯、独立数值核对和诊断/敏感性接收。","conditions":["原始数据和方案可追溯","独立统计控制/参考切片","环境和访问适配"],"evidenceBasis":"BixBench3用给定方法和原始数据评估研究尺度分析；最高系统的任务平均产物通过比例为0.48，不能读为48%研究全部完成。；统计方法适用性本身需验收，算出一个数不能代替选择正确方法。","localStatus":"本站本任务AI运行数、成功率、实际成本及人工介入量均为null；代表例为交付设计，未执行。","gaps":["生物计算证据不等于临床试验分析全可靠","未覆盖各类Bayes/纵向/复杂设计的全部方法"],"mode":"bounded-delivery","rationale":"BixBench3直接展示按给定方法处理原始生物数据的长链执行；StatQA显示方法适用不能只靠计算。可委派执行并保留科学/临床设计边界，不把生物产物平均分等同整项研究或试验正确。","frontier":"2026研究尺度分析已有公开任务/产物grader；依赖链、规模、格式和恢复仍能导致不完整，设计选择并未由该基准评测。","nextValidation":"冻结授权数据/方案，先校准错分析集、忽略聚类和删失负控制；代理盲跑后由独立计算检查关键估计和图表；采样模拟另存误差与失败。","futureSignals":["扩大：未见数据与方案偏离下，能追溯转换、识别不适用并重现主/敏感性输出。","收缩：只在参考表可见时正确、静默改分析集或伪造缺失结果；缩到可核对的数据/局部计算。"]},"workflow":{"before":"数据清理、统计程序、图表和解释拆开，版本和分母容易漂移。","now":"代理围绕估计目标组织转换、计算、诊断与解释；反馈来自数据控制和模型检查，稳定复现用固定项目，缺口返回方案层。","retained":["估计目标及方案偏离取舍","科学/临床结论采纳"],"added":["分析集计数桥接","端到端干净重跑","不适用和未完成出口"]},"humanResponsibility":"研究接收者确认估计目标、偏离和使用含义；代理可执行转换/拟合/诊断/表图。临床用途由有职责的研究团队接收，不把LLM复述当安全裁决。","failureHandling":"缺失/不一致数据列具体记录与影响；不收敛报告识别/数值问题；盲态受限时给可执行脚本和已知检查，不越权读取分组或补造结果。","example":{"input":"设计例：纵向12周试验，按原随机组分析，给基线/访视/缺失与补救治疗记录；锁定连续终点模型与缺失情景。","output":"分析集桥接、参数化项目、估计差/区间及诊断、缺失敏感性与可重建表图；尚未执行。","failure":"只分析完整病例却称意向性总体，或把多访视记录当独立扩大样本。"},"claims":[{"id":"M08-BIO-EXEC","statement":"BixBench3用给定方法和原始数据评估研究尺度分析；最高系统的任务平均产物通过比例为0.48，不能读为48%研究全部完成。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-M-BIX3","locator":"§2.1/2.2；§3.2；§4.2/4.4；附录A.5 TCER-1例的Research Objective/Method Guidance/Required Outputs","relation":"partial","supports":"真实原始生物数据到结构产物的领域执行证据。 附录公开例给出mRNA/小RNA原始输入、样本/重复、参考版本及归一化/差异产物合同，可作为实验室分支具体入口。","limit":"不评价研究设计；按任务平均产物比例不同于原始138产物统一分母，也非临床推断可靠性。"}],"progressIds":["P-M-BIX3"],"configuration":"v1；20研究/138产物/13模型；GPT5.6Sol最高任务平均0.48；固定方法指导，32vCPU/128GB、24小时、InspectReAct；产物阈值0.80专家校准。","limitation":"不评价研究设计；按任务平均产物比例不同于原始138产物统一分母，也非临床推断可靠性。","evidenceStatus":"active"},{"id":"M08-METHOD","statement":"统计方法适用性本身需验收，算出一个数不能代替选择正确方法。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-M-STATQA","locator":"Overview；paper task definition/error analysis","relation":"partial","supports":"原作者专门评价任务/列/方法选择和适用性。","limit":"历史有限方法集/合成问法，不代表完整统计职业成功率。"}],"progressIds":["P-M-STATQA"],"configuration":"NeurIPS2024；11,623构建样例，方法前提/元数据/标注限定；GPT4o等历史模型，本文不继承其分数。","limitation":"历史有限方法集/合成问法，不代表完整统计职业成功率。","evidenceStatus":"active"},{"id":"M08-ADOPT","statement":"锁定方案后，可委派完整可重算分析；关键接收从原始到分析集、估计量到解释均要独立核查。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-M-BIX3","locator":"§3.2限制、§4.1方法指导","relation":"adjacent","supports":"按已定方法执行与主动决定研究方法分开。","limit":"临床/调查/方法学模拟为显式迁移，需要对应控制。"}],"progressIds":[],"configuration":null,"limitation":"临床/调查/方法学模拟为显式迁移，需要对应控制。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开资料与工具入口已核对；下列论文配置、开源实现和本任务实际运行分别记录。本轮不安装、不调用付费服务，也不继承外部成绩。","tools":[{"name":"BixBench3 runner/grader","url":"https://github.com/EdisonScientific/BixBench3/tree/d0e0bbb41222335b1b8878f533a58466a3a782dc","version":"d0e0bbb…；release v1.0.0","access":"Python3.12/uv、GCP配额和计费、模型API；部分CellRanger需许可；本轮未运行。"},{"name":"StatQA","url":"https://github.com/HKUSTDial/StatQA/tree/3d5f6b5b6926600dc952f39d79dfdef82999aeeb","version":"固定代码3d5f6b5…","access":"公开方法适用性任务；不等同本研究全流程。"}]},"variants":[{"title":"临床纵向/生存分析","inputDifference":"分析集、时间零点、删失、intercurrent events及盲态。","assessmentChange":"可委派锁定方案执行，重大偏离/新方法回到方案接收；不自动解盲。","acceptanceChange":"原随机组/分析集桥接、删失定义、模型诊断和同估计目标敏感性。"},{"title":"调查/缺失数据","inputDifference":"权重、整群、抽样设计与不应答。","assessmentChange":"可委派加权推断和缺失情景，不套独立样本公式。","acceptanceChange":"设计权重、方差估计/有效样本量、缺失机制与偏差敏感性。"},{"title":"统计方法研究复现","inputDifference":"比较新估计/检验，需理论和模拟。","assessmentChange":"可委派计算实验与反例；新方法性质用M04/M03另核，不因模拟好看成立。","acceptanceChange":"预设数据生成机制、覆盖/偏差/功效、MonteCarlo误差、失败条件和基线公平。"},{"title":"中期/安全性报告","inputDifference":"仅授权窗口/人群可看分组，可能有停止边界。","assessmentChange":"分阶段委派限定数据计算，停止/临床判断按研究职责接收。","acceptanceChange":"独立统计控制、时间点/alpha边界/暴露分母，访问与报告版本；不把普通终期分析复用为中期。"},{"title":"按既定设计的统计推断与报告","inputDifference":"设计/估计目标和分析数据已给定，任务不是专门比较新方法、调查加权或临床生存分析。","assessmentChange":"可整包委派拟合、诊断、敏感性和可复算说明；按实际设计选择方法，不额外假设特殊数据机制。","acceptanceChange":"分析集—估计量—区间/检验—表图/正文同源，核方法前提及独立关键数；特殊抽样/临床/组学输入另选相应变体。"},{"title":"实验室/组学分析","inputDifference":"原始测序/检测与样本表、处理条件、生物重复/技术重复、批次和参考资料需变成统计产物。","assessmentChange":"BixBench3给定方法的原始生物数据执行与此最贴近；可整包委派质量控制、处理/归一化和分析产物，设计选择仍需另核。","acceptanceChange":"样本表到数据对应及生物实验单位、技术重复合并/依赖、参考版本/特征宇宙、归一化/批次与对比方向、多重比较和失败样本均需检查；独立重算关键对比，不能把细胞/reads当独立受试单位。"}],"sourceIds":["WU-M-BIX3","WU-M-STATQA"],"evidenceLinks":[{"sourceId":"WU-M-BIX3","supports":"真实原始生物数据到结构产物的领域执行证据。；按已定方法执行与主动决定研究方法分开。","notCovered":"不评价研究设计；按任务平均产物比例不同于原始138产物统一分母，也非临床推断可靠性。；临床/调查/方法学模拟为显式迁移，需要对应控制。"},{"sourceId":"WU-M-STATQA","supports":"原作者专门评价任务/列/方法选择和适用性。","notCovered":"历史有限方法集/合成问法，不代表完整统计职业成功率。"}],"resourceRefs":[],"career":{"direction":"应用统计、生物统计、研究复现","evidence":"分析集/代码/环境与数值控制、诊断敏感性","nextStep":"明确处理了什么设计及缺失/偏差，不用会作图代替推断。"},"team":{"bottleneck":"转换、方法选择还是复现版本限制交付？","allocation":"代理承担整包计算与解释，独立路径核对重大结果，方案争议集中处理。","nextStep":"比较清洗/计算/复核总成本和静默错误。"},"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-M09","domainId":"math","title":"数学教学包：可诊断理解的材料、提问与反馈","goal":"围绕课程目标交付数学正确、层次可用并可观察学生理解的教学/辅导包。","trigger":"课程需要新单元、差异化辅导或误解诊断，已有课程要求和学生先修范围。","scope":"包括目标/先修、讲义/题目/评分、分层提示和误解诊断；材料正确与教学效果分开，后者用W-M13接收。不是把模型解题成绩写成学生学习。","inputs":[{"name":"课程与学生","detail":"课程目标、先修、大学/高阶或基础范围、学生已有作答与常见误解；不假定所有学生已有相同背景。"},{"name":"教学与评价要求","detail":"允许证明强度、符号/教材约定、学习时段与材料形式、独立作答规则、无AI评价题和评分标准。"},{"name":"反馈边界","detail":"提示何时给出/如何递进、不能泄露的评价答案、哪些作答需再诊断；教师可接收的用语/无障碍需求。"}],"deliverables":["目标—先修—活动—评价对照、正确讲义及分层例/非例","题库、独立解答/证明、评分rubric及可接受替代法","基于实际/模拟错误的诊断问题与提示分支，不把猜测当学生事实","试教/反馈记录入口、未核对知识点和效果验证设计"],"acceptance":[{"criterion":"数学和课程范围正确","method":"独立解题/证明检查所有关键题与反例，核对定义域/边界和先修；可用精确程序或形式化核部分内容，但不强制Lean。","failure":"把sqrt(x²)=x教成全实数公式；用尚未教的定理掩盖证明。"},{"criterion":"提示促进可观察推理","method":"用预先准备的错解/多解轨迹试走提示，检查先问理由、针对性反馈及替代路线；评价题答案不混进辅导提示。","failure":"“苏格拉底式”措辞实际逐字给最终答案；错把计算失误当概念缺陷。"},{"criterion":"评分和诊断可靠","method":"独立双评分一组正确替代、部分理解、偶然猜中和表面抄写作答；分歧可定位，反馈必须引用真实作答。","failure":"正确非标准证明被拒；把模型对学生的印象当学习诊断。"},{"criterion":"交付不冒充效果","method":"核对输出标签：材料审查、试教可用、独立学习效果分别标识；效果另收无AI/延迟/迁移结果。","failure":"因为模型会做题或学生完成练习快就写教学有效。"}],"assessment":{"recommendation":"可整包委派课程材料和辅导候选，并用独立数学检查与作答轨迹接收；正式声称学习效果须另验证。","conditions":["目标/先修清楚","题目可独立检查","有学生作答或诊断控制"],"evidenceBasis":"Tutor CoPilot的现场试验研究真人导师接收AI建议后的课题掌握与辅导行为，而不是模型单独解题。","localStatus":"本站本任务AI运行数、成功率、实际成本及人工介入量均为null；代表例为交付设计，未执行。","gaps":["学校辅导证据不是大学高阶数学效果","具体学生/课程需实际授权试教"],"mode":"staged-delivery","rationale":"真实教学试验支持把解释方式和工具使用方式纳入设计，同时显示辅助成绩与独立学习不同。因此可委派可检查教学包，不从学校试验推大学长期效果。","frontier":"已有真人辅导中的实时AI建议试验；它支持教学组织变化，仍不同于无人完成整门大学课。","nextValidation":"选一个大学课程单元，先用正确/错解/替代解校准材料与评分，再在授权教学中收独立作答和反馈；学习效果用M13另设计。","futureSignals":["扩大：新学生/新概念中数学正确、提示能定位误解，独立迁移结果有证据。","收缩：频繁给答案、拒绝合法解或幻造学生状态；缩到已核教材改编。"]},"workflow":{"before":"教师独立写讲义、题库、讲解与批改，各处目标和错误样本难复用。","now":"代理围绕课程目标组织材料、例/非例、提示和评分；实际作答反馈驱动修订，题目生成与答案独立检查并行但不强迫固定顺序。","retained":["课程目标与学生支持责任","最终课程使用及公平性取舍"],"added":["区别辅助作答与独立掌握","保存评分/提示版本","反例和合法替代答案控制"]},"humanResponsibility":"课程负责人确定学习目标和使用范围，独立数学/教学接收者验证内容与诊断；代理可完成材料和反馈设计，不把责任分配当永久能力上限。","failureHandling":"数学错误返回相应材料/题库版本；没有真实学生信息时标诊断假设；学生持续误解时改解释/活动，不继续重复同答案。","example":{"input":"设计例：大学微积分“导数为零与局部极值”，给先修和学生错解；需要例、非例、提示和独立评价。","output":"讲义/题库、x³等反例、多个合法证明/解释、提示分支及评分；独立学习结果尚未测。","failure":"把导数为零当极值充分条件；AI答案出现在无AI评估题的提示里。"},"claims":[{"id":"M09-TUTOR","statement":"Tutor CoPilot的现场试验研究真人导师接收AI建议后的课题掌握与辅导行为，而不是模型单独解题。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-M-TUTORCOPILOT","locator":"RCT design / Results / Limitations","relation":"partial","supports":"人机教学协作的原始效果研究。","limit":"真人建议辅助、特定学生和课题；不外推大学或长期保持。"}],"progressIds":["P-M-TUTOR"],"configuration":"arXiv v1；预注册真人导师/学生试验，意向性分析课题掌握；辅助者与学生分开，本文不复制分数为本站效果。","limitation":"真人建议辅助、特定学生和课题；不外推大学或长期保持。","evidenceStatus":"active"},{"id":"M09-ADOPT","statement":"正确材料、可区分误解的提示与独立评分可整包委派；学习效果另用独立学生结果接收。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-M-TEACH-RCT","locator":"Design / unassisted exam / limitations","relation":"adjacent","supports":"辅助表现不能替代独立学习证据。","limit":"高中短期证据迁移到大学材料为待验证设计。"}],"progressIds":[],"configuration":null,"limitation":"高中短期证据迁移到大学材料为待验证设计。","evidenceStatus":"active"},{"id":"M09-ENTRY","statement":"Khanmigo官方提供教师备课/评价材料与学习辅导入口，角色和地区/学校访问不同；可用性不等于研究中同版本或学习效果。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-M-KHAN-ENTRY","locator":"For teachers / For learners / FAQ classroom access and eligibility","relation":"method","supports":"真实产品入口及访问前提。","limit":"未调用；官方功能说明不是大学内容正确或学习有效的成绩。"}],"progressIds":[],"configuration":null,"limitation":"使用时锁定实际版本/访问方式，不能继承学校试验效果。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开资料和工具入口已核对；论文配置、当前实现声明及本任务成绩分别记录，未运行本任务。","tools":[{"name":"Tutor CoPilot研究及方法","url":"https://arxiv.org/html/2410.03017v1","version":"v1 2024-10-03","access":"论文/研究资料入口；不冒称有自部署同配置产品。"},{"name":"课程内容的精确核查","url":"https://lean-lang.org/doc/reference/latest/ValidatingProofs/","version":"适用形式子题需锁定工具链","access":"可选形式核查入口；自然论证可按领域方法接收，不为所有课程强制形式化。"},{"name":"Khanmigo官方产品","url":"https://www.khanmigo.ai/","version":"持续服务，2026-10-20读取；具体部署版使用时记录","access":"教师/学习者/学校入口不同，受地区/账户资格影响；课堂学生访问不能由教师个人注册自动授予，未开通服务。"}]},"variants":[{"title":"高阶证明课程","inputDifference":"题目多为自然论证、学生解法差异大。","assessmentChange":"可委派题目/引导与评分候选；不可把模型judge直接当分数。","acceptanceChange":"领域数学独立检验关键推导、合法替代证明和部分理解。"},{"title":"个性化辅导","inputDifference":"使用学生真实作答，解释随误解变化。","assessmentChange":"可委派反馈/提示候选，具体诊断必须有作答证据。","acceptanceChange":"独立作答、针对错误的追问；不能仅用满意度或交互次数证明理解。"},{"title":"大学基础课程单元","inputDifference":"如微积分/线性代数基础，重点概念、计算与初步论证，不要求高阶证明课程输入。","assessmentChange":"可交材料/提示/评价候选，学习效果仍另收。","acceptanceChange":"先修/定义域/例与非例/计算和迁移题匹配；独立正确性与多种合理解法，不以模型答题能力当学生效果。"},{"title":"学生项目指导","inputDifference":"输入为学生的研究/建模/证明项目目标及阶段产物，而非一套已定高阶课程讲义。","assessmentChange":"可委派阶段反馈、资源和检查候选；学生独立贡献/理解靠本人解释与新任务验证。","acceptanceChange":"反馈引用具体产物与出处，能解释假设/错误和局限；AI生成报告不能替代学生独立掌握。"}],"sourceIds":["WU-M-TUTORCOPILOT","WU-M-TEACH-RCT","WU-M-KHAN-ENTRY"],"evidenceLinks":[{"sourceId":"WU-M-TUTORCOPILOT","supports":"人机教学协作的原始效果研究。","notCovered":"真人建议辅助、特定学生和课题；不外推大学或长期保持。"},{"sourceId":"WU-M-TEACH-RCT","supports":"辅助表现不能替代独立学习证据。","notCovered":"高中短期证据迁移到大学材料为待验证设计。"},{"sourceId":"WU-M-KHAN-ENTRY","supports":"产品/角色访问入口。","notCovered":"不证明学习或大学课程效果。"}],"resourceRefs":[],"career":{"direction":"高校数学教学与辅导","evidence":"目标/教材/题目、数学核查、学生真实作答与评分分歧","nextStep":"分清会讲解和已证明学习有效，补真实教学观察。"},"team":{"bottleneck":"内容准备慢还是学生理解诊断不准？","allocation":"代理生成/调整教学包，数学接收者校内容，教学团队以真实作答判断反馈。","nextStep":"用一单元记录错题、评分分歧和独立迁移结果。"},"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-M10","domainId":"math","title":"程序与系统验证：规格、反例和证明产物相互对照","goal":"为指定实现交付可追溯的形式规格与验证结果，解释哪些实际要求已覆盖、哪些仍未证。","trigger":"关键程序需验证、已有规格需维护，或模型检查给出反例需要定位真实缺陷。","scope":"需求到规格、环境抽象、代码/模型变更和验证同包；固定证明义务可以整包搜索，系统级规格需分阶段接收。不把Nexus数学成果当工业保证。","inputs":[{"name":"真实需求与对象","detail":"固定代码/硬件模型、目标属性（安全性/活性/数值范围等）、接口/状态、外部环境假设、并发/内存模型和实现映射。"},{"name":"验证环境","detail":"指定Dafny/Rocq/Frama-C或模型检查器版本、solver/编译器、可信库/公理及允许改动区；超时/有限界说明。"},{"name":"独立控制","detail":"正例与能触发溢出/空集/并发等反例、已有测试/规格及旧版证明；不能只检查Agent写的assert。"}],"deliverables":["需求—规格—实现/模型映射与假设/未编码需求清单","完整验证项目、锁定工具/solver、证明脚本/模型检查状态及原始日志","可重放反例及其映射：真实代码缺陷、规格错误或抽象伪反例","验证驱动代码/规格差异与回归、升级后的证明维护/信任基础记录"],"acceptance":[{"criterion":"证明的属性对应真实需求","method":"独立对照需求、前/后置条件、状态/环境假设；检查可满足性和vacuity，确认不是把期望写进前提。","failure":"requires false导致全通过；数学整数模型掩盖机器溢出。"},{"criterion":"形式结果实际完整","method":"干净重建并检查admit/axiom/assume/disabled proof、solver/库依赖；有限模型/超时/unknown各自标识，不用“无反例”混成无界证明。","failure":"跳过失败义务、增加未批准公理；仅有限展开却声称全局活性。"},{"criterion":"反例与实现连接","method":"重放trace及输入，核对环境可达性；真实缺陷变成独立回归，伪反例记录需要加强的抽象。","failure":"反例只在不可能环境出现；为消除反例删真实需求。"},{"criterion":"变更保持目标和信任边界","method":"比较固定规格/代码diff，重新运行受影响证明/模型检查和实现测试；记录编译器/FFI/硬件连接的保证缺口。","failure":"证明旧代码，部署新代码；依赖版本变而缓存仍绿。"}],"assessment":{"recommendation":"固定规格/证明义务可整包委派生成与修复；从业务需求到系统规格先分阶段接收，按实际保证范围验收。","conditions":["需求与实现映射可审","验证工具和信任基础固定","独立缺陷/空洞证明控制"],"evidenceBasis":"AutoRocq在程序来源的641条SV-COMP证明义务中报告证明198条；分母是引理，不是完整程序。；DafnyBench评价模型给固定程序/规格补验证提示，并检查最终程序，而非自由修改规格求绿灯。","localStatus":"本站本任务AI运行数、成功率、实际成本及人工介入量均为null；代表例为交付设计，未执行。","gaps":["并发/硬件/大型规格没有本包直接AI完成率","证明不覆盖未建模需求或完整可信编译链"],"mode":"staged-delivery","rationale":"DafnyBench和AutoRocq提供程序专属证明证据，支持把细节搜索交给代理；它们并未证明任意生产系统规格正确，因此独立验收需求/模型连接而不只看证明通过。","frontier":"已公开从C验证义务到Rocq脚本及Dafny注解的代理路径；新仓库工具链与论文旧版本不同，需重放。","nextValidation":"用固定小型C/Dafny模块，先用vacuous requires、溢出和错误不变量校准验收；隐藏参考证明，检验完整义务、代码映射和一个修改后的维护。","futureSignals":["扩大：未见真实模块/变更中规格覆盖、反例映射与完整义务持续通过，维护成本可控。","收缩：只补局部引理而漏主目标、依赖大量assume或现实模型不符；缩到局部义务。"]},"workflow":{"before":"需求专家写规格，开发改代码，验证者补证明，变化后返工。","now":"代理可提出规格/不变量、探索反例、修改代码并响应验证器；可信需求/规格边界独立检查，固定义务无需逐步批准。","retained":["真实属性与环境假设确认","保证用于部署的接收责任"],"added":["空洞规格/未证义务检查","反例到代码的映射","证明与实现版本同源"]},"humanResponsibility":"需求/验证接收者确认期望属性及模型含义，代理可承担证明、反例和变更搜索；验证工具承担形式判断，部署负责人确认未建模条件。","failureHandling":"unknown/timeout保留具体义务与预算，不作unsafe或safe结论；反例先辨真伪，不通过弱化规格消除；环境缺失交可重放最小项目。","example":{"input":"设计例：固定32位数组求和函数，约定长度/元素范围和溢出行为；要证明结果与数学和一致并保持边界安全。","output":"代码—规格映射、溢出义务/不变量/证明或trace、锁定项目与变更回归；本例未运行。","failure":"只在无限整数模型证明求和，却对32位实现声称无溢出；加入false前提通过。"},"claims":[{"id":"M10-PROGRAM","statement":"AutoRocq在程序来源的641条SV-COMP证明义务中报告证明198条；分母是引理，不是完整程序。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-M-AUTOROCQ","locator":"§5.1/§6.1/§7","relation":"direct","supports":"程序验证专属生成证明与实际检查证据。","limit":"不等于198完整工业系统通过；未覆盖全部需求与编译/硬件链。"}],"progressIds":["P-M-PROGRAM-PROOFS"],"configuration":"arXiv v1；GPT4.1 temperature0；论文Rocq8.18/CoqPyt1.0.0、Frama-C WP提取；641义务含不溢出/功能/不变量，当前仓库9.0.0另记。","limitation":"不等于198完整工业系统通过；未覆盖全部需求与编译/硬件链。","evidenceStatus":"active"},{"id":"M10-ANNOTATION","statement":"DafnyBench评价模型给固定程序/规格补验证提示，并检查最终程序，而非自由修改规格求绿灯。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-M-DAFNY","locator":"dataset construction / evaluation","relation":"partial","supports":"固定规格下的程序验证注解任务证据。","limit":"有限程序、历史模型；不保证自动写对需求或整个工业系统。"}],"progressIds":["P-M-DAFNY"],"configuration":"arXiv v1；750余现有Dafny程序移除提示，原模型/反馈条件以论文为准；本文不转抄分数。","limitation":"有限程序、历史模型；不保证自动写对需求或整个工业系统。","evidenceStatus":"active"},{"id":"M10-ADOPT","statement":"接收必须包括需求/模型映射和完整义务，不能由局部证明通过推出现实系统安全。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-M-AUTOROCQ","locator":"§5生成义务；§7案例；limitations","relation":"adjacent","supports":"区分程序提取假设、引理与完整实现保证。","limit":"系统级规格、并发/硬件变体需新任务验证。"}],"progressIds":[],"configuration":null,"limitation":"系统级规格、并发/硬件变体需新任务验证。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开资料和工具入口已核对；论文配置、当前实现声明及本任务成绩分别记录，未运行本任务。","tools":[{"name":"AutoRocq","url":"https://github.com/NUS-Program-Verification/AutoRocq/tree/24b92683d049aaa9ecf0cd19e247535e762f43b9","version":"24b92683…；仓库Rocq9.0.0，论文8.18","access":"开源Python/opam/Docker、模型API；不把新工具链继承旧成绩。"},{"name":"DafnyBench","url":"https://github.com/sun-wendy/DafnyBench","version":"论文v1；代码执行前锁commit及Dafny/solver版本","access":"公开ground_truth/hints_removed/eval；未安装运行。"}]},"variants":[{"title":"固定局部证明义务","inputDifference":"规格与代码已可信，只缺证明/不变量。","assessmentChange":"可整包委派直到完整核验；接收范围就是该义务。","acceptanceChange":"防改规格/加假设/漏依赖，不能扩成全程序完成。"},{"title":"并发或硬件系统模型","inputDifference":"时序、公平性、缓存/通信/硬件抽象影响性质。","assessmentChange":"可交形式模型与反例候选，缺专属AI实测；分阶段接收模型。","acceptanceChange":"安全/活性、状态界与公平性/环境假设，trace到真实状态映射；软件算术证据不转移。"},{"title":"规格/代码升级","inputDifference":"依赖、接口和实现一起演进。","assessmentChange":"可委派证明维护及影响分析，先锁定不允许弱化的属性。","acceptanceChange":"重放受影响义务和负控制，区分新旧版本覆盖。"},{"title":"顺序软件规格与验证","inputDifference":"固定顺序软件模块、机器整数/内存语义和目标属性，不以并发/硬件抽象为前提。","assessmentChange":"可委派规格/证明/反例候选；需求对应先接收，已有可信规格下按完整义务接收。","acceptanceChange":"查需求—前后置/状态—代码对应、可达性/vacuity、完整义务与实现测试，边界/溢出及反例重放。"}],"sourceIds":["WU-M-AUTOROCQ","WU-M-DAFNY"],"evidenceLinks":[{"sourceId":"WU-M-AUTOROCQ","supports":"程序验证专属生成证明与实际检查证据。；区分程序提取假设、引理与完整实现保证。","notCovered":"不等于198完整工业系统通过；未覆盖全部需求与编译/硬件链。；系统级规格、并发/硬件变体需新任务验证。"},{"sourceId":"WU-M-DAFNY","supports":"固定规格下的程序验证注解任务证据。","notCovered":"有限程序、历史模型；不保证自动写对需求或整个工业系统。"}],"resourceRefs":[],"career":{"direction":"形式方法、程序验证、验证工程","evidence":"规格/实现映射、证明/反例、信任基础及维护差异","nextStep":"补目标语言/系统语义，展示一次空洞证明如何被拒。"},"team":{"bottleneck":"规格建模还是证明细节/维护耗时？","allocation":"代理做候选模型/证明与反例，领域接收者审属性和现实连接。","nextStep":"分开统计规格错误、义务失败和工具chain阻塞。"},"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-M11","domainId":"math","title":"密码安全论证包：威胁模型、协议与证明/攻击对照","goal":"对一个锁定密码构造/协议交付明确安全游戏或符号模型、可检查论证或攻击及精确适用范围。","trigger":"研究/工程团队需评估构造、协议变体或现有论证，避免自然语言安全主张无模型。","scope":"安全定义、攻击者、假设、协议模型、证明/反例为一包；代码执行和物理泄漏另按W-M12验收；非授权实际攻击不纳入。","inputs":[{"name":"方案与安全目标","detail":"构造/协议版本、角色/消息、密钥/nonce生成和复用条件、参数；安全游戏/攻击者能力/信道与腐化模型。"},{"name":"形式环境与先验","detail":"符号或计算模型、允许假设/理想化、目标EasyCrypt/Tamarin/CryptoVerif等版本、已知攻击/论文及允许改动区。"},{"name":"研究边界","detail":"目标证明/攻击、可用计算/搜索预算；完整安全模型由可信输入提供或分阶段核对，不从可编译反推。"}],"deliverables":["原方案—安全定义/模型—形式声明对应和假设清单","可重建完整形式开发/证明、攻击trace或条件性未解义务","安全游戏/辅助引理的依赖及差异，非平凡性/可达性检查","适用参数/攻击复杂性/理想化和代码落地缺口；来源及后续修订"],"acceptance":[{"criterion":"安全模型忠实","method":"独立核对角色、密钥、nonce、消息绑定和攻击者查询/腐化能力；用已知不安全变体作为负控制，验证模型能暴露攻击。","failure":"删去重放/腐化能力使协议“安全”；加理想信道而原实现没有。"},{"criterion":"完整形式结果不藏未证部分","method":"固定环境重放全部目标，核对admit/axiom/假设/solver与证明依赖；游戏跃迁和引理连接到最终目标，局部通过单列。","failure":"多个引理通过但最终安全界未证；把构造安全假设直接当最终定理。"},{"criterion":"攻击/反例在原模型内","method":"重放trace/代码，检查前提、复杂性、概率、参数和预算；区分符号攻击与可实现计算攻击。","failure":"攻击需要超出允许查询或不现实时间，却写破坏目标安全级别。"},{"criterion":"保证与部署不混淆","method":"核对报告只声称已证模型/参数；理论保证到代码/随机源/侧信道/密钥管理缺口必须可见。","failure":"IND-CPA证明被说成认证加密/抗侧信道；nonce重复时仍套原安全界。"}],"assessment":{"recommendation":"安全目标/模型可锁定时，可整包委派形式论证候选；模型生成、完整证明和实现安全分别接收。","conditions":["威胁模型与安全定义可核","工具和假设固定","有已知攻击/不安全变体控制"],"evidenceBasis":"CrypFormBench专门评估密码模型解释、生成、补全、转换和修正，语法可执行与安全语义需分开看。；ShannonProver v1在专家给定安全模型及引理级分解后自动构建EasyCrypt证明，含部署协议相关个案。","localStatus":"本站本任务AI运行数、成功率、实际成本及人工介入量均为null；代表例为交付设计，未执行。","gaps":["10月代码的项目级搜索超出v1已验证范围","无本例可调用工具的稳定性数据"],"mode":"staged-delivery","rationale":"密码专属评估和Shannon个案支持自动证明工程；它们都不能替代模型选择与部署条件。因此给代理完整论证目标，按语义/最终证明和有效攻击接收，保留未解出口。","frontier":"论文v1在专家给模型与引理分解下完成证明工程；10月固定源码进一步声明项目级分解搜索，属于可检实现入口，需另试验，不能冒称论文已验证更宽能力。","nextValidation":"固定一个公开协议或AEAD安全游戏，隐藏参考证明，对照已知nonce复用不安全变体；独立重建全部目标及模型对应，保存未解引理和总预算。","futureSignals":["扩大：未见协议/游戏分解下，完整安全目标在固定攻击者模型中通过，独立密码分析一致。","收缩：只会补语法，漏nonce/腐化假设、改变目标或跳过最终目标；缩到固定引理补全。"]},"workflow":{"before":"密码专家写安全模型/游戏，证明工程师手工分解和填tactic，改模型后再交接。","now":"代理可搜索中间游戏、引理和证明并响应checker；来源/安全定义与最终游戏连接独立核对，不固定所有密码任务用同一工具。","retained":["安全目标及攻击者能力确认","安全假设和最终部署采纳"],"added":["游戏/引理与最终目标版本同源","符号/计算/实现保证分开","已知攻击负控制"]},"humanResponsibility":"密码研究接收者确认模型、假设与可用范围；代理可探索模型候选、证明与反例，形式工具核查适用目标，发布者承担安全主张的范围责任。","failureHandling":"不能证明时列最小未解引理和已证范围；发现攻击保留可重放输入/trace；模型歧义先给相异解释及影响，不偷偷削弱攻击者。","example":{"input":"设计例：锁定使用唯一nonce的AEAD保密/完整性游戏，再对nonce可能复用的协议变体分析；消息格式、查询与腐化条件明定。","output":"安全模型/目标、完整证明或有效攻击trace、参数界和部署限制；没有本例运行。","failure":"证明只覆盖唯一nonce却声明支持重复nonce；完成辅助引理却省略最终游戏跃迁。"},"claims":[{"id":"M11-CRYPTO-EVAL","statement":"CrypFormBench专门评估密码模型解释、生成、补全、转换和修正，语法可执行与安全语义需分开看。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-M-CRYPFORM","locator":"§3任务；§5评估；§6语义错误","relation":"direct","supports":"密码形式分析专属的能力/失败证据。","limit":"不是生产代码、硬件泄漏或任意新密码安全性；历史模型/题集限定。"}],"progressIds":["P-M-CRYPFORM"],"configuration":"v1；9模型、7工具语言，符号与计算模型；任务类型及分母依论文分开，本文不把总体分当任务完成概率。","limitation":"不是生产代码、硬件泄漏或任意新密码安全性；历史模型/题集限定。","evidenceStatus":"active"},{"id":"M11-SHANNON-RESULT","statement":"ShannonProver v1在专家给定安全模型及引理级分解后自动构建EasyCrypt证明，含部署协议相关个案。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-M-SHANNON-PAPER","locator":"Abstract；evaluation/case studies；AppendixC","relation":"direct","supports":"明确的密码证明工程成果与输入边界。","limit":"不是独立设计完整安全模型，也不是代码/物理安全证明；部分污染控制材料不公开。"}],"progressIds":["P-M-SHANNON"],"configuration":"arXiv2607.02847v1；专家供模型/分解；EasyCrypt checker与多分支调度；不同任务预算/可获得材料按原文，不推本例成本。","limitation":"不是独立设计完整安全模型，也不是代码/物理安全证明；部分污染控制材料不公开。","evidenceStatus":"active"},{"id":"M11-SHANNON-ENTRY","statement":"10月固定源码提供游戏/辅助引理分解搜索与完整文件检查入口，比v1论文输入范围更宽的声明需另验证。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-M-SHANNON-CODE","locator":"What does Shannon automate / project-level guide","relation":"direct","supports":"当前实现可检查及运行入口。","limit":"源码方法声明不是该宽范围已测成功率。"}],"progressIds":[],"configuration":null,"limitation":"源码方法声明不是该宽范围已测成功率。","evidenceStatus":"active"},{"id":"M11-ADOPT","statement":"给出可核对安全目标后，委派完整论证候选并以模型、最终目标和攻击控制接收；代码部署另验。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-M-CRYPFORM","locator":"semantic alignment/error analysis","relation":"adjacent","supports":"模型语义失配决定验收边界。","limit":"本AEAD变体为设计例，未执行。"}],"progressIds":[],"configuration":null,"limitation":"本AEAD变体为设计例，未执行。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开资料和工具入口已核对；论文配置、当前实现声明及本任务成绩分别记录，未运行本任务。","tools":[{"name":"Shannon Prover","url":"https://github.com/SkyShannonProver/shannon-prover/tree/60701dbe37ee72075e60e655a076e282efe7485e","version":"60701dbe…；vendored EasyCrypt及locked Python","access":"macOS/Linux、opam/Python≥3.12/uv及模型key；严格live eval需Linux+bubblewrap，未运行。"},{"name":"CrypFormBench","url":"https://github.com/Secbrain/CrypFormBench/tree/b5540654533dbb7a491ece75e50c8da67defc604","version":"b5540654…；论文v1","access":"公开数据/实验及多语言工具，非统一即用协议安全产品。"}]},"variants":[{"title":"符号协议模型","inputDifference":"秘密/认证目标与Dolev-Yao式理想化。","assessmentChange":"可交模型/trace候选，不套计算安全证明。","acceptanceChange":"角色/消息绑定、可达性和攻击者能力，有限搜索界与无界保证分开。"},{"title":"固定计算安全引理","inputDifference":"安全游戏/假设已可信且只缺proof。","assessmentChange":"可整包委派固定引理，与M10程序或M01数学义务不同上下文。","acceptanceChange":"EasyCrypt完整检查、允许假设和最终目标覆盖；不以子引理通过当完整协议。"},{"title":"新构造/密码分析","inputDifference":"模型和构造都可变，可能研究新攻击。","assessmentChange":"探索候选与反例；需要独立文献/复杂性及原模型核对。","acceptanceChange":"假设/参数/复杂性/攻击成功定义，未解不等于安全。"},{"title":"既有构造的计算安全模型","inputDifference":"对已有原语/协议定义或核对计算安全游戏、查询/腐化能力与安全界，不要求先提出新构造或仅补局部引理。","assessmentChange":"可委派模型/游戏及论证候选，语义先核再查最终安全目标。","acceptanceChange":"假设/参数/攻击者查询、nonce/随机性和游戏跃迁逐项对应原方案；局部证明和完整安全界分开。"}],"sourceIds":["WU-M-CRYPFORM","WU-M-SHANNON-PAPER","WU-M-SHANNON-CODE"],"evidenceLinks":[{"sourceId":"WU-M-CRYPFORM","supports":"密码形式分析专属的能力/失败证据。；模型语义失配决定验收边界。","notCovered":"不是生产代码、硬件泄漏或任意新密码安全性；历史模型/题集限定。；本AEAD变体为设计例，未执行。"},{"sourceId":"WU-M-SHANNON-PAPER","supports":"明确的密码证明工程成果与输入边界。","notCovered":"不是独立设计完整安全模型，也不是代码/物理安全证明；部分污染控制材料不公开。"},{"sourceId":"WU-M-SHANNON-CODE","supports":"当前实现可检查及运行入口。","notCovered":"源码方法声明不是该宽范围已测成功率。"}],"resourceRefs":[],"career":{"direction":"密码研究、形式密码分析","evidence":"安全模型、完整目标/证明或有效攻击与参数/假设说明","nextStep":"补密码安全定义、工具语言和已知攻击背景。"},"team":{"bottleneck":"模型、游戏分解还是tactic成本限制研究？","allocation":"代理承担证明工程/候选攻击，接收者核对安全模型、完整覆盖和部署适用。","nextStep":"用一个公开案例完整重放，再选未泄漏引理/协议变体测迁移。"},"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-M12","domainId":"math","title":"密码实现交付：标准构件、密钥条件与目标平台验证","goal":"将已接受的密码方案接入指定系统，交付互操作、失败处理、密钥/nonce条件和实现安全证据。","trigger":"需要采用/迁移标准密码库、替换不安全用法或验证目标平台实现，方案选择已有依据。","scope":"聚焦标准构件集成/验证和迁移；新原语设计、理论安全证明另接收。测试向量与形式保证不等于任意硬件无泄漏，AI代码辅助属相邻证据迁移。","inputs":[{"name":"方案/系统契约","detail":"目标原语/协议/参数/版本和标准测试向量，兼容旧格式/对端，允许的库及平台/编译器。"},{"name":"密钥和随机性","detail":"key生成/存储/轮换/撤销、nonce唯一性/持久化/并发条件、随机源、错误处理与访问授权；敏感值不得写日志。"},{"name":"验证与迁移","detail":"目标设备/CPU指令/编译选项，现有实现和基线、互操作/回退要求、泄漏测量权限与可接受方法。"}],"deliverables":["使用审核构件的实现/集成补丁、依赖/编译锁定和保证覆盖说明","标准向量/差分/互操作/篡改失败测试与密钥/nonce生命周期检查","目标平台时序/泄漏诊断及其统计/覆盖限制，或明确尚未完成项","分批迁移、版本协商/兼容回退、审计/故障恢复与已知风险"],"acceptance":[{"criterion":"功能/互操作正确","method":"独立标准向量、边界/空消息/大消息、差分实现、对端互操作；篡改标签/重放/截断按契约拒绝。","failure":"能加解密互通却接受坏tag；自测同一错误算法彼此一致。"},{"criterion":"密钥/nonce条件实际维持","method":"在并发、进程重启、重试和轮换下检查唯一性与生命周期；失败不泄露明文/密钥，随机源故障可观察。","failure":"重启后计数nonce归零；错误路径写密钥日志或释放未认证明文。"},{"criterion":"理论/代码/平台证据分开","method":"核对所选库对应源码/编译产物和保证范围；目标平台做适用的静态/动态泄漏检查并记录统计条件，未测项不宣称无侧信道。","failure":"HACL*源码保证被外推到所有编译器/调用包装/硬件；一次无显著泄漏被当证明。"},{"criterion":"迁移和回退可运行","method":"授权测试环境演练旧新数据/对端、key版本、部分失败和回退；验证不能降级到不安全算法或混用格式。","failure":"回退丢失key标识导致旧数据不可读；协商降级重开漏洞。"}],"assessment":{"recommendation":"可委派使用已审标准构件的实现与验证候选；仅在功能、key/nonce条件、平台检查与迁移演练分别通过后接收。","conditions":["方案/库已有依据","目标平台及生命周期可检查","可执行独立向量/差分/迁移验证"],"evidenceBasis":"HACL*官方固定库说明有限代码的功能/内存/secret-independence保证；CrypFormBench只提供相邻形式模型能力证据。本任务尚无直接AI端到端实现、生命周期或平台泄漏评估，不把工具方法冒充成绩。","localStatus":"本站本任务AI运行数、成功率、实际成本及人工介入量均为null；代表例为交付设计，未执行。","gaps":["没有直接AI端到端实现安全评估，不能升格为已测可靠","高保证源码到目标编译物/设备及key系统的连接待核"],"mode":"staged-delivery","rationale":"HACL*等工具提供高保证构件，不能证明代理集成无误；密码形式分析只是相邻能力证据。可委派候选补丁和检查执行，关键采用由真实系统证据决定，而非代码像标准或模型声称安全。","frontier":"可用高保证库与形式密码代理是不同资产；本轮没有直接覆盖AI端到端生成、侧信道测试和生产迁移的试验，保留这一迁移缺口。","nextValidation":"在授权样板系统集成固定AEAD库，加入坏tag、并发重启nonce复用、key轮换和版本回退负控制；独立检查真实编译产物与平台泄漏范围。","futureSignals":["扩大：未见系统/平台下库保证连接清楚，生命周期/迁移/泄漏检查可独立复现。","收缩：自动生成原语、隐含nonce/随机假设或降级回退；缩到已审库包装与测试候选。"]},"workflow":{"before":"密码选择、开发接入、运维key管理和安全测试分别交接，保证容易在边界丢失。","now":"代理围绕目标系统组织标准库集成、向量/差分、生命周期和迁移实验；确定性向量直接执行，异常定位到调用/状态而非重造算法。","retained":["安全目标和key所有权","目标平台保证与正式迁移采纳"],"added":["库保证到调用包装映射","并发重启/轮换负控制","平台泄漏与协议安全分开"]},"humanResponsibility":"安全/系统接收者确认标准库选择、密钥职责和平台范围，代理可实现集成与检查；正式迁移由已有权限主体承担。","failureHandling":"向量不符先拒收功能；nonce或随机源条件未满足则拒收安全接入；无法做平台测量明确交付限制，不用理论证明补平台空缺。","example":{"input":"设计例：服务把存储加密迁到固定AEAD库，key版本显式保存，nonce跨并发/重启唯一；需旧数据可读与坏tag拒绝。","output":"集成补丁、向量/差分/坏tag/重启/轮换/回退记录及目标平台限制；没有本例运行。","failure":"本地往返测试通过，但重启重复nonce，或认证失败仍返回明文。"},"claims":[{"id":"M12-VERIFIED-LIB","statement":"HACL*固定代码提供限定算法的内存安全、功能正确和secret independence说明，生产集成入口另指向packages。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-M-HACL","locator":"README HACL* / ValeCrypt / Status","relation":"method","supports":"可复用高保证构件及保证范围。","limit":"不是AI成绩；只抵抗某些定时侧信道，包装/编译/硬件条件仍需核查。"}],"progressIds":[],"configuration":null,"limitation":"不是AI成绩；只抵抗某些定时侧信道，包装/编译/硬件条件仍需核查。","evidenceStatus":"active"},{"id":"M12-TRANSFER","statement":"可让代理整包产出标准库集成与验证材料，但没有直接证据把密码形式模型成绩转成端到端实现安全可靠性。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-M-CRYPFORM","locator":"§3任务定义及§6模型语义错误","relation":"adjacent","supports":"相邻模型/语义能力提示需要核对安全前提。","limit":"该来源不评AI实现、key生命周期、侧信道或迁移；推荐是待验证工程迁移。"}],"progressIds":[],"configuration":null,"limitation":"该来源不评AI实现、key生命周期、侧信道或迁移；推荐是待验证工程迁移。","evidenceStatus":"active"},{"id":"M12-ADOPT","statement":"接收依赖真实平台上独立功能、key/nonce和迁移控制，库形式保证不能单独覆盖集成。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-M-HACL","locator":"README verification scope / production packages","relation":"method","supports":"限定代码保证与集成条件分开。","limit":"本AEAD迁移为设计例，平台测量需另授权与提供环境。"}],"progressIds":[],"configuration":null,"limitation":"本AEAD迁移为设计例，平台测量需另授权与提供环境。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开资料和工具入口已核对；论文配置、当前实现声明及本任务成绩分别记录，未运行本任务。","tools":[{"name":"HACL*固定研究库","url":"https://github.com/hacl-star/hacl-star/tree/504c2987452f87fe44bce9b9f12e19d6e051761f","version":"504c2987…；F*/KaRaMeL链按项目锁定","access":"开源研究代码/生成C；实际生产发布/绑定查hacl-packages再固定版本，不使用浮动main冒称可复现。"},{"name":"生产packages入口","url":"https://github.com/cryspen/hacl-packages/","version":"官方研究库指向，具体release尚未为目标系统选定","access":"尚未构建/目标平台测试；工具可获取不代表任务通过。"}]},"variants":[{"title":"目标硬件/侧信道","inputDifference":"有功耗/电磁/时序泄漏威胁。","assessmentChange":"可委派工具配置、分析和候选修复；未有本例AI测量成绩，不能只给安全标签。","acceptanceChange":"实际设备/编译器、采集方法、统计功效/攻击验证和覆盖限制；secret independence不等于全物理泄漏。"},{"title":"后量子迁移","inputDifference":"参数/证书大小、握手、混合方案和兼容约束变化。","assessmentChange":"标准/库/协议版本先核定，再交候选实现与迁移实验，不自定义组合。","acceptanceChange":"跨对端、降级/回退、key格式/尺寸与资源消耗；理论PQC证明不能代替互操作。"},{"title":"标准构件常规集成","inputDifference":"在目标系统使用已接受的经典或既定标准库/协议；不要求后量子替换、尺寸变化或旧版迁移。","assessmentChange":"可交标准构件包装与独立向量/互操作/失败检查候选，安全接收仍查实际key/nonce条件。","acceptanceChange":"固定库/平台/接口，标准向量、差分、坏tag/截断、key/nonce生命周期及保证覆盖；同一错误实现的往返一致不够。"},{"title":"既有协议与密钥系统的版本迁移","inputDifference":"经典算法/协议/参数或key系统版本更新，需旧数据/对端兼容与可回退，不必是后量子方案。","assessmentChange":"可分阶段委派迁移候选及测试；先明确新旧保证、状态和回退允许范围。","acceptanceChange":"旧新数据/对端、key版本/轮换、部分失败及降级控制重放，回退不得恢复已知不安全模式。"}],"sourceIds":["WU-M-HACL","WU-M-CRYPFORM"],"evidenceLinks":[{"sourceId":"WU-M-HACL","supports":"可复用高保证构件及保证范围。；限定代码保证与集成条件分开。","notCovered":"不是AI成绩；只抵抗某些定时侧信道，包装/编译/硬件条件仍需核查。；本AEAD迁移为设计例，平台测量需另授权与提供环境。"},{"sourceId":"WU-M-CRYPFORM","supports":"相邻模型/语义能力提示需要核对安全前提。","notCovered":"该来源不评AI实现、key生命周期、侧信道或迁移；推荐是待验证工程迁移。"}],"resourceRefs":[],"career":{"direction":"密码工程、标准库集成、安全实现","evidence":"固定构件、生命周期及平台/迁移实际证据","nextStep":"补安全模型/低层实现/目标平台，不把向量通过当安全审计。"},"team":{"bottleneck":"安全构件选择还是生命周期/迁移出错？","allocation":"代理负责候选集成与检查，密码/系统人员接收key与平台条件。","nextStep":"记录每类失效及独立验证成本，先检最易破坏的key/nonce状态。"},"reviewedAt":"2026-10-20","evidenceStatus":"active"},{"id":"W-M13","domainId":"math","title":"教学效果交付：独立作答、延迟保持与迁移证据","goal":"判断一套数学教学/AI辅导方案对目标学生的学习有什么可支持效果，并给可重算研究与改进建议。","trigger":"课程试用了辅导材料/AI，需要区分练习完成更好与离开工具后真正掌握，避免只看满意度。","scope":"教学研究设计、实施记录、独立评价、保持/迁移及群体差异为一包；高校课程使用学校研究作为相邻依据，不据模型答题分数推学生学习。","inputs":[{"name":"课程/干预对照","detail":"目标学生、课程/先修、方案/工具/导师版本、分配/参与机制和对照条件，原材料与运行成本。"},{"name":"评价和时间","detail":"主要学习目标、无AI独立题、延迟时间和新情境迁移题；评分rubric、盲评与合法替代，先于看结果冻结。"},{"name":"实施/数据","detail":"授权学生数据、缺课/退出、实际使用/交互、班级聚类、导师差异、污染和测量版本；隐私处理按既定制度。"}],"deliverables":["预设效果研究/课程改进方案、分配/对照与测量说明","独立即时/延迟/迁移作答与盲评分，参与/缺失计数桥接","可重算意向性/适用分析、效应/区间、群体/使用差异和敏感性","内容与教学方式改进包，分别说明材料正确、使用体验、独立学习及证据限制"],"acceptance":[{"criterion":"测的是独立学习","method":"独立审题覆盖目标和新情境；无工具评价独立监测，延迟和迁移不复用已泄漏练习答案，盲评合法替代。","failure":"用AI可用时练习分充当无AI掌握；题库答案在干预中泄漏。"},{"criterion":"比较与估计可信","method":"独立核对分配/对照、基线与聚类、缺失/退出、污染和ITT；随机试验与观察关联分开，置信区间与不确定性完整。","failure":"只看活跃用户；群体基线不平衡未处理；小样本显著性当普遍提升。"},{"criterion":"组合干预归因有限","method":"检查实际AI用量、导师/平台练习与内容是否同变；若对照无法分离AI贡献，明确只支持组合效果。","failure":"Khan练习+AI比无练习好就宣称全归AI；导师接建议的研究写成无人辅导。"},{"criterion":"保持/迁移按实际范围解释","method":"核对评价时点、学生群体与课程层级；长期不等于两年内持续用工具，停止干预后的保持单列；未测就交下一轮可执行设计。","failure":"短期测试写终身效果；学校结果直接推高阶大学数学。"}],"assessment":{"recommendation":"可委派研究方案、评价材料、数据分析和改进候选；学习效果以独立学生数据接收，缺长期数据时交付下一轮测量设计。","conditions":["学生数据/干预授权明确","对照与主要学习结果可评","无AI/延迟/迁移测量可实施"],"evidenceBasis":"数学现场试验把工具辅助练习与无工具考试分开，普通生成式界面可能提高练习表现却损害独立学习。；两学年Khanmigo试验报告真实学校成绩与使用参与；效果类似无AI练习，不能把组合效果全归AI。","localStatus":"本站本任务AI运行数、成功率、实际成本及人工介入量均为null；代表例为交付设计，未执行。","gaps":["大学数学长期保持的直接AI证据未被本包覆盖","组合平台/导师效果不能简单归因AI"],"mode":"staged-delivery","rationale":"真实数学教学RCT证明辅助表现、独立考试和使用参与是不同对象；两学年试验仍未分离所有AI边际贡献。因此代理可组织完整研究与分析，效果结论受设计、真实使用和测量范围限制。","frontier":"已有课堂/真人导师及两学年学校试验；正负或有限效果与干预形式相关，不等同任意AI大学教学效果。","nextValidation":"为一大学数学单元预先设计对照和独立/延迟/迁移题，先验证评分与泄漏负控制；经既有教学授权采数后再估效果，不通过运行模型补学生证据。","futureSignals":["扩大：多学期/新群体独立保持与迁移可复现，组合因素可辨、退出和成本完整。","收缩：只有练习时分数/满意度好、无AI退步或对照无法分离；缩小到已证时间窗和干预。"]},"workflow":{"before":"教学工具用量、课堂成绩与老师评价各自汇报，缺独立学习和对照。","now":"代理可设计测量、检查干预执行/缺失并生成可重算分析；反馈针对具体误解与参与问题调整课程，下一轮按相同学习目标检验。","retained":["教育目标/研究伦理与学生支持","课程变化采纳责任"],"added":["独立无AI/延迟/迁移证据","实际参与和组合因素","评分/干预版本与失效解释"]},"humanResponsibility":"课程研究负责人确定可接受设计、学生支持和正式课程决策；代理可组织研究与计算，盲评和独立作答提供效果证据。","failureHandling":"无对照/缺失严重时交描述性分析与可行改进/补测设计，不强行因果；发现无AI表现变差则收窄工具使用方式并记录反面结果。","example":{"input":"设计例：大学微积分两种提示方案，主要结果为两周后无AI的新情境极值题；班级/导师影响和工具使用单列。","output":"预设评价/分配、独立作答与盲评、效应区间/缺失敏感性及课程改进；本例尚无学生数据。","failure":"用学生聊天次数或模型做题正确率宣布学习进步；只分析持续使用者。"},"claims":[{"id":"M13-UNASSISTED","statement":"数学现场试验把工具辅助练习与无工具考试分开，普通生成式界面可能提高练习表现却损害独立学习。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-M-TEACH-RCT","locator":"Design / Main Results / Limitations","relation":"direct","supports":"直接学习而非解题分数的因果研究。","limit":"高中特定课程/短期，不能普遍推大学或长期保持；数值不抄作本站效果。"}],"progressIds":["P-M-LEARNING"],"configuration":"PNAS2025作者稿；土耳其高中2023–24学期、近千学生、GPT4普通/约束辅导/对照；独立评分无AI考试；非长期追踪。","limitation":"高中特定课程/短期，不能普遍推大学或长期保持；数值不抄作本站效果。","evidenceStatus":"active"},{"id":"M13-LONGER","statement":"两学年Khanmigo试验报告真实学校成绩与使用参与；效果类似无AI练习，不能把组合效果全归AI。","basis":"reported-result","sourceLinks":[{"sourceId":"WU-M-KHAN2Y","locator":"Abstract / working paper35620","relation":"partial","supports":"较长学校现场与真实使用的反证/补充。","limit":"中学补救课程，持续使用期不等于停用后的保持，非大学效果。"}],"progressIds":["P-M-KHAN2Y"],"configuration":"2026-08工作论文；18所田纳西中学、两年集群随机，Khan Academy+Khanmigo补救练习；参与与ITT/使用分析分开。","limitation":"中学补救课程，持续使用期不等于停用后的保持，非大学效果。","evidenceStatus":"active"},{"id":"M13-ADOPT","statement":"学习效果须用独立、延迟和迁移结果及可解释对照接收；模型成绩只能辅助查内容。","basis":"editorial-inference","sourceLinks":[{"sourceId":"WU-M-TEACH-RCT","locator":"unassisted exam / limits","relation":"adjacent","supports":"把借助工具完成与学习区分。","limit":"大学/高阶课程及两周例是研究设计，未实测。"}],"progressIds":[],"configuration":null,"limitation":"大学/高阶课程及两周例是研究设计，未实测。","evidenceStatus":"active"},{"id":"M13-ENTRY","statement":"Khanmigo官方提供教师备课/评价材料与学习辅导入口，角色和地区/学校访问不同；可用性不等于研究中同版本或学习效果。","basis":"tool-method","sourceLinks":[{"sourceId":"WU-M-KHAN-ENTRY","locator":"For teachers / For learners / FAQ classroom access and eligibility","relation":"method","supports":"真实产品入口及访问前提。","limit":"未调用；官方功能说明不是大学内容正确或学习有效的成绩。"}],"progressIds":[],"configuration":null,"limitation":"使用时锁定实际版本/访问方式，不能继承学校试验效果。","evidenceStatus":"active"}],"availability":{"status":"mixed","detail":"公开资料和工具入口已核对；论文配置、当前实现声明及本任务成绩分别记录，未运行本任务。","tools":[{"name":"数学学习RCT材料","url":"https://hamsabastani.github.io/education_llm.pdf","version":"PNAS2025，实施GPT4研究配置","access":"论文/设计参考，不承诺原干预界面可复制。"},{"name":"Khanmigo长期现场研究","url":"https://www.nber.org/papers/w35620","version":"working paper35620 2026-08-17","access":"研究入口；具体产品/学校访问另核，未开通/购买服务。"},{"name":"Khanmigo官方产品","url":"https://www.khanmigo.ai/","version":"持续服务，2026-10-20读取；具体部署版使用时记录","access":"教师/学习者/学校入口不同，受地区/账户资格影响；课堂学生访问不能由教师个人注册自动授予，未开通服务。"}]},"variants":[{"title":"大学高阶课程","inputDifference":"内容、学生、自主证明和评估均不同于学校试验。","assessmentChange":"可委派研究/测量与分析，效果为显式迁移假设。","acceptanceChange":"高阶理解/证明迁移、合法多解及延迟无AI评价，不借学校效应值。"},{"title":"真人导师辅助","inputDifference":"AI只建议导师，并非直接教学生。","assessmentChange":"可以研究人机辅导组合，不能当无人教师。","acceptanceChange":"导师采纳/策略、学生独立结果和导师差异分别分析。"},{"title":"大学基础课程效果评价","inputDifference":"目标为大学基础概念/计算/初步论证，学生/课程不同于学校试验和高阶证明课。","assessmentChange":"可委派测量/对照和分析，效果需该课程独立学生数据，不能继承学校效应值。","acceptanceChange":"无AI独立、延迟保持与新情境迁移、基线/聚类/缺失/实际使用和盲评分；内容正确与学习效果分开。"}],"sourceIds":["WU-M-TEACH-RCT","WU-M-KHAN2Y","WU-M-KHAN-ENTRY"],"evidenceLinks":[{"sourceId":"WU-M-TEACH-RCT","supports":"直接学习而非解题分数的因果研究。；把借助工具完成与学习区分。","notCovered":"高中特定课程/短期，不能普遍推大学或长期保持；数值不抄作本站效果。；大学/高阶课程及两周例是研究设计，未实测。"},{"sourceId":"WU-M-KHAN2Y","supports":"较长学校现场与真实使用的反证/补充。","notCovered":"中学补救课程，持续使用期不等于停用后的保持，非大学效果。"},{"sourceId":"WU-M-KHAN-ENTRY","supports":"产品/角色访问入口。","notCovered":"不证明学习或大学课程效果。"}],"resourceRefs":[],"career":{"direction":"教学研究、课程评价与学生指导","evidence":"预设设计、独立评价、学生数据/评分及改进证据","nextStep":"补研究设计与教育测量，不用平台活跃度代替效果。"},"team":{"bottleneck":"效果未知还是测量只看工具内表现？","allocation":"代理做方案/数据和分析，教师组织真实独立评价并采纳课程改进。","nextStep":"先验证测量与反面结果处理，再评估长期投入。"},"reviewedAt":"2026-10-20","evidenceStatus":"active"}],"progress":[{"id":"P-C-ORGANIZATION","domainId":"computing","title":"代码产出增加后，工作转向组织、接收与监控Agent","kind":"field-observation","originalPublishedAt":"2026-10-16","revisionAt":null,"summary":"Anthropic内部观察显示，Agent做得更多会改变人的工作内容，增加上下文维护、请求分流与监控。","whatChanged":"变化集中在交付之间的组织：规格、可见上下文、互审和异常分流进入日常工作。","result":"研究给出内部工作和招聘材料及访谈，观察到Agent管理工作增加；这是一家公司采用后的变化。","configuration":"内部工作数据与访谈；采用非随机；代码量等产出代理指标不代表全部业务价值。","before":"给人拆分实现步骤，再集中审阅提交。","after":"可以委派完整修改；同时需要让任务状态和异常可见，避免所有Agent输出都请求逐条阅读。","action":"为一条维护流程记录待接收补丁、重复异常和复核耗时；让注意力集中在真实决策上。","limitation":"组织观察不能推出你的团队应减几人，也不证明所有复核必须永远由人完成。","workUnitIds":["W-C01","W-C02","W-C03"],"sourceIds":["WU-EMPLOYMENT"],"revisionNote":"此条未另记资料修订；按首次公开日阅读。","assessmentChangedAt":null,"evidenceStatus":"active"},{"id":"P-C-HARNESS","domainId":"computing","title":"模型不变，改进工具和检查流程也能改善交付","kind":"workflow","originalPublishedAt":"2026-10-09","revisionAt":null,"summary":"Harness²通过比较运行轨迹来修改上下文、工具和检查安排，展示运行流程本身可以优化。","whatChanged":"从只换模型或加提示，扩展到保存失败轨迹、试改流程并重新执行。","result":"指定Code配置的平均检查得分由73.4变为76.5；得分是每题检查项通过比例的汇总，不是完整项目通过率。","configuration":"Gemini 3.5 Flash＋OpenCode执行、Opus 4.8编辑；40个Code任务、三条随机顺序经验流、k=1。含确定性和LLM/VLM检查；k=1无最终多产物选择，编辑/候选另耗预算。","before":"失败后直接重试或换模型，缺少对漏读、漏执行和漏检查的定位。","after":"用轨迹定位瓶颈并试改流程，再按相同目标验收。","action":"先检查Agent是否真的读输入、运行测试和查看产物；记录改流程后的收益与全部成本。","limitation":"流内经验相关；论文未验证本站队列或当前仓库，不能据最终执行步数推总成本下降。","workUnitIds":["W-C01","W-C02","W-C03"],"sourceIds":["WU-HARNESS"],"revisionNote":"此条未另记资料修订；按首次公开日阅读。","assessmentChangedAt":null,"evidenceStatus":"active"},{"id":"P-C-TB4","domainId":"computing","title":"测量修订：资源和验收错误也会制造失败","kind":"measurement","originalPublishedAt":"2026-08-28","revisionAt":null,"summary":"Terminal-Bench 4.0修正环境、题目与验收，让测到的失败更接近任务本身。","whatChanged":"评估Agent前，先检查任务说明、运行资源和验收是否合理。","result":"本次调整资源条件、修订题目并移除部分任务；它改变测量口径，不代表模型突然获得新能力。","configuration":"Terminal-Bench 4.0；环境和题集变化按主版本处理，相关试次需重跑。","before":"直接比较不同环境的榜单，或把所有超时都当能力失败。","after":"版本化输入、环境和验收；正确解、关键缺陷和绕过控制分别检查。","action":"保留一个正确替代实现和几个已知缺陷；规则变更重评产物，环境变更重跑。","limitation":"不把旧新题集分数连接成能力曲线；该基准预算不是业务SLA。","workUnitIds":["W-C03","W-C02"],"sourceIds":["WU-TB4","WU-CONTINUOUS"],"revisionNote":"此条未另记资料修订；按首次公开日阅读。","assessmentChangedAt":null,"evidenceStatus":"active"},{"id":"P-C-APPS","domainId":"computing","title":"从生成界面到真正完成用户操作","kind":"workflow","originalPublishedAt":"2026-03-24","revisionAt":null,"summary":"长时间应用构建案例把规划、实现与操作评价连接，评价者要真正使用应用才能发现未接通流程。","whatChanged":"验收Agent用Playwright实际试玩，发现矩形填充只落起止点、实体删除条件等问题，再把缺陷交回修订；验收关注可操作行为。","result":"在“2D复古游戏制作工具”案例中，单Agent输出的实体不响应输入；完整运行系统的结果能移动角色并试玩，仍出现平台重叠和关卡大墙等问题。","configuration":"原文Opus 4.5案例：单Agent约20分钟/9美元，完整系统约6小时/200美元；规划器还把一句话请求扩为16项功能、10个迭代。不是同预算、同范围的因果对照。","before":"生成文件或漂亮截图就交付。","after":"实际操作核心流程，失败返回实现循环，交付行为证据。","action":"选一个真实操作，例如保存、重新打开和修改；把结果状态纳入验收。","limitation":"案例支持工作法，不提供所有应用的一般完成率。","workUnitIds":["W-C02"],"sourceIds":["WU-APPS"],"revisionNote":"此条未另记资料修订；按首次公开日阅读。","assessmentChangedAt":null,"evidenceStatus":"active"},{"id":"P-M-RELEASE","domainId":"math","title":"公开研究产物让AI数学能力可以逐项检查","kind":"capability","originalPublishedAt":"2026-10-06","revisionAt":"2026-10-15","summary":"OpenAI公开内部前沿模型产生的数学结果，并提供部分Lean形式化和修订机制。","whatChanged":"观察对象从竞赛答对多少，扩展到具体研究论证、证明文件和后续修订。","result":"有可查研究结果与部分形式化产物，并公布尝试和推理摘要等信息；不是只有榜单分数。","configuration":"内部前沿模型；不同结果的形式化与学术接收状态不同。10月15日更新主要为资助与合作安排。","before":"用题库成绩推断是否能参与研究。","after":"逐项查看命题、产物、依赖、修订与独立接收。","action":"选与你的问题相近的一份产物，先核对问题和假设，再决定尝试证明搜索或独立反例检查。","limitation":"公开不等于全部结果已获认可，内部系统能力也不保证当前公开产品可直接复现。","workUnitIds":["W-M01","W-M03","W-M04"],"sourceIds":["WU-MATH-RELEASE"],"revisionNote":"10月15日更新主要涉及资金承诺；数学研究产物的原始公开日仍为10月6日，未据此刷新能力进展。","assessmentChangedAt":null,"evidenceStatus":"active"},{"id":"P-M-NEXUS","domainId":"math","title":"证明搜索开始在开放问题上交付可检查结果","kind":"capability","originalPublishedAt":"2026-05-21","revisionAt":null,"summary":"Nexus研究配置把证明、反例及对象/参数探索接到形式反馈；公开的是成果集合。","whatChanged":"W-M01承接固定目标证明；W-M04承接可调对象/参数与题意核对的研究产物，不合并两者验收。","result":"论文报告353个Erdős问题中解决9个、492个OEIS猜想中证明44个，并公开Lean证明。","configuration":"arXiv v1；基础代理Gemini3.1Pro，完整代理加入进化搜索/AlphaProof；Erdős未成功则3000 episodes停止；成功后团队核对题意。","before":"人逐步找引理、写证明，模型主要提供建议。","after":"给定形式目标后，Agent可搜索并响应内核反馈；人检查题意与研究意义。","action":"从固定Erdős125公开文件校准题意和核验；新变体另立目标与预算，保存未解决结果。","limitation":"原论文问题集和成功选择有边界；事后九题比较不能当新问题概率；本站未重放，公开结果仓库不是Nexus API。","workUnitIds":["W-M01","W-M03","W-M04"],"sourceIds":["WU-NEXUS"],"revisionNote":"本样稿增加W-M04研究范围与固定成果定位；未改变2026-05-21事件身份，非新能力发布。","assessmentChangedAt":"2026-05-21","evidenceStatus":"active"},{"id":"P-M-ALETHEIA","domainId":"math","title":"找证明、找反例、返回未解决，成为同一研究循环","kind":"workflow","originalPublishedAt":"2026-02-11","revisionAt":null,"summary":"DeepMind展示Aletheia等研究协作：生成候选、检查、修订与检索相互连接。","whatChanged":"研究可同时追求证明和反驳，失败返回也能帮助缩小问题。","result":"公开Feng26案例由AI在无人介入条件下生成一篇计算算术几何结构常数（eigenweights）的论文；作者归为Level 2“达到可发表质量”，已投期刊，并非已获接收。","configuration":"Deep Think/Aletheia研究系统与专家协作；自然语言检查器；未逐篇复核所链研究。","before":"只要求模型证明既定猜想，容易保留确认偏差。","after":"并行尝试证明与反例，工具核验具体主张，再把有效部分交回研究。","action":"把定义、预算和允许工具写清；要求可复算反例或明确未解决范围。","limitation":"作者分级与公开产物支持具体研究贡献线索，不等于全部数学自主完成。本站未逐篇复验，已投稿不同于同行接收，自然语言核查也不同于形式证明内核通过。","workUnitIds":["W-M03","W-M01"],"sourceIds":["WU-ALETHEIA"],"revisionNote":"仅补齐日期语义，保留原研究事件和适用范围；无新运行。","assessmentChangedAt":"2026-02-11","evidenceStatus":"active"},{"id":"P-F-TEMPLATES","domainId":"finance","title":"报告、模型和备忘录可以按整包工作委派","kind":"workflow","originalPublishedAt":"2026-05-05","revisionAt":null,"summary":"官方金融Agent模板把资料读取、模型维护、对账和研究交付组织为有工具与数据接入的工作。","whatChanged":"原先分散在抄数、Excel更新、报告写作的步骤，可以围绕同一交付与接收标准循环。","result":"发布了包含模型维护、总账对账及研究等具体参考架构，交付覆盖多个文件与应用。","configuration":"产品/模板发布说明；需按本机构数据、规则、权限与接收流程适配；本站未运行这些模板。","before":"逐个派取数、改表、写稿任务，多次重复交接相同背景。","after":"给定报告、模型和目标，Agent更新资料与产物；冲突集中进入接收。","action":"先选一份工作簿更新或对账快照，用来源追溯、重算和异常清单验收。","limitation":"模板可用不等于公司流程已可靠，模型榜单或客户引语不能直接算净收益。","workUnitIds":["W-F01","W-F02","W-F03"],"sourceIds":["WU-FINANCE-AGENTS"],"revisionNote":"此条未另记资料修订；按首次公开日阅读。","assessmentChangedAt":null,"evidenceStatus":"active"},{"id":"P-F-FINCH","domainId":"finance","title":"财务验收从单个答案转向完整工作簿","kind":"measurement","originalPublishedAt":"2025-12-15","revisionAt":null,"summary":"FINCH的任务围绕跨表、跨文件的财务工作，验收关注整个请求有没有完成。","whatChanged":"不能只比较一个数或要求与参考表布局完全相同；公式、覆盖和非预期修改都影响接收。","result":"作者以人工检查完整产物作为质量参照，并校准自动评价；可接受等价公式和布局。","configuration":"FINCH v1的组合工作流与评价方法；不是截至今天所有模型能力排行。","before":"一个答案正确就算交付，或按单元格完全一致误拒合法实现。","after":"关键数值重算、依赖检查、范围差异和可读性共同验收。","action":"为工作簿保留更新前版本；同时验证新数字、公式仍可用和其它内容未被误改。","limitation":"未运行该基准；自动裁判也可能错，关键错数不能靠平均分补偿。","workUnitIds":["W-F01","W-F02"],"sourceIds":["WU-FINCH"],"revisionNote":"此条未另记资料修订；按首次公开日阅读。","assessmentChangedAt":null,"evidenceStatus":"active"},{"id":"P-A-APP","domainId":"analysis","title":"分析应用有了可见查询与可保存修正的参考实现","kind":"workflow","originalPublishedAt":"2026-09-10","revisionAt":null,"summary":"公开数据分析应用示例把数据发现、只读查询、结果解释与修正保存连接起来。","whatChanged":"用户不必只相信一段总结，可以检查查询，并决定哪些业务修正值得保留。","result":"有可读实现与工具流程，说明如何把分析Agent接入具体数据环境；不是分析能力试验成绩。","configuration":"Agents API数据分析应用参考；本站未安装或执行该示例。","before":"在聊天里问数，查询与口径修正散落在上下文。","after":"查询与结果可见；经确认的修正显式保存，后续分析引用明确口径。","action":"选一个指标报表，要求同时交查询、结果和口径；不要自动保存未经确认的推断。","limitation":"工程参考需适配数据库和权限，不证明任意分析都正确。","workUnitIds":["W-A01","W-A02"],"sourceIds":["WU-DATA-APP"],"revisionNote":"本轮仅补齐日期语义；保留原首发日、既有适用单元及来源身份，没有新执行或可比较能力提升。","assessmentChangedAt":null,"evidenceStatus":"active"},{"id":"P-A-INTERNAL","domainId":"analysis","title":"业务问题到报告，数据Agent可以自己循环查数与修正","kind":"field-observation","originalPublishedAt":"2026-01-29","revisionAt":null,"summary":"OpenAI内部数据Agent展示从发现数据、运行查询到发布报告的闭环，并强调表的生成逻辑与业务语义。","whatChanged":"人的指令可以聚焦分析目标；Agent自行查表、检验中间结果并修正查询。","result":"内部工程案例公开了上下文层、查询结果核验与报告流程；没有给出本网站任务的通过率。","configuration":"公司内部专用数据、权限和工作流；含参考查询与结果比较的评估方式。","before":"人反复找表、写SQL、发现连接错误，再回到聊天解释背景。","after":"Agent基于数据语义组织查询循环；分析者确认问题、口径和实际决策。","action":"建立关键指标的参考切片与查询，验收相同结果而非SQL字符串一致。","limitation":"案例不提供通用正确率，也不能从观察数据自动创造因果识别。","workUnitIds":["W-A01","W-A02","W-A03"],"sourceIds":["WU-DATA-AGENT"],"revisionNote":"本轮仅补齐日期语义；保留原首发日、既有适用单元及来源身份，没有新执行或可比较能力提升。","assessmentChangedAt":null,"evidenceStatus":"active"},{"id":"P-C-VALIDATOR-AUDIT","domainId":"computing","title":"软件评测审计使验收器本身进入核对范围","kind":"measurement","originalPublishedAt":"2026-07-08","revisionAt":null,"revisionNote":"本样稿新增一手审计入口；没有新模型运行。","assessmentChangedAt":null,"summary":"OpenAI审计指出SWE-Bench Pro的题意和测试可能影响输赢。","whatChanged":"接收不再只看测试绿灯，要核对独立行为与合法替代。","result":"厂商审计证据改变验收设计，不是本CSV后端已修复。","configuration":"OpenAI为参评厂商；原审计方法与任务范围见原文。","before":"默认任何测试失败都代表模型能力失败。","after":"先排除验收定义和环境错误，再解释产物输赢。","action":"给CSV验收加入漏行/越权负控制和合法替代实现。","limitation":"无可比能力提升；因此assessmentChangedAt为null，不凭新增阅读日期刷新能力。","workUnitIds":["W-C02"],"sourceIds":["WU-SWE-PRO-AUDIT"],"evidenceStatus":"active"},{"id":"P-F-FINCH-V5","domainId":"finance","title":"Finch v5分开完整财务工作流人评与自动评价","kind":"measurement","originalPublishedAt":"2025-12-15","revisionAt":"2026-04-15","revisionNote":"使用v5论文方法和结果；保留旧P-F-FINCH的v1历史范围，不覆盖旧记录。","assessmentChangedAt":null,"summary":"完整workflow结果与自动judge结果需要分别报告。","whatChanged":"支持W-F01接收整包文件，不能用自动judge替代关键数字和公式核对。","result":"GPT-5.1 Pro产品代理人评66/172、自动72/172；是历史workflow口径。","configuration":"v5；172 workflows/384 tasks；产品代理外部浏览开启，历史聊天关闭；不同API baseline不合并。","before":"容易以一个答案或自动评分代表财务交付。","after":"以完整请求为单位，核对实际文件及独立计算。","action":"为联动簿先保存来源、公式与非目标区域基线。","limitation":"不声称v1到v5模型能力提高，文档修订日不充当能力变化日；本例未测。","workUnitIds":["W-F01"],"sourceIds":["WU-FINCH-V5"],"evidenceStatus":"active"},{"id":"P-F-SPREADSHEET2","domainId":"finance","title":"业务表格测量区分单元格修改与整表正确","kind":"measurement","originalPublishedAt":"2026-06-29","revisionAt":null,"revisionNote":"本轮新增该任务族证据，未重跑基准。","assessmentChangedAt":null,"summary":"局部修改分高仍可能无法交付整表。","whatChanged":"W-F01验收增加重算、跨sheet联动及非目标改动检查。","result":"论文中Opus4.6在金融建模100项的目标单元格匹配比例任务均值为89.69%，整表34/100通过；总集321包含其他类别，图表另用rubric。","configuration":"arXiv v1；总集321任务，金融建模子类100任务；统一SWE-agent scaffold，最多50轮。89.69%为100项中每任务目标单元格计算值匹配比例的平均，34/100为整张输出工作簿所有单元格均匹配参考的任务比例；图表rubric另计。","before":"一个新增数字正确就接收模型。","after":"最终可重算工作簿和相关联动都要成立。","action":"加入驱动扰动后情景/图表一致性检查。","limitation":"无可比能力提升；不同口径分数差不能写成模型退步或提升，assessmentChangedAt为null。","workUnitIds":["W-F01"],"sourceIds":["WU-SPREADSHEETBENCH2"],"evidenceStatus":"active"},{"id":"P-F-UNDERWRITE","domainId":"finance","title":"核保评估开始检验企业准则、工具与多轮交互","kind":"capability","originalPublishedAt":"2026-01-31","revisionAt":null,"revisionNote":"本次收录论文v1；不是今日新增实验。","assessmentChangedAt":null,"summary":"UNDERWRITE把机构准则、数据库和模拟申请交互接入评测，使核保任务能按具体决策条件研究。","whatChanged":"从保险知识问答进入带私有规则和工具反馈的核保候选判断。","result":"论文报告专门任务环境中的结果和重复运行可靠性缺口；不把文中存在差异的榜单点值录作当前能力。","configuration":"13模型、300任务、最多50轮；模拟企业与用户，ReAct/LangGraph；原配置与当前部署不同。","before":"通用保险问答难以反映机构准则与实际工具依赖。","after":"可按准则适配、缺项、幻造产品和交互恢复检查候选判断。","action":"固定本机构准则与代表申请，用可追溯建议和例外作为验收对象。","limitation":"模拟环境不等于真实承保；论文模型时点和机构输入分布不可直接迁移。","workUnitIds":["W-F04"],"sourceIds":["WU-UNDERWRITE"],"evidenceStatus":"active"},{"id":"P-C-MOBILE-STATE","domainId":"computing","title":"AndroidWorld把移动任务验收接到真实模拟器状态","kind":"measurement","originalPublishedAt":"2024-05-23","revisionAt":null,"revisionNote":"首次公开材料；本次核查不是新的能力发布日期。","assessmentChangedAt":null,"summary":"动态参数/持久状态检查使手机操作可以按结果验收；它没有评估新app代码。","whatChanged":"动态参数/持久状态检查使手机操作可以按结果验收；它没有评估新app代码。","result":"AndroidWorld公开在Android模拟器运行的参数化app任务，结果可由持久状态核验。","configuration":"AndroidWorld原论文：116任务、20app、动态参数；公开参考环境Pixel6/API33；模型/预算按原论文，不转写为开发成绩。","before":"仅看截图/点击轨迹难判定手机任务是否真正完成。","after":"任务可随机实例化并检查终态，操作与成功评价得以分开；这不测新app编写。","action":"为移动功能准备断网/后台/重进控制，核对实际存储与服务状态；另用目标真机补生命周期/外设。","limitation":"既有Android应用操作，不是新app开发、iOS或真机资源测试。","workUnitIds":["W-C05"],"sourceIds":["WU-C-ANDROID","WU-C-ANDROID-REPO"],"evidenceStatus":"active"},{"id":"P-C-ML-TRAIN","domainId":"computing","title":"MLE-bench公开训练与实验任务资产","kind":"workflow","originalPublishedAt":"2024-10-09","revisionAt":null,"revisionNote":"首次公开材料；本次核查不是新的能力发布日期。","assessmentChangedAt":null,"summary":"离线Kaggle工程任务有直接模型实验与公开运行资产。","whatChanged":"离线Kaggle工程任务有直接模型实验与公开运行资产。","result":"MLE-bench发布75个Kaggle工程任务及评分/运行资产，并实测Agent的数据准备、训练和实验搜索。","configuration":"原论文：75竞赛，包含o1-preview+AIDE等模型/脚手架；本文不混用榜单分数，当前仓库暂停新提交。","before":"代码生成题难以衡量数据准备、训练资源和最终模型效果。","after":"离线ML工程可以按任务预算与竞赛结果评价；不覆盖线上服务或业务收益。","action":"冻结业务保留集/泄漏控制与计算预算，先对简单基线，再接收完整训练/失败账本。","limitation":"离线竞赛与预算特定；模型训练得分不继承为服务/收益。","workUnitIds":["W-C06"],"sourceIds":["WU-C-MLE","WU-C-MLE-REPO"],"evidenceStatus":"active"},{"id":"P-C-DBA-STATE","domainId":"computing","title":"DBA-Bench以活跃负载区分诊断、恢复和约束","kind":"measurement","originalPublishedAt":"2026-07-24","revisionAt":"2026-10-10","revisionNote":"文档版本/结果表述修订；没有据此推定新的模型能力跃迁。","assessmentChangedAt":null,"summary":"v2在活跃PostgreSQL/TiDB比较有界/持续运维状态，揭示诊断与恢复及安全的差距。","whatChanged":"v2在活跃PostgreSQL/TiDB比较有界/持续运维状态，揭示诊断与恢复及安全的差距。","result":"DBA-Bench v2在活跃PostgreSQL/TiDB检验231有界与10持续场景，观察到诊断、恢复和约束满足的差距。","configuration":"v2：PostgreSQL/TiDB、活跃负载；231有界/10持续场景，8自动基线；具体模型/预算与场景约束依原表，不推算目标库成功率。","before":"根因或单步成功无法说明数据库业务已恢复且安全。","after":"评价扩到有负载状态和连续事件，能发现恢复后退化和违规动作。","action":"在目标DB冻结事务/权限与负载探针；恢复后再施加一次事件，检查先前修复仍成立。","limitation":"预印本受限场景；具体库、权限与负载仍需独立状态验收。","workUnitIds":["W-C08","W-C16","W-C18","W-C28"],"sourceIds":["WU-C-DBA"],"evidenceStatus":"active"},{"id":"P-C-EMBED-SIM","domainId":"computing","title":"EmbedAgent公开固件、电路和跨平台任务","kind":"workflow","originalPublishedAt":"2025-04-19","revisionAt":"2026-01-23","revisionNote":"文档版本/结果表述修订；没有据此推定新的模型能力跃迁。","assessmentChangedAt":null,"summary":"原研究与实现把代码和虚拟电路行为联合检查，不同平台成绩不能互换。","whatChanged":"原研究与实现把代码和虚拟电路行为联合检查，不同平台成绩不能互换。","result":"EmbedAgent公开固件、电路与迁移任务的虚拟电路验收，原实验显示平台/语言条件影响完成表现。","configuration":"论文v3：126案例、9组件、3平台、10LLM；有/无原理图、迁移与检索/编译反馈设置分开，本文不转用比例。","before":"编译或通用编程题无法检查电路连接和设备行为。","after":"代码、原理图与跨平台迁移被放入同一可执行评价环境；仿真仍不同于实物。","action":"选定目标板卡/接线和SDK，让Agent交候选；用外部刺激/时序及升级中断控制接收实物结果。","limitation":"虚拟电路与小板卡不能证明工业实物、实时或电气可靠性。","workUnitIds":["W-C10"],"sourceIds":["WU-C-EMBED","WU-C-EMBED-REPO"],"evidenceStatus":"active"},{"id":"P-C-PHISH-RCT","domainId":"computing","title":"钓鱼队列随机试验报告人加Agent研判改善","kind":"measurement","originalPublishedAt":"2025-11-17","revisionAt":null,"revisionNote":"首次公开材料；本次核查不是新的能力发布日期。","assessmentChangedAt":null,"summary":"研究报告队列优先级/解释改变分析师注意力与研判表现。","whatChanged":"研究报告队列优先级/解释改变分析师注意力与研判表现。","result":"问卷中排序/解释改变专业分析师的邮件研判与注意力；原语料Agent零FP/FN后合成错误用于行为研究，resolve-benign收益属于组合反事实而非持续邮箱处置观测。","configuration":"论文v1 §2，Qualtrics标准化只读邮件材料；无AI随机队列、知情AI队列、不知情AI队列三组；含合成判定/解释错误及resolve-benign反事实。招募/剔除/队列条件详见C30-PHISH，不是现场邮箱响应试验。","before":"随机队列与材料研判难区分排序、解释及跳过良性协议的作用。","after":"三组干预和合成判定错误可区分这些条件；跳过良性的协议只作反事实，不构成无人遏制结果。","action":"用本地真假/待证与自然错误分布验证排序/解释，单列自动关单的漏报风险；不照搬问卷反事实。","limitation":"167招募、标准化只读材料/Qualtrics，25封队列源自93语料；有合成错误与反事实，非生产长期无人响应。","workUnitIds":["W-C30"],"sourceIds":["WU-C-PHISH"],"evidenceStatus":"active"},{"id":"P-C-IT-RCT","domainId":"computing","title":"IT管理员RCT覆盖登录、策略和设备排障","kind":"measurement","originalPublishedAt":"2024-11-01","revisionAt":"2024-11-12","revisionNote":"文档版本/结果表述修订；没有据此推定新的模型能力跃迁。","assessmentChangedAt":null,"summary":"试验比较Entra/Intune中Copilot辅助管理员与对照；v2修订耗时数字。","whatChanged":"试验比较Entra/Intune中Copilot辅助管理员与对照；v2修订耗时数字。","result":"IT管理员RCT在Entra/Intune的登录、策略与设备排障比较辅助准确性/耗时；v2修正了耗时数字。","configuration":"论文v2：Entra/Intune有/无Security Copilot随机组，登录/设备策略/设备排障三情景；题型分开，未将修订当新能力。","before":"管理员需分别收集身份、策略和设备信息以回答/排查问题。","after":"Copilot可在这些管理情景辅助信息综合；实验答题改善不等于实际设备已恢复。","action":"选真实测试设备/账户，核对原用户任务、权限/数据和重启结果；将诊断、操作与恢复分开记录。","limitation":"辅助管理员任务，不是无人终端配置/维修试验。","workUnitIds":["W-C13","W-C23"],"sourceIds":["WU-C-ITRCT"],"evidenceStatus":"active"},{"id":"P-C-LONG-STATE","domainId":"computing","title":"OSWorld2.0挑战跨工具长流程和隐藏状态","kind":"measurement","originalPublishedAt":"2026-06-28","revisionAt":"2026-07-13","revisionNote":"文档版本/结果表述修订；没有据此推定新的模型能力跃迁。","assessmentChangedAt":null,"summary":"108个长流程把二元完成与部分分分开，动态约束/状态成为失败来源。","whatChanged":"108个长流程把二元完成与部分分分开，动态约束/状态成为失败来源。","result":"OSWorld2.0发布108个长流程，分开二元完成与部分评分，实测暴露动态信息、约束和隐藏状态处理失败。","configuration":"论文v2：108流程、500步主评价；完整/部分分分开；原实验模型/思考与批调用设置按论文，本文不混合成绩。","before":"短桌面任务较少体现跨来源、长时约束与中途变化。","after":"长流程把局部步骤和最终完成分开，能检查多步任务中遗漏状态/验证的问题。","action":"项目状态从工单/产物核实；注入过期记录与依赖变化，检查重排及未确认承诺，不用文档完成度验收。","limitation":"跨应用一般任务，不能把部分评分或总分转成项目管理率。","workUnitIds":["W-C22"],"sourceIds":["WU-C-OS2"],"evidenceStatus":"active"},{"id":"P-C-REPRO-RESEARCH","domainId":"computing","title":"PaperBench把论文复现拆为可检查贡献","kind":"workflow","originalPublishedAt":"2025-04-02","revisionAt":"2025-04-07","revisionNote":"文档版本/结果表述修订；没有据此推定新的模型能力跃迁。","assessmentChangedAt":null,"summary":"要求从论文到代码并执行实验，部分贡献与完整复现需分开。","whatChanged":"要求从论文到代码并执行实验，部分贡献与完整复现需分开。","result":"PaperBench提供20篇论文的贡献rubric和实际复现实验，原模型结果是部分贡献得分，不能视为整篇成功率。","configuration":"论文v3：20篇ICML2024论文、8316子项，作者参与rubric与独立judge校准；原最佳Claude3.5 Sonnet(New)+开放脚手架，本文不转用得分。","before":"写出代码或类似论文文本难以判断关键实验是否复现。","after":"贡献被连到可检查实现和实际运行，复现差异可以逐项暴露；LLM裁判仍有误差。","action":"选一项可预算贡献冻结rubric，独立重跑核心实验/消融，记录未复现项、预算和差异原因。","limitation":"贡献分与LLM裁判均有限，不是完整论文复现率。","workUnitIds":["W-C25"],"sourceIds":["WU-C-PAPER"],"evidenceStatus":"active"},{"id":"P-C-ALGORITHM","domainId":"computing","title":"AlphaEvolve公开算法搜索与部分成果","kind":"workflow","originalPublishedAt":"2025-05-14","revisionAt":"2025-06-16","revisionNote":"5月官方首次公布成果；6月白皮书补方法/产物说明，不重置能力首次公开日。","assessmentChangedAt":null,"summary":"专用系统以代码搜索和任务评价器得到算法/计算优化成果。","whatChanged":"专用系统以代码搜索和任务评价器得到算法/计算优化成果。","result":"AlphaEvolve报告演化代码搜索在算法/计算基础设施得到改进，并公开部分结果供检查。","configuration":"2025白皮书专用AlphaEvolve，LLM代码修改+演化+任务评价器；单成果不带未报告成功分母，本文不猜算力/重试。","before":"把LLM写代码与算法评价分开，难以持续利用反馈搜索新候选。","after":"专用系统把代码变异、评价与演化连接，可产出可检查算法；公开成果不等于整搜索系统可得。","action":"冻结算法语义、独立评价器和预算，先校准错误/合法替代，再搜索候选；推广与复杂性论证单独检查。","limitation":"专用研究系统与选择性成果，不保证公开Agent复现或解决任意新题。","workUnitIds":["W-C14"],"sourceIds":["WU-C-EVOLVE","WU-C-EVOLVE-ART","WU-C-EVOLVE-ANNOUNCE"],"evidenceStatus":"active"},{"id":"P-C-VISUAL-REPAIR","domainId":"computing","title":"SWE-bench Multimodal把视觉JS缺陷单列评估","kind":"measurement","originalPublishedAt":"2024-10-04","revisionAt":null,"revisionNote":"原论文首次公开；本次补录不刷新能力日期。","assessmentChangedAt":null,"summary":"617个视觉JavaScript缺陷任务把视觉和跨语言泛化问题显式化。","whatChanged":"加入带图像的用户界面软件缺陷，而非继续只测文本Python仓库。","result":"原研究观察到通用修复系统在视觉JS任务上退化，视觉/语言分布需单独评价。","configuration":"论文v1：617任务、17JS库，SWE-agent等当时配置；不是2026模型排行或本站运行。","before":"Python文本问题基准难覆盖图像/渲染与视觉交互缺陷。","after":"视觉JS问题及其检查成为独立任务入口，不能把原榜单直接外推到前端。","action":"前端变更用真实浏览器重放操作与渲染，独立查持久状态、焦点和失败态；不用截图相似替代功能。","limitation":"历史模型与库级缺陷；不覆盖所有完整页面、可访问性或跨浏览器产品交付。","workUnitIds":["W-C04"],"sourceIds":["WU-C-VISUAL"],"evidenceStatus":"active"},{"id":"P-FIN-FIRST","domainId":"finance","title":"财务研究验收可以同时查答案与证据","kind":"measurement","originalPublishedAt":"2026-09-21","revisionAt":null,"revisionNote":"原始公开身份保留；本站2026-10-20核查不算新增能力事件。","assessmentChangedAt":null,"summary":"FinFIRST把金融检索的最终答案与出处、期间、定义和计算一起验收，原子高分与完整题通过仍有差距。","whatChanged":"代理需要自己找资料、对齐主体和期间、计算并给可追踪证据；验收能定位答案正确但出处或口径不成立的情况。","result":"作者在123题统一工具评估中报告：GPT-5.6-Sol有88题全部标准满足（71.54%）；另一配置Claude-Opus-5的87.59%是701个原子标准的得分，二者不是同一分母。","configuration":"FinFIRST v1：123题、701原子标准、15配置；统一ReAct搜索/网页/Python，温度1.0，每配置一次指定运行，GLM-5.1裁判；GPT-5.6-Sol的Strict Pass为88/123。","before":"只核最终数字的财务问答，难以区分真实研究与碰巧命中答案。","after":"同一次任务同时检查取数、来源、计算和答案；完整接收要求所有必需标准满足。","action":"选一项有明确截止日的公司指标研究，要求原始披露、口径桥和重算；另设主体/期间/地域错配反例校准验收。","limitation":"作者指定运行与自动裁判结果，非独立复现；论文Table2示例混用美加收入与美国市场分母，需独立审查参考口径，不能当本业务正确率。","workUnitIds":["W-F03"],"sourceIds":["WU-FIN-FINFIRST"],"evidenceStatus":"active"},{"id":"P-FIN-SCENARIOS","domainId":"finance","title":"FinGAIA展示金融取材与计算子能力，角色名需核对实题","kind":"measurement","originalPublishedAt":"2025-07-23","revisionAt":null,"revisionNote":"原始公开身份保留；本站2026-10-20核查不算新增能力事件。","assessmentChangedAt":null,"summary":"金融角色标签下的公开例包含财报取数、新闻识别、回撤计算和基金持仓查找；不能仅按标签当作专业决策测试。","whatChanged":"多模态资料、网页与计算可以由同一Agent协调；评价其工作适用性时必须检查实际输入和终点，而非继承角色名称。","result":"公开例Fig6核对财报事实，Fig7识别新闻音频中的公司，Fig8计算历史回撤，Fig9含图像计数、基金排序和持仓查找；可参考这些子能力，不能推授信、欺诈检测或配置优化。","configuration":"FinGAIA v1，407题/10Agent的历史零样本网页/附件环境；实际样例见Appendix A。角色分类不作专业能力映射，正文及附表分母不同，不引用统一职业完成率。","before":"只给固定上下文的知识或计算题，无法检查主动取材与专业问题组织。","after":"可把取材/历史计算用作局部参照；完整职业交付须另有对应的专业案例和最终状态验收。","action":"用真实职责对应的脱敏信用或风险案例，独立核对事实、计算和反证；审批、调查行动及长期结果另设状态验证。","limitation":"旧网页/附件配置；实际样例与角色名存在差距，正文与附表口径不一致，不作为本业务专业成功率。","workUnitIds":["W-F10","W-F16","W-F21"],"sourceIds":["WU-FIN-FINGAIA"],"evidenceStatus":"active"},{"id":"P-FIN-TAX-ARTIFACT","domainId":"finance","title":"复杂税表有可检查模型产物，验收须看整份","kind":"measurement","originalPublishedAt":"2025-07-22","revisionAt":"2026-10-19","revisionNote":"项目原论文发表于2025-07-22；当前固定仓库快照为2026-10-19。v2注明2026-06发布但未列日，不能把快照日冒称新能力首次公开。","assessmentChangedAt":null,"summary":"TaxCalcBench固定TY25案例保留模型生成税表、逐行判定和整份严格结果，便于检查真实产物。","whatChanged":"税务Agent的候选输出可沿完整税表逐行追溯，不只展示回答或局部算式；整份通过与部分字段正确被分开。","result":"固定快照中ty25-ca-001的一份gpt-6.1-sol high无web结果报告整份严格通过，并保存115.41秒和$0.093068估算生成成本；它是一份选中成功产物。","configuration":"ty25-ca-001，gpt-6.1-sol high无web；报告严格通过、115.41s、$0.093068(litellm_estimate)，仅保存完成次成本，未本站复跑。","before":"只看答案或若干行数值，难以发现资格、跨表和州/联邦衔接差异。","after":"保存输入、输出及逐行评分后，可以核对整份候选是否一致，再单独验证申报回执。","action":"针对同法域同税年的新案例加入资格边界和跨表差错；隐藏参考输出盲验整份，再记录全部重试、人时和提交状态。","limitation":"项目版本和保存产物核查；按最佳thinking展示的单次保存结果不支持一般稳定性或端到端申报完成率。","workUnitIds":["W-F12"],"sourceIds":["WU-FIN-TAXCALCBENCH","WU-FIN-TAX-SAVED-CASE"],"evidenceStatus":"active"},{"id":"P-FIN-ACT-PRACTICE","domainId":"finance","title":"精算从知识题扩大到可执行表格与R实作","kind":"measurement","originalPublishedAt":"2026-07-27","revisionAt":null,"revisionNote":"原始公开身份保留；本站2026-10-20核查不算新增能力事件。","assessmentChangedAt":null,"summary":"INS-ActBench把精算评估扩展到表格和R代码的执行结果，可直接检查计算产物。","whatChanged":"模型需要形成可执行表格或代码并得到指定输出；精算知识题、实作题和完整业务模型分别评价。","result":"INS-ActBench Table3的GPT-5.5在636道表格题平均67.7%、516道R题输出评分71.1%。表格按答案格均分，R以执行输出核对；研究还观察到工具未启动、函数/引用错误和模型计算错误，三者不能靠代码可运行一并排除。","configuration":"v1 Table3/§4.1–4.4；9模型，Practice零样本；此行为GPT-5.5，表格636/R516题，数值相对误差0.2%，表格多答案格取均值；原题/工具配置结果，未本站重放。","before":"仅用知识题或改编案例选择题，不能证明能构建和运行数值产物。","after":"可以直接检查表格/代码输出；仍需用经验、定价、负债或资本任务各自的假设与校准接收整包。","action":"选定一种精算交付，冻结数据与假设，独立重算关键现金流或准备金；加入尾部、暴露和时间错位输入，记录整包接收而非仅格均分。","limitation":"改编考试题和指定输出；不涵盖长期准备金充分性或各地监管。","workUnitIds":["W-F20"],"sourceIds":["WU-FIN-INS-ACTBENCH"],"evidenceStatus":"active"},{"id":"P-FIN-ACCOUNTING-PRODUCT","domainId":"finance","title":"企业财务Agent已进入产品，现场接收仍需独立核对","kind":"workflow","originalPublishedAt":"2025-07-22","revisionAt":null,"revisionNote":"原始公开身份保留；本站2026-10-20核查不算新增能力事件。","assessmentChangedAt":null,"summary":"Intuit Enterprise Suite发布会计及财务Agent，覆盖交易分类、对账及财务分析等产品工作流。","whatChanged":"分类、对账和分析可通过产品中的Agent接到既有业务数据；用户需核对可用订阅、地区、连接和实际操作权限。","result":"厂商发布了嵌入企业财务系统的Agent产品入口；这是产品范围自报，未提供本账簿的完整正确率、重试和人工介入分母。","configuration":"Intuit 2025-07-22 Enterprise Suite原始产品发布；核查产品覆盖声明，未注册、购买或运行本站账套；客户节时比例不转写为任务完成率。","before":"需自行把资料、计算与财务系统连接起来组织自动化。","after":"存在厂商集成入口，可作为会计与预测工具候选；收款Agent不能直接充当资金部门对外付款能力。","action":"在适用产品沙盒中给一套带跨期、重复及缺失凭证的账簿，独立核对分类、银行/子总账差异和更正回写；保存每次人工介入。","limitation":"厂商发布而非独立成功率；地区订阅、连接和使用条件须向实际产品核对。","workUnitIds":["W-F02","W-F06","W-F07","W-F09"],"sourceIds":["WU-FIN-INTUIT-AGENT"],"evidenceStatus":"active"},{"id":"P-FIN-INSURANCE-AGENT","domainId":"finance","title":"保险Agent已有业务接入，产品准入仍有限","kind":"field-observation","originalPublishedAt":"2026-08-03","revisionAt":null,"revisionNote":"部署和产品说明，后续资料为范围补充；不据网站修改推能力跃迁。","assessmentChangedAt":null,"summary":"Guidewire发布保单变更Agent及核心流程接入，但该功能Early Access，未公布自动核保决定可靠性。","whatChanged":"保单变更Agent可接到核心保险流程，交付对象开始包含业务状态；Early Access范围不等于完整自动核保。","result":"Guidewire发布保单变更Agent及核心流程接入，但该功能Early Access，未公布自动核保决定可靠性。","configuration":"厂商产品/部署自报；具体模型、预算和错误分母未公开；不是本站或独立复现。","before":"资料、判断与状态分开交接。","after":"可围绕最终业务状态组织Agent与专业系统；真实权限、人工环节和失败必须仍可核对。","action":"沙盒准备标准接受、拒绝、超权、并发报价和再保未确认五类案件，独立核对理由、条件及最终状态；记录重试与介入。","limitation":"再保真实确认与复杂承保判断尚缺直接证据；厂商自报不是独立可靠性评估。","workUnitIds":["W-F22"],"sourceIds":["WU-FIN-GW-AGENT"],"evidenceStatus":"active"},{"id":"P-FIN-CLAIMS-LIFECYCLE","domainId":"finance","title":"低复杂赔案已有Agent处理到支付决定前的生产自报","kind":"field-observation","originalPublishedAt":"2025-11-03","revisionAt":"2026-03-18","revisionNote":"部署和产品说明，后续资料为范围补充；不据网站修改推能力跃迁。","assessmentChangedAt":null,"summary":"Allianz自报澳洲低额食品变质Agent处理到最终支付决策前，仍有人工决定；德国宠物自动化是另一系统，不等于同一Agent自动付款。","whatChanged":"低额食品变质赔案已有Agent处理流程的生产自报，但最终支付决策与人工环节仍需逐案核对，不能由处理文字推定自动付款。","result":"Allianz自报澳洲低额食品变质Agent处理到最终支付决策前，仍有人工决定；德国宠物自动化是另一系统，不等于同一Agent自动付款。","configuration":"厂商产品/部署自报；具体模型、预算和错误分母未公开；不是本站或独立复现。","before":"资料、判断与状态分开交接。","after":"可围绕最终业务状态组织Agent与专业系统；真实权限、人工环节和失败必须仍可核对。","action":"用两轮损失估计、部分已付、超权准备金、支付失败和新证据重开组成脱敏沙盒，独立重算估计/调整并核对批准、支付及关闭状态。","limitation":"个案持续估计准确性、长期结果和跨险种迁移尚缺直接证据；厂商自报不是独立可靠性评估。","workUnitIds":["W-F23"],"sourceIds":["WU-FIN-ALLIANZ-NEMO","WU-FIN-ALLIANZ-2026"],"evidenceStatus":"active"},{"id":"P-FIN-EXPERIENCE","domainId":"finance","title":"赔案文本到精算变量已有可检查研究原型","kind":"measurement","originalPublishedAt":"2026-06-04","revisionAt":null,"revisionNote":"首次公开研究，未重放；不以本次资料核对刷新能力日期。","assessmentChangedAt":null,"summary":"研究原型把赔案文本转换成有时间与来源的精算变量，再连接分层分析；经验基础不必局限于结构化表。","whatChanged":"可把文本变量候选作为独立经验交付接收，增设原文和时间盲审。","result":"作者报告合成样本的核心变量专家量表评阅及下游示范；不是实赔或全险种准确率。","configuration":"GPT-4o-mini原型，合成FHIR与文本，双临床评阅量表；不报告量表为百分比。","before":"经验分析多依结构化数据。","after":"可试作带原文及时间回链的文本变量候选，先核对再用于模型。","action":"固定一批去标识赔案和保单暴露，注入缺失、时间错位与未结案，双盲核对变量后用冻结假设回看。","limitation":"合成数据和专家量表不是实赔准确率；分层准备金示范不证明真实准备金充分或寿险适用。","workUnitIds":["W-F24"],"sourceIds":["WU-FIN-CLAIM-VARIABLES"],"evidenceStatus":"active"},{"id":"P-M-SHADOW","domainId":"math","title":"shadow辅助命题检查可编译但错题意的声明","kind":"measurement","originalPublishedAt":"2026-08-29","revisionAt":null,"revisionNote":"本轮核对原始资料并加入任务级引用；没有本站模型运行。","assessmentChangedAt":null,"summary":"语义评价比编译更贴近原目标。","whatChanged":"用能区分原含义的检查，保留漏判范围。","result":"ShadowBench在178题上，Claude Code(Opus4.8)+Numina-Lean-Agent编译率61.8%，SA-Pass为11.2%；同一配置可编译远多于语义检查通过。","configuration":"arXiv v1；178研究生/研究级题；Claude Code+Opus4.8，Medium thinking，加Numina-Lean-Agent搜索/编译；每题单次运行；Table3指标与Table10配置。隐藏shadow checker不可自部署；非Agent采样参数不混入此行。","before":"把compile rate近似题意正确率。","after":"用能区分原含义的检查，保留漏判范围。","action":"检查结论当假设等错声明。","limitation":"不是模型突然变强；assessmentChangedAt为空。","workUnitIds":["W-M05"],"sourceIds":["WU-M-SHADOW"],"evidenceStatus":"active"},{"id":"P-M-ORAGENT","domainId":"math","title":"运筹Agent评价从公式扩到可执行决策质量","kind":"measurement","originalPublishedAt":"2026-06-18","revisionAt":null,"revisionNote":"本轮核对原始资料并加入任务级引用；没有本站模型运行。","assessmentChangedAt":null,"summary":"真实多文件/规则到方案被纳入评价。","whatChanged":"按独立原约束与目标质量接收方案。","result":"ORAgentBench的GPT-5.4/Codex配置在107任务中报告35.51%通过；部分可行方案仍达不到质量阈值，可行和高质量交付不同。","configuration":"v1；107任务/14完整model-agent行；OpenAI用Codex其余Claude Code，high；45分钟交互、5分钟求解，离线Harbor/SCIP，默认四个OR skills。 此结果为GPT-5.4/Codex(high)的单次完整行；通过需可行且归一化质量q>0.4，非任意排程成功率。","before":"只验公式/代码或有可行解。","after":"按独立原约束与目标质量接收方案。","action":"保留可行性和质量两个结果，增加动态变体。","limitation":"不提供现场净收益；不能外推全部OR或本例概率。","workUnitIds":["W-M02"],"sourceIds":["WU-M-ORAGENT"],"evidenceStatus":"active"},{"id":"P-M-SCICODE-AUDIT","domainId":"math","title":"科学代码重评显示数值验收本身需要领域校准","kind":"measurement","originalPublishedAt":"2026-08-05","revisionAt":null,"revisionNote":"本轮核对原始资料并加入任务级引用；没有本站模型运行。","assessmentChangedAt":null,"summary":"原文匹配重评（含明确逐行例外）显示验收定义会改变科学代码测量。","whatChanged":"先校准科学规格、容差和金标准，再解释能力。","result":"SciCode-Verified作者的匹配重评报告12模型快照的主问题准确区间约由9–27%到69–92%；多数行沿同一评测设置，修订版两条DeepSeek V4 Preview行另有较早harness/三次均值例外。题意/金标准/容差影响科学编码测量。","configuration":"SciCode-Verified arXiv v1 Table1/§3；64可判定主问题/287子题，with-background，pass@1设计，两套固定科学Python环境OR判分。多数行是一次运行；修订版DeepSeek V4 Pro Preview与DeepSeek V4 Flash Preview两行是较早harness revision的三次运行均值，不是所有12行都仅改数据的一次同harness对照。本站未重跑。","before":"旧基准低分直接当科学能力上限。","after":"先校准科学规格、容差和金标准，再解释能力。","action":"以解析/制造解验证验收器，再评新数值实现。","limitation":"作者自报的独立审计；区间跨模型，不是本仿真或职业成功率。两条修订版Preview行的harness/重复例外须保留，不能把每行差值全部归于只换数据；修订测量也不是模型升级。","workUnitIds":["W-M06"],"sourceIds":["WU-M-SCICODE-VERIFIED"],"evidenceStatus":"active"},{"id":"P-M-STATQA","domainId":"math","title":"统计方法是否适用被纳入专属评估","kind":"measurement","originalPublishedAt":"2024-05-26","revisionAt":null,"revisionNote":"本轮核对原始资料并加入任务级引用；没有本站模型运行。","assessmentChangedAt":null,"summary":"算出数值前的列与方法选择也是错误来源。","whatChanged":"同时核对方法前提和数据生成。","result":"原作者评价任务、相关列、候选方法与适用前提。","configuration":"StatQA历史模型、构建数据和有限方法集。","before":"只验统计计算答案。","after":"同时核对方法前提和数据生成。","action":"给缺失/聚类/配对等不适用样例。","limitation":"不把旧分数当今天模型或完整统计职业能力。","workUnitIds":["W-M08"],"sourceIds":["WU-M-STATQA"],"evidenceStatus":"active"},{"id":"P-M-SAPAI","domainId":"math","title":"方案到SAP已有统计师评分的领域研究","kind":"capability","originalPublishedAt":"2026-03-19","revisionAt":null,"revisionNote":"本轮核对原始资料并加入任务级引用；没有本站模型运行。","assessmentChangedAt":"2026-03-19","summary":"研究比较完整计划中的描述与统计推理内容。","whatChanged":"有了直接候选计划证据，仍需可执行设计与独立核算。","result":"9方案/27SAP，描述项与推理项不同，存在遗漏和幻造。","configuration":"预印本v1；GPT5/Sonnet4/Gemini2.5Pro，逐节pipeline，46项统计师双评分。","before":"只能从一般文档生成推断SAP可行。","after":"有了直接候选计划证据，仍需可执行设计与独立核算。","action":"核对估计目标/缺失/中期及功效而非只审语言。","limitation":"预印本清单项质量不等于整计划通过或合规；未给本站效果。","workUnitIds":["W-M07"],"sourceIds":["WU-M-SAPAI"],"evidenceStatus":"active"},{"id":"P-M-BIX3","domainId":"math","title":"原始生物数据到研究产物出现可执行长链评估","kind":"capability","originalPublishedAt":"2026-08-26","revisionAt":null,"revisionNote":"本轮核对原始资料并加入任务级引用；没有本站模型运行。","assessmentChangedAt":"2026-08-26","summary":"按指定方法的研究尺度计算可以评价最终结构产物。","whatChanged":"可以委派较长计算研究候选，同时验完整性/链路失效。","result":"20研究/138产物，原文最高任务平均产物比例0.48；不是整研究通过率。","configuration":"13模型，32vCPU/128GB，24h，方法指导；产物阈值0.80校准。","before":"只看短问答或单一步分析。","after":"可以委派较长计算研究候选，同时验完整性/链路失效。","action":"从锁定方案/数据到分析集和结果逐层核对。","limitation":"不评价研究问题/方法选择，不能外推临床试验效度。","workUnitIds":["W-M08"],"sourceIds":["WU-M-BIX3"],"evidenceStatus":"active"},{"id":"P-M-DAFNY","domainId":"math","title":"程序验证注解已有可重放任务评价","kind":"measurement","originalPublishedAt":"2024-06-12","revisionAt":null,"revisionNote":"本轮核对原始资料并加入任务级引用；没有本站模型运行。","assessmentChangedAt":null,"summary":"固定程序/规格下补全验证提示不同于数学解题。","whatChanged":"检查程序的完整验证状态而非只看代码文本。","result":"原作者用移除提示的Dafny程序评价模型补全。","configuration":"DafnyBench v1，750余程序、历史模型/反馈。","before":"以通用写代码推测有形式保证。","after":"检查程序的完整验证状态而非只看代码文本。","action":"固定规格并用空洞前提控制。","limitation":"不证明业务规格正确或整个系统安全。","workUnitIds":["W-M10"],"sourceIds":["WU-M-DAFNY"],"evidenceStatus":"active"},{"id":"P-M-PROGRAM-PROOFS","domainId":"math","title":"程序验证代理交付可重放Rocq证明义务","kind":"capability","originalPublishedAt":"2025-11-21","revisionAt":null,"revisionNote":"本轮核对原始资料并加入任务级引用；没有本站模型运行。","assessmentChangedAt":"2025-11-21","summary":"程序来源的证明义务有专属代理成果。","whatChanged":"采用程序验证专属实验，保留需求/模型/实现连接边界。","result":"SV-COMP641义务中报告198通过；义务不是程序。","configuration":"AutoRocq v1/GPT4.1 temp0、Rocq8.18；新源码9.0.0另需运行。","before":"从数学形式证明猜测工业验证能力。","after":"采用程序验证专属实验，保留需求/模型/实现连接边界。","action":"重放固定模块全部义务与一个代码变更。","limitation":"局部引理和Linux模块案例不等于完整内核验证。","workUnitIds":["W-M10"],"sourceIds":["WU-M-AUTOROCQ"],"evidenceStatus":"active"},{"id":"P-M-CRYPFORM","domainId":"math","title":"密码模型的生成/转换开始独立于通用编码测量","kind":"measurement","originalPublishedAt":"2026-06-24","revisionAt":null,"revisionNote":"本轮核对原始资料并加入任务级引用；没有本站模型运行。","assessmentChangedAt":null,"summary":"符号与计算安全模型有专属形式分析评价。","whatChanged":"同时看角色/攻击者/目标及工具状态。","result":"原作者比较五类任务与语义/可执行错误。","configuration":"v1；9模型、7工具语言，任务分母不同。","before":"只按代码能编译推密码分析可用。","after":"同时看角色/攻击者/目标及工具状态。","action":"加入已知攻击与错误nonce绑定控制。","limitation":"不覆盖生产代码/硬件泄漏或现实部署可靠性。","workUnitIds":["W-M11"],"sourceIds":["WU-M-CRYPFORM"],"evidenceStatus":"active"},{"id":"P-M-SHANNON","domainId":"math","title":"密码引理证明工程已有EasyCrypt代理案例","kind":"capability","originalPublishedAt":"2026-07-03","revisionAt":null,"revisionNote":"本轮核对原始资料并加入任务级引用；没有本站模型运行。","assessmentChangedAt":"2026-07-03","summary":"给定模型和引理分解后可以委派证明脚本搜索。","whatChanged":"可搜索证明并完整重验，模型选择/部署仍另核。","result":"原作者报告ChaChaPoly1305/MEE-CBC相关证明工程案例。","configuration":"ShannonProver v1、EasyCrypt、多分支调度；私有污染控制和公开材料分开。","before":"专家手工填写全部tactic。","after":"可搜索证明并完整重验，模型选择/部署仍另核。","action":"固定最终目标，先重放公开案例再试未见义务。","limitation":"10月源码更宽分解能力不继承v1已测范围。","workUnitIds":["W-M11"],"sourceIds":["WU-M-SHANNON-PAPER"],"evidenceStatus":"active"},{"id":"P-M-TUTOR","domainId":"math","title":"真人辅导AI建议开始按学生学习结果评价","kind":"capability","originalPublishedAt":"2024-10-03","revisionAt":null,"revisionNote":"本轮核对原始资料并加入任务级引用；没有本站模型运行。","assessmentChangedAt":"2024-10-03","summary":"研究观察建议是否改变教学行为与掌握。","whatChanged":"可以评价对真实辅导与独立学生结果的影响。","result":"原作者预注册现场试验，导师可获实时建议。","configuration":"TutorCoPilot v1，人机组合；不是无人教师。","before":"只看AI回答是否像好解释。","after":"可以评价对真实辅导与独立学生结果的影响。","action":"把真人建议与学生直接使用工具分开。","limitation":"学校/课题掌握，不推大学长期保持。","workUnitIds":["W-M09"],"sourceIds":["WU-M-TUTORCOPILOT"],"evidenceStatus":"active"},{"id":"P-M-LEARNING","domainId":"math","title":"数学学习试验区分练习表现与无工具考试","kind":"measurement","originalPublishedAt":"2025-06-25","revisionAt":null,"revisionNote":"本轮核对原始资料并加入任务级引用；没有本站模型运行。","assessmentChangedAt":null,"summary":"工具内答题变好不等于独立掌握。","whatChanged":"无AI独立结果成为接收依据。","result":"普通AI界面与教学约束界面结果不同，短期无AI考试另测。","configuration":"PNAS2025；土耳其高中近千学生，GPT4研究界面；原文独立评分。","before":"用练习完成率或模型正确率推学习。","after":"无AI独立结果成为接收依据。","action":"分开即时、延迟和迁移，不编长期改善。","limitation":"来源为作者稿/正式版关联；不是大学数学或长期保持。","workUnitIds":["W-M13"],"sourceIds":["WU-M-TEACH-RCT"],"evidenceStatus":"active"},{"id":"P-M-KHAN2Y","domainId":"math","title":"两学年学校试验将实际使用和AI边际贡献带入判断","kind":"measurement","originalPublishedAt":"2026-08-17","revisionAt":null,"revisionNote":"本轮核对原始资料并加入任务级引用；没有本站模型运行。","assessmentChangedAt":null,"summary":"较长真实教学研究不能省略参与及组合因素。","whatChanged":"同时报告真实参与、对照和可归因范围。","result":"原论文报告组合方案成绩及低频使用，收益类似无AI练习。","configuration":"NBER35620；18中学/两年集群随机，Khan Academy与Khanmigo。","before":"提供AI入口便预期自动改善长期学习。","after":"同时报告真实参与、对照和可归因范围。","action":"大学应用先设计自己的独立与延迟测量。","limitation":"持续干预两年不等于停用后保持；无大学外推。","workUnitIds":["W-M13"],"sourceIds":["WU-M-KHAN2Y"],"evidenceStatus":"active"}],"outlooks":[{"id":"O-COMPUTING","domainId":"computing","title":"2029–2031：从单份修改到持续交付的软件Agent","horizon":"未来3–5年：2029–2031","basis":"整包应用案例、真实修复接收和运行流程优化，是更长软件任务可委派的正面信号。以下是本站据此提出的条件情景，不是这些研究给出的日期预测。","sourceIds":["WU-APPS","WU-PR","WU-HARNESS","WU-EMPLOYMENT","WU-METR"],"scenarios":[{"id":"baseline","title":"基准情景：完整修改成为常用交付","outcome":"在上下文与测试较成熟的系统中，Agent承接功能、修复和回归的一体交付，维护者从逐步编写转向目标、异常与接收；审查的一部分继续自动化。","signals":["未见仓库和变更类型仍保持可验收交付","返工、接收和线上恢复成本随使用改善","环境/上下文变化能被检测并更新"],"roles":"工程岗位更常覆盖需求、系统取舍与运行结果；初级培养需从只写代码升级为理解与验证完整修改。","action":"现在沉淀真实行为契约、失败样本和环境；按净交付周期衡量收益，不以代码量替代。"},{"id":"accelerated","title":"加速情景：软件成为按目标持续运行的系统","outcome":"若长链恢复、跨仓库迁移和自动验证持续进步，Agent可在权限与预算内规划、实现、迁移和维护较大系统；人可用更少逐步干预探索新产品，团队边界可能重组。","signals":["跨多系统、数周目标能在真实变化后继续交付","独立评估能发现自生成测试漏掉的重要错误","完整人工介入和故障成本持续下降，且系统可迁移到新业务"],"roles":"稀缺工作可能转向选目标、产品机会、资源和影响取舍；不能预设人工代码审查永远是主要瓶颈。","action":"把工具和职责做成可替换接口；试验更大交付，但按实际触发信号扩围，并安排成员转向新的业务机会。"},{"id":"constrained","title":"受限情景：编码快，集成与组织仍慢","outcome":"模型产物更强，但遗留系统、隐性需求、权限和恢复成本使自主交付扩展较慢；固定程序与AI协作长期并存。","signals":["未知依赖和需求变动仍引发大幅返工","审查和事故处理抵消生成提速","同一Agent换系统后需要重新铺上下文"],"roles":"系统知识、集成和沟通能力仍有高价值；工具投入首先服务复现、测试和资料。","action":"优先修复反复出现的集成障碍；保持可回退的部署和岗位安排，避免依据榜单单独裁定编制。"}],"watchNext":["实际未见仓库的完整交付与维护记录","跨月净人工时间、回滚与事故情况","任务/环境/模型版本变化后的迁移表现"],"evidenceStatus":"active"},{"id":"O-MATH","domainId":"math","title":"2029–2031：从辅助解题到可检验的研究伙伴","horizon":"未来3–5年：2029–2031","basis":"开放问题的形式证明与专业研究产物表明，AI已不只局限于套路题。数学的可验证性可推动能力扩展；跨领域独立选题与持续发现是否普及仍需观察。以下为条件情景。","sourceIds":["WU-NEXUS","WU-MATH-RELEASE","WU-ALETHEIA","WU-LEAN"],"scenarios":[{"id":"baseline","title":"基准情景：更多证明与模型工作整包委派","outcome":"形式库和工具完善后，AI承担更大比例的引理搜索、推导核查、建模代码与验证；研究者能并行探索更多候选，日常工作从手推每步转向问题结构和结果理解。","signals":["不同来源未见命题的搜索覆盖扩大","自然语言到形式声明错误减少且可独立检查","证明和失败记录可重放，合理预算内有持续成果"],"roles":"数学人员需要连接问题、形式语言与解释；验证工程和研究协作能力重要，单纯手算的相对价值可能下降。","action":"建立可检查的问题集、形式库与数值对照；同时练习读懂AI产物和发现被弱化的条件。"},{"id":"accelerated","title":"加速情景：自主提出并推进研究方向","outcome":"若AI能在新领域定位重要问题、建立所缺理论与验证链，并持续获得独立接收，研究单元可能扩大为选题到证明或反例的完整计划，成为接近通用研究能力的重要证据。","signals":["多家独立团队复核新的重要结果，非只重复熟悉题型","可迁移到缺少现成形式库的领域","完整尝试分布、成本与后续修订透明，而非只见成功题"],"roles":"人的工作可能集中于研究目标、价值选择、概念理解和与现实系统连接；证明搜寻与部分审查本身也可能高度自动化。","action":"为更强系统准备可复用问题、语义验证与贡献记录；支持快速尝试新方向，按实际成果重新配置研究职责。"},{"id":"constrained","title":"受限情景：可验证子题强，开放研究迁移慢","outcome":"AI在库内证明和数值工具上不断进步，但形式化成本、选题与研究意义判断限制更广泛自主发现。","signals":["成果集中于少数可形式化题型","新领域经常缺定义、引理或可靠语义转换","专家解释与校核成本仍随产物量增长"],"roles":"领域理论、形式化桥接和问题设计仍是瓶颈；不会只剩通用提示技巧。","action":"优先填高复用基础库与语义缺口，按研究问题选择AI辅助范围，不把未解题全部解释为AI不可用。"}],"watchNext":["新结果的独立复核、修订及学术接收","公开尝试分母、资源与未解任务","跨领域、跨表示的题意保持和长期研究循环"],"evidenceStatus":"active"},{"id":"O-FINANCE","domainId":"finance","title":"2029–2031：从文件助手到连续财务与研究流程","horizon":"未来3–5年：2029–2031","basis":"跨文件金融模板与工作簿评估已提供完整交付的起点。更强推理与编程可把财务操作进一步串联，但组织数据与接收流程决定扩展速度。以下为编辑情景，非投资预测。","sourceIds":["WU-FINANCE-AGENTS","WU-FINCH","WU-EMPLOYMENT"],"scenarios":[{"id":"baseline","title":"基准情景：常规更新和对账形成持续流水线","outcome":"报告更新、对账、异常分流和备忘录刷新逐步合并；多数重复数值链由可追溯程序维护，AI处理变化与说明。","signals":["跨期、重述和异构文件可稳定处理","关键数字与公式错误能在采纳前被检测","完整复核时间低于人工基线"],"roles":"财务与研究岗位更重口径、异常、模型假设和业务解释；基础训练仍须懂账和公式，才能诊断错误。","action":"先固化来源追溯与重算，比较异常类型和接收耗时；让成员学会处理复杂差异。"},{"id":"accelerated","title":"加速情景：企业财务状态可持续模拟和执行","outcome":"若跨系统数据、权限与恢复形成可靠闭环，Agent可持续对账、更新模型并提出经营动作；在明确授权范围内自动执行低影响动作，部分管理分析也可委派。","signals":["跨部门系统状态可核对且异常恢复有证据","新业务口径可被学习、解释并通过独立验收","端到端错误和人工接管成本随规模下降"],"roles":"工作重心可能转向经营目标、资本取舍、客户与制度安排；不预设所有分析或审核会永久由人执行。","action":"把账务动作与研究建议区分授权，建立小范围可恢复试点；将节省时间用于新的经营分析，并持续重新评估职责。"},{"id":"constrained","title":"受限情景：数据与口径碎片化限制整合","outcome":"模型会算会写，但数据授权、旧表格、重述和机构规则使完整流程难迁移；局部文件辅助继续有价值。","signals":["接入与清洗成本高于重复业务收益","异常口径没有明确责任人","表格/系统变化频繁破坏自动链路"],"roles":"数据治理、财务制度和系统集成仍是主要投入；单靠换更强模型收益有限。","action":"先治理高频来源与主键，减少重复口径争议；按可核查业务收益逐步投资，避免模板一上就当整岗替代。"}],"watchNext":["实际月结和研究更新的端到端记录","异常接管、错数与恢复成本","跨机构/新业务可迁移范围及采纳后的结果"],"evidenceStatus":"active"},{"id":"O-ANALYSIS","domainId":"analysis","title":"2029–2031：从回答数据问题到运行学习循环","horizon":"未来3–5年：2029–2031","basis":"分析Agent已展示问题到查询与报告的循环；编程和数学能力提升可能扩展自动诊断、实验分析和模拟。能计算与能识别因果仍是不同条件，以下为情景推演。","sourceIds":["WU-DATA-AGENT","WU-DATA-APP","WU-NEXUS","WU-SRM"],"scenarios":[{"id":"baseline","title":"基准情景：常规分析按问题整包交付","outcome":"指标、分解和标准实验检查可由Agent连贯完成，分析师减少搬表和改SQL时间；重复计算固化，新的业务问题继续探索。","signals":["新问题能选对数据并保持指标语义","输出连同查询、口径和检查可重放","缺数据时能正确缩窄结论而非编故事"],"roles":"分析岗位更常兼顾业务问题、数据语义和实验设计；只会生成图表的差异化下降。","action":"沉淀指标语义和真实反例，要求报告回答“下一步什么信息会改变决定”。"},{"id":"accelerated","title":"加速情景：Agent持续提出、检验并更新业务假设","outcome":"如果跨工具执行、统计推理和目标理解持续改善，Agent可主动发现变化、设计可检验假设、运行授权实验并更新策略，形成跨周的学习循环。","signals":["在新业务中提出的假设有可验证增益","能识别随机化、混杂与外部变化，不只追显著性","动作、结果和策略更新有透明记录，失效时能恢复"],"roles":"人的职责可能转向选择目标、约束和价值取舍；部分分析与实验规划也可由AI承担，不必把手工审每张图当长期目标。","action":"用明确授权的小实验建立从问题到结果的闭环，保留干预和数据生成记录，按净效果扩大范围。"},{"id":"constrained","title":"受限情景：答案快了，决策学习仍慢","outcome":"数据缺失、指标漂移、实验难实施或组织不能行动，会限制分析自动化的实际价值。","signals":["报表增多但决策不变","关键数据无法说明生成机制","观察相关性反复被误写为因果效果"],"roles":"数据工程、业务协作与实验条件建设仍有高价值；分析的瓶颈未必是推理模型。","action":"追踪一份报告是否改变行动以及结果，优先补采集和设计；减少不被使用的图表与重复报告。"}],"watchNext":["未见业务问题的口径保持和结果核验","实验与实际动作后的可识别效果","报告数量之外的决策周期、错误行动与学习收益"],"evidenceStatus":"active"}],"sources":[{"id":"WU-PR","title":"AI参与的修复PR：哪些被项目接收，哪些没有","url":"https://arxiv.org/html/2602.00164v1","publishedAt":"2026-01-29","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"§2样本、表1、§3.2失败原因","scope":"五类Agent参与的8106个修复PR中5267已合并、2113关闭未合并、726仍开放；提供修复交付与项目接收的观察。","limitations":"非随机样本，Agent配置与项目不同；合并不等同全部正确或净省时，不用于本站成功率。"},{"id":"WU-APPS","title":"把长时间应用开发交付给Agent","url":"https://www.anthropic.com/engineering/harness-design-long-running-apps","publishedAt":"2026-03-24","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"规划/生成/评价分工、实际应用操作及评价者限制","scope":"展示围绕可用应用的实现、操作评价与修正循环，支持以用户行为验收功能。","limitations":"作者工程案例；未复现演示应用，不能推出任意项目可靠率或多Agent必优。"},{"id":"WU-HARNESS","title":"Harness²：固定执行模型，改进运行流程","url":"https://arxiv.org/html/2610.14224v1","publishedAt":"2026-10-09","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"§2.2、§3.1表1、附录B/D","scope":"WorkBuddy-Bench Code上的平均检查得分变化及工具/上下文/检查流程的优化方法。","limitations":"检查项均分不是项目通过率；经验流内相关，候选与编辑成本另计；未验证本站任务。"},{"id":"WU-EMPLOYMENT","title":"Anthropic内部的工作与职责变化","url":"https://www.anthropic.com/research/employment-at-anthropic","publishedAt":"2026-10-16","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"Bottlenecks and process；采用非随机限制","scope":"单公司现场观察到Agent组织、上下文维护和注意力分配成为新的工作。","limitations":"公司业务及采用选择特殊；产出代理量不等于完整业务价值，不外推行业岗位数量。"},{"id":"WU-TB4","title":"Terminal-Bench 4.0的资源与验收修订","url":"https://www.tbench.ai/news/terminal-bench-4-0","publishedAt":"2026-08-28","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"Calibrating resources；Removing and fixing tasks；Versioning","scope":"环境资源、题目和验收器会改变测量结果；不同环境版本需重跑。","limitations":"这是测量修订，不是模型能力突破；基准预算不能照搬成业务时限。"},{"id":"WU-CONTINUOUS","title":"持续维护评测：何时复用、重评和重跑","url":"https://www.tbench.ai/news/continuous-benchmarks","publishedAt":"2026-07-30","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"Versioning and Publishing Tasks；CI on New and Improved Tasks","scope":"保存产物，按输入环境或验收变化分别重跑或重评；用参考、空操作与作弊控制校准。","limitations":"方法论依据，不提供本站工作可靠率。"},{"id":"WU-EVAL","title":"Agent评估：以交付与最终状态验收","url":"https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents","publishedAt":"2026-01-09","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"The structure of an evaluation；grader类型；任务设计","scope":"区分任务、试次、产物和轨迹；以明确目标与可通过参考解校准验收。","limitations":"方法建议不是某领域能力试验；模型裁判需独立校准。"},{"id":"WU-NEXUS","title":"AlphaProof Nexus：形式证明搜索进入开放数学问题","url":"https://arxiv.org/html/2605.22763v1","publishedAt":"2026-05-21","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"摘要、§2输入输出、§5与§6、附录A.1/B.2","scope":"论文报告在353个Erdős问题中解决9个、492个OEIS猜想中证明44个，并公开Lean产物；展示搜索和内核反馈循环。","limitations":"选题、形式库和搜索预算影响覆盖；未逐条重放论文证明，也不代表自然语言题意已自动核验。"},{"id":"WU-MATH-RELEASE","title":"OpenAI公开新的数学研究产物","url":"https://openai.com/index/sharing-ai-progress-in-mathematics/","publishedAt":"2026-10-06","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"正文结果发布、Lean材料与修订协议；10月15日资金更新","scope":"内部前沿模型的研究结果、部分Lean形式化及尝试信息被公开，可逐项检查。","limitations":"并非每项结果均获学界接受或有形式证明；内部模型不等于公众产品；10月15日更新不是新能力发布。"},{"id":"WU-ALETHEIA","title":"Deep Think在专业数学研究中的生成、核查与修订","url":"https://deepmind.google/blog/accelerating-mathematical-and-scientific-discovery-with-gemini-deep-think/","publishedAt":"2026-02-11","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"The Frontier of Pure Mathematics；贡献分类；Expanding to Physics and Computer Science","scope":"研究案例包含找证明、找反例与跨领域联系；区分自主和协作贡献，并允许返回未解决。","limitations":"自然语言核查器不是形式内核；发布时作者未声称其最高两档重大突破，未逐篇复核案例。"},{"id":"WU-LEAN","title":"Lean官方：怎样验证一份证明","url":"https://lean-lang.org/doc/reference/latest/ValidatingProofs/","publishedAt":"未标注（持续更新文档）","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"检查形式陈述、内核与公理依赖相关章节","scope":"固定目标后检查证明与传递公理，另核对形式陈述是否等于原问题。","limitations":"工具方法而非模型能力数据；latest会变，实际运行须锁定Lean和Mathlib版本。"},{"id":"WU-SCIPY","title":"SciPy线性规划：约束、整数性和退出状态","url":"https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.linprog.html","publishedAt":"未标注（持续更新文档）","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"Parameters: bounds/integrality；Returns: status/slack/con；Examples","scope":"提供可程序化求解和残差/状态检查的工具方法，区分不可行、无界与数值失败。","limitations":"不是AI测试；仅覆盖该求解器支持的问题，不能把线性规划验收外推到任意数值模型。"},{"id":"WU-FINANCE-AGENTS","title":"金融Agent模板：模型维护、对账与研究交付","url":"https://www.anthropic.com/news/finance-agents","publishedAt":"2026-05-05","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"New agent templates for finance work；跨应用交付","scope":"官方公开金融工作参考架构，包含报告到模型、总账对账和研究备忘录等跨文件交付。","limitations":"发布说明和模板存在不等于已验证用户公司的流程、正确率或投资收益；客户节时引语未作证据。"},{"id":"WU-FINCH","title":"FINCH：按完整财务工作簿验收","url":"https://arxiv.org/html/2512.13168v1","publishedAt":"2025-12-15","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"§2.2工作流组合；§2.3验收；§3.2.1自动与人工比较","scope":"跨表工作流应检查数值公式、覆盖、非预期改动与可读性，并接受等价实现。","limitations":"旧版本基准方法，未运行其任务；不能把自动裁判一致性或旧模型成绩充当当前现场能力。"},{"id":"WU-DATA-AGENT","title":"OpenAI内部分析Agent：从业务问题到查询与报告","url":"https://openai.com/index/inside-our-in-house-data-agent/","publishedAt":"2026-01-29","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"How it works；Context；Evals；Guide the Goal, Not the Path","scope":"展示数据发现、多轮查询、复核和发布的闭环；口径与生成数据的代码是必要上下文。","limitations":"内部专用系统；没有给出本站三项分析的可靠率，也没有证明任意因果问题可自动识别。"},{"id":"WU-DATA-APP","title":"数据分析应用参考：查询可见，修正明确保存","url":"https://developers.openai.com/cookbook/examples/agents_api/apps/data_analyst/readme","publishedAt":"2026-09-10","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"目标、数据/工具、示例边界与记忆保存","scope":"可实施的数据分析应用参考，支持发现schema、只读查询、呈现SQL和控制修正保存。","limitations":"工程示例而非能力或节时试验；未执行示例代码，使用时仍须适配身份、数据库与口径。"},{"id":"WU-SRM","title":"样本比例异常：实验看似有效也可能有数据问题","url":"https://www.microsoft.com/en-us/research/publication/diagnosing-sample-ratio-mismatch-in-online-controlled-experiments-a-taxonomy-and-rules-of-thumb-for-practitioners/","publishedAt":"2019-07","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"作者摘要中的SRM定义、错误决策与诊断范围","scope":"实验分配比例异常是需定位原因的数据质量信号，不能跳过诊断直接接受效果结论。","limitations":"长期适用的方法背景，不是2026年AI进展；SRM无异常也不能证明所有识别假设成立。"},{"id":"WU-METR","title":"任务时间跨度：软件任务趋势及其适用范围","url":"https://metr.org/time-horizons/","publishedAt":"未标注（最后更新2026-09-08）","checkedAt":"2026-10-19","nextReview":"2026-10-26","status":"active","locator":"页首维护状态；Methodological Details；FAQ","scope":"在给定软件类任务分布上，按人类专家任务时长描述模型可完成难度，适合观察长链能力趋势。","limitations":"此页已停止主动更新；时间跨度不是Agent自主运行时长，不直接对应行业自动化或AGI日期。"},{"id":"WU-OPENHANDS","title":"OpenHands Agent Canvas：固定版本实现入口","url":"https://github.com/OpenHands/OpenHands/blob/f3caa2778553662c22715bfb52b5eb6edf975d5b/README.md","publishedAt":"2026-10-19","checkedAt":"2026-10-19","nextReview":"2026-11-02","status":"active","locator":"README及同commit下已记录的配置/测试/源码路径","scope":"固定仓库版本的实现资料只读核查","limitations":"未执行该最新仓库的模型工作；不可把README展示结果当本站复现"},{"id":"WU-RDAGENT","title":"Microsoft RD-Agent：研究循环实现","url":"https://github.com/microsoft/RD-Agent/blob/484776c211e4fbbeef03e0ec00d6bbee7362a4f4/README.md","publishedAt":"2026-09-23","checkedAt":"2026-10-19","nextReview":"2026-11-02","status":"active","locator":"README及同commit下已记录的配置/测试/源码路径","scope":"固定仓库版本的实现资料只读核查","limitations":"未执行该最新仓库的模型工作；不可把README展示结果当本站复现"},{"id":"WU-MATHLIB","title":"Mathlib：固定工具链与构建入口","url":"https://github.com/leanprover-community/mathlib4/blob/e3e75e86143b43849829fdcd44a39a6b0fe35234/README.md","publishedAt":"2026-10-19","checkedAt":"2026-10-19","nextReview":"2026-11-02","status":"active","locator":"README及同commit下已记录的配置/测试/源码路径","scope":"固定仓库版本的实现资料只读核查","limitations":"未执行该最新仓库的模型工作；不可把README展示结果当本站复现"},{"id":"WU-COGNITION","title":"Cognition：内部提交代码的Devin采用声明","url":"https://cognition.com/blog/series-d","publishedAt":"2026-05-27","checkedAt":"2026-10-19","nextReview":"2026-11-02","status":"active","locator":"What’s Next段","scope":"厂商对自身工程团队提交代码的自述","limitations":"89%是原文代码提交口径；未披露本站需要的工作量分母、统计窗口、复核人时与独立审计"},{"id":"WU-ARISTOTLE","title":"Harmonic Aristotle产品能力声明","url":"https://aristotle.harmonic.fun/","publishedAt":"未标注；2026-10-19观察","checkedAt":"2026-10-19","nextReview":"2026-11-02","status":"active","locator":"Fully Agentic / Library-Ready Code","scope":"厂商声明可从英文问题或Lean项目开展证明工作","limitations":"本次只核对原始声明，没有调用服务或核验其广泛正确性"},{"id":"WU-DEEPSWE-V11","title":"DeepSWE v1.1：仓库变更及独立行为验收","url":"https://deepswe.datacurve.ai/","publishedAt":"2026-09-22（页面结果更新）","checkedAt":"2026-10-20","nextReview":"2026-11-03","status":"active","locator":"v1.1任务分布；Run DeepSWE / Quickstart / Agents and models；https://deepswe.datacurve.ai/run","scope":"作者发布的仓库级任务与重放入口；113任务、91仓库、5语言。","limitations":"未在本站运行；任务分布不等于企业后端覆盖率，网页会更新，执行前需固定仓库与runner提交。"},{"id":"WU-SWE-PRO-AUDIT","title":"OpenAI：SWE-Bench Pro评测问题审计","url":"https://openai.com/index/separating-signal-from-noise-coding-evaluations/","publishedAt":"2026-07-08","checkedAt":"2026-10-20","nextReview":"2026-11-03","status":"active","locator":"SWE-Bench Pro任务审计、过严测试/题意缺失/测试覆盖问题","scope":"模型厂商对参评软件基准的原始审计说明。","limitations":"作者是参评厂商；其问题比例估计不等于整个基准无效，也不证明本站验收器有相同缺陷。"},{"id":"WU-FINCH-V5","title":"Finch v5：完整财务工作流的人评与自动评价","url":"https://arxiv.org/html/2512.13168v5","publishedAt":"2025-12-15；v5 2026-04-15","checkedAt":"2026-10-20","nextReview":"2026-11-03","status":"active","locator":"Evaluation / Human evaluation / Main results；官方实现 https://github.com/FinWorkBench/Finch","scope":"作者报告172个复合workflow、384任务，区分产品代理、API及人工/自动评价。","limitations":"模型与配置属于论文历史版本；公开的是数据及评价管道，xlwings/本机Excel是环境依赖；不是现成通用财务代理。"},{"id":"WU-SPREADSHEETBENCH2","title":"SpreadsheetBench 2 v1：业务工作簿整表验收","url":"https://arxiv.org/html/2606.29955v1","publishedAt":"2026-06-29","checkedAt":"2026-10-20","nextReview":"2026-11-03","status":"active","locator":"Task taxonomy / Evaluation / Main results；https://github.com/RUCKBReasoning/SpreadsheetBench-2","scope":"321项端到端业务工作簿任务，区分修改单元格、整表正确和图表rubric。","limitations":"数值与图表口径不同，产品另测30例子集；未验证10月新模型。输入/golden/输出需重算，图表检查另需Excel/WPS COM。"},{"id":"WU-NEXUS-125","title":"Nexus Erdős125：固定公开证明产物","url":"https://github.com/google-deepmind/alphaproof-nexus-results/blob/0647711a71183c1ea492ad60860776617ce1ea88/APNOutputs/ErdosProblems/erdos_125.variants.positive_lower_density.lean","publishedAt":"2026-06-05（固定提交）","checkedAt":"2026-10-20","nextReview":"2026-11-03","status":"active","locator":"L42-L54定义；L342-L370 target_theorem_0；同提交 NaturalLanguageProofs/ErdosProblems/erdos125.pdf、lean-toolchain及lake-manifest.json","scope":"可读取的特定数学产物；三进制/四进制0、1数集和集的正下密度命题。","limitations":"只核对文本，没有本站Lean重放；定理的False答案须连同目标声明解释，不能把邻近问题或变体当原题自动通过。"},{"id":"WU-ARISTOTLE-SDK210","title":"Aristotle官方分发包2.1.0与调用入口","url":"https://pypi.org/project/aristotlelib/2.1.0/","publishedAt":"2026-06-18","checkedAt":"2026-10-20","nextReview":"2026-11-03","status":"active","locator":"Quick Start / Authentication / CLI与API；服务 https://aristotle.harmonic.fun/","scope":"官方SDK分发的安装、鉴权和证明工作调用方式。","limitations":"需要账户/API key；本轮未调用；SDK可安装不代表研究系统Nexus可调用，不继承IMO评估预算或成绩。"},{"id":"WU-UNDERWRITE","title":"UNDERWRITE：在企业准则与工具环境中评估核保Agent","url":"https://arxiv.org/html/2602.00456v1","publishedAt":"2026-01-31","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"The Underwrite Environment；Evaluating Models；Table 1及pass^k分析","scope":"商业财产责任险的合成企业环境，包含准则、数据库和模拟用户；评估回答正确性与交互。","limitations":"不是真实承保/赔付结果；正文最高分与Table 1口径存在差异，此处不转录榜单分数，也不外推当前最强模型。"},{"id":"WU-IAB","title":"Cooper IAB：保险文件理解的厂商评估","url":"https://www.askcooper.ai/labs/insurance-agent-benchmark","publishedAt":"页面未标首次日期；实验窗2026-08-13—09-09","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"§4 How we built IAB；§5评分定义；§6失败实例","scope":"166份保险文件的字段、条款、跨文件与缺失值处理；厂商自报单次评估，含文档系统对照。","limitations":"平均judge分含部分得分，不是核保或理赔流程通过率；文档访问受限，本站未独立复验。"},{"id":"WU-CLAIMS-DEMO","title":"AWS保险理赔Agent示例：可检查的组织与集成入口","url":"https://github.com/aws-samples/sample-agentic-insurance-claims-processing-demo/tree/0bab9ac348e830b016c3faaa2a31ab6618dbf385","publishedAt":"首次日期未标；固定提交2026-10-09","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"README What It Does、Demo Limitations & Production Considerations、API Reference","scope":"寿险身故理赔参考实现，提供样例保单与模拟场景、流程和运行依赖。","limitations":"README核对，未运行；预置保单、纯文本文档和示例路径不证明真实身份/证据真实性或一般理赔准确率。"},{"id":"WU-C-VISUAL","title":"SWE-bench Multimodal 原论文","url":"https://arxiv.org/abs/2410.03859v1","publishedAt":"2024-10-04","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Abstract；任务定义与实验","scope":"视觉JavaScript缺陷；跨语言与视觉泛化研究","limitations":"历史模型/任务，不是当前前端、可访问性或产品完成率。"},{"id":"WU-C-PW","title":"Playwright accessibility testing","url":"https://playwright.dev/docs/accessibility-testing","publishedAt":"未标注单页发布日期","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Introduction；Example tests；WCAG limitations","scope":"浏览器与axe检查方法","limitations":"工具方法，不是AI成绩；自动扫描不能覆盖所有可访问性问题。"},{"id":"WU-C-PWMCP","title":"Playwright MCP 官方实现","url":"https://github.com/microsoft/playwright-mcp","publishedAt":"持续更新仓库；本次2026-10-20访问","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"README：tools、browser configuration","scope":"Agent浏览器操作与可见状态入口","limitations":"可用实现不证明任务正确；需固定执行版本并隔离账户数据。"},{"id":"WU-C-ANDROID","title":"AndroidWorld 原论文","url":"https://arxiv.org/abs/2405.14573v1","publishedAt":"2024-05-23","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Abstract；benchmark and robustness analysis","scope":"动态参数化Android应用操作评测","limitations":"操作既有app，不评估新app编写、iOS、耗电或真实设备驱动。"},{"id":"WU-C-ANDROID-REPO","title":"AndroidWorld 官方环境","url":"https://github.com/google-research/android_world","publishedAt":"持续更新仓库；本次2026-10-20访问","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"README：Installation/Quickstart/Docker","scope":"API33模拟器、状态任务和M3A运行入口","limitations":"默认legacy模型；Docker仍标experimental，未在本站安装。"},{"id":"WU-C-ANDROID-LIFE","title":"Android 活动生命周期测试","url":"https://developer.android.com/guide/components/activities/testing","publishedAt":"未标注单页发布日期","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"ActivityScenario；activity lifecycle transitions","scope":"设备事件和生命周期测试方法","limitations":"平台验收方法不是AI完成性证据；模拟器不能代替全部真机外设。"},{"id":"WU-C-MLE","title":"MLE-bench 原论文","url":"https://arxiv.org/abs/2410.07095v1","publishedAt":"2024-10-09","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Abstract；dataset与resource scaling","scope":"Kaggle训练、数据准备与实验交付","limitations":"离线竞赛不覆盖在线收益/模型运维；历史配置不代表现行最好系统。"},{"id":"WU-C-MLE-REPO","title":"MLE-bench 官方实现与可比性公告","url":"https://github.com/openai/mle-bench","publishedAt":"持续更新仓库；2026-04-24公告","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"README：Leaderboard update；Setup/agents","scope":"训练环境、评分器、运行入口","limitations":"官方暂停新榜单提交；不同预算、数据访问和脚手架成绩不可混用。"},{"id":"WU-C-MLFLOW","title":"MLflow Tracking 官方说明","url":"https://mlflow.org/docs/latest/ml/tracking/","publishedAt":"未标注单页发布日期","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Runs；Models；Tracking Runs","scope":"参数、代码版本、指标与产物记录","limitations":"记录工具不证明训练/部署正确；执行前应固定依赖。"},{"id":"WU-C-SPIDER","title":"Spider2 官方任务与维护公告","url":"https://github.com/xlang-ai/Spider2","publishedAt":"持续更新仓库；2026-08-12公告","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"News；Overview；Spider2-DBT quickstart","scope":"企业SQL与DuckDB/dbt仓库级任务","limitations":"Snowflake账号暂停公告；DBT本地路径与云环境不能等同。"},{"id":"WU-C-DBA","title":"DBA-Bench v2","url":"https://arxiv.org/abs/2607.22165v2","publishedAt":"2026-07-24；v2修订2026-10-10","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Abstract；live PostgreSQL/TiDB与persistent scenarios","scope":"有负载的状态式DB运维评估","limitations":"原文预印本；诊断、恢复与安全结果须分开，不能外推目标生产库成功率。"},{"id":"WU-C-PITR","title":"PostgreSQL18 连续归档与PITR","url":"https://www.postgresql.org/docs/18/continuous-archiving.html","publishedAt":"PostgreSQL18文档，单页未标发布日期","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"25.3.1/25.3.2/25.3.5/25.3.8","scope":"基础备份、WAL、时间点恢复与限制","limitations":"原生机制而非AI评测；配置文件需另外保护，备份存在不等于已还原。"},{"id":"WU-C-AIOPS","title":"AIOpsLab 官方框架","url":"https://github.com/microsoft/AIOpsLab","publishedAt":"持续更新仓库；本次2026-10-20访问","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"README：Problems；Fault/Workload/Evaluator","scope":"可注入故障、生成负载并观察恢复状态的试验床","limitations":"框架不是SRE端到端可靠率；拓扑、权限和业务不变量需另配。"},{"id":"WU-C-CYBENCH","title":"Cybench v4","url":"https://arxiv.org/abs/2408.08926v4","publishedAt":"2024-08-15；v4修订2025-04-12","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Abstract；任务环境与agent scaffolds","scope":"受控CTF执行与任务级能力评测","limitations":"CTF不代表生产检测、合规、完整审计或授权范围外扫描。"},{"id":"WU-C-PHISH","title":"Phishing Triage Agent 随机对照试验","url":"https://arxiv.org/abs/2511.13860v1","publishedAt":"2025-11-17","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Abstract；RCT treatment and outcomes","scope":"AI增强分析师钓鱼队列研判","limitations":"厂商相关研究；有人参与的队列实验不是自主事件响应全流程。"},{"id":"WU-C-PHISH-ACCESS","title":"Microsoft Phishing Triage Agent 接入说明","url":"https://learn.microsoft.com/en-us/defender-xdr/phishing-triage-agent","publishedAt":"未标注单页发布日期","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Prerequisites；Setup；SCU consumption","scope":"商业安全产品账号、权限、容量条件","limitations":"许可/容量有要求，未在本站登录或开通；存在产品不等于已验证本单位邮箱。"},{"id":"WU-C-EMBED","title":"EmbedAgent v3","url":"https://arxiv.org/abs/2506.11003v3","publishedAt":"2025-04-19；v3修订2026-01-23","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Abstract；Embedbench roles and evaluation","scope":"编程、原理图和跨平台迁移仿真评测","limitations":"126例、小型板卡与虚拟电路，不是工业实时、EMC或实物认证。"},{"id":"WU-C-EMBED-REPO","title":"EmbedAgent 官方实现","url":"https://github.com/icip-cas/EmbedAgent","publishedAt":"持续更新仓库；本次2026-10-20访问","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"README：run/evaluate；Wokwi","scope":"可检查虚拟电路与固件任务入口","limitations":"仿真不等于实物；工具与API费用未实测。"},{"id":"WU-C-COMPANY","title":"TheAgentCompany v3","url":"https://arxiv.org/abs/2412.14161v3","publishedAt":"2024-12-18；v3修订2025-09-10","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Abstract；simulated company environment","scope":"办公系统、跨工具任务与沟通评测","limitations":"模拟同事/公司，不能推断真实利益协调或完整项目经理责任。"},{"id":"WU-C-COMPANY-REPO","title":"TheAgentCompany 官方环境","url":"https://github.com/TheAgentCompany/TheAgentCompany","publishedAt":"持续更新仓库；本次2026-10-20访问","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"README：evaluation；task roles/result-based grading","scope":"项目与文档等公司环境的可运行任务资产","limitations":"需隔离部署、模型API与环境资源；未在本站运行。"},{"id":"WU-C-SWT","title":"SWT-Bench 官方实现","url":"https://github.com/logic-star-ai/swt-bench","publishedAt":"NeurIPS2024；仓库持续更新","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"README：Overview；Unit Test vs Reproduction Script；gold validation","scope":"缺陷前失败、修复后通过的测试生成评测","limitations":"复现脚本模式不跑全套测试；不能把覆盖率当检出率或业务正确性。"},{"id":"WU-C-REVIEW","title":"ReviewBench 官方语料与评测实现","url":"https://github.com/review-bench/ReviewBench","publishedAt":"持续更新仓库；本次2026-10-20访问","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"README：golden findings；corpus；offline review","scope":"有人工审阅发现的真实PR检查资产","limitations":"参考发现集合不是缺陷全集，LLM裁判不能代替复现；本文不引用其榜单成绩。"},{"id":"WU-C-OS2","title":"OSWorld2.0 v2","url":"https://arxiv.org/abs/2606.29537v2","publishedAt":"2026-06-28；v2修订2026-07-13","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Abstract：108 long-horizon workflows；binary/partial","scope":"跨应用长流程与隐藏状态的任务评估","limitations":"二元完成与部分分不能混用；不是企业IT恢复专属成功率。"},{"id":"WU-C-ITRCT","title":"IT管理员 Security Copilot RCT v2","url":"https://arxiv.org/abs/2411.01067v2","publishedAt":"2024-11-01；v2修订2024-11-12","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Abstract；Entra/Intune三类情景","scope":"登录、设备策略与故障的辅助管理员实验","limitations":"人加Copilot，不是自主远程维修；v2修正时间数字，不是新能力跃迁。"},{"id":"WU-C-PAPER","title":"PaperBench v3","url":"https://arxiv.org/abs/2504.01848v3","publishedAt":"2025-04-02；v3修订2025-04-07","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Abstract；replication rubrics与judge","scope":"从论文到实现和实验的复现评测","limitations":"部分复现得分不是论文完整复现率；不检验新颖性与所有理论结论。"},{"id":"WU-C-ITACCESS","title":"Security Copilot inclusion现行说明","url":"https://learn.microsoft.com/en-us/copilot/security/security-copilot-inclusion","publishedAt":"未标注单页发布日期","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Eligibility；products/permissions；agents deployment","scope":"Entra/Intune等集成与许可/容量条件","limitations":"有资格不等于租户已启用；不提供恢复实测。"},{"id":"WU-C-OSREPO","title":"OSWorld 官方实现","url":"https://github.com/xlang-ai/OSWorld","publishedAt":"持续更新仓库；本次2026-10-20访问","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"README setup/tasks/evaluation","scope":"桌面环境、任务与执行检查入口","limitations":"OSWorld原始/Verified与2.0任务集须区分；未在本站安装。"},{"id":"WU-C-EVOLVE","title":"AlphaEvolve 白皮书","url":"https://arxiv.org/abs/2506.13131v1","publishedAt":"2025-06-16","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Abstract；evolution/evaluators与algorithm discoveries","scope":"代码搜索、自动评价与算法成果","limitations":"专用研究系统和任务；不推断公开Agent均可复现，也不是自由理论发现率。"},{"id":"WU-C-EVOLVE-ART","title":"AlphaEvolve 公开产物","url":"https://github.com/google-deepmind/alphaevolve_results","publishedAt":"持续更新仓库；本次2026-10-20访问","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"README与具体算法产物目录","scope":"可检查算法/数学成果入口","limitations":"公开成果不等于完整搜索系统/训练配置开放，本站未重放。"},{"id":"WU-C-KSERVE","title":"KServe0.20发布说明","url":"https://github.com/kserve/website/blob/main/blog/2026-08-06-KServe-0.20-release.md","publishedAt":"2026-08-06","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"traffic splitting/canary/observability","scope":"模型服务流量切分与部署机制","limitations":"工具机制不代表AI能稳定完成模型运维。"},{"id":"WU-C-EVOLVE-ANNOUNCE","title":"Google DeepMind AlphaEvolve首次成果公布","url":"https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/","publishedAt":"2025-05-14","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"2025-05-14发布；algorithm discovery/computational infrastructure","scope":"专用研究系统的算法与计算成果首次公开身份","limitations":"厂商研究披露，不能代表公开服务已可获得或任意任务完成率。"},{"id":"WU-C-PG-HA","title":"PostgreSQL18 Failover","url":"https://www.postgresql.org/docs/18/warm-standby-failover.html","publishedAt":"PostgreSQL18文档，单页未标日期","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"26.3：旧主隔离、检测边界、备机重建","scope":"复制切换与恢复冗余的方法","limitations":"不是AI成绩；完整检测/通知工具不由PostgreSQL单独提供。"},{"id":"WU-C-IR-NIST","title":"NIST SP800-61r3","url":"https://csrc.nist.gov/pubs/sp/800/61/r3/final","publishedAt":"2025-04","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Table2 ID.RA-01/03/04/05；Table3 检测、分析、处置与恢复","scope":"风险路径、事件结论、证据与业务恢复方法","limitations":"专业建议，不是AI自主响应实测。"},{"id":"WU-C-CONTROL-NIST","title":"NIST SP800-53A r5","url":"https://csrc.nist.gov/pubs/sp/800/53/a/r5/final","publishedAt":"2022-01","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"assessment methods/procedures","scope":"可检查对象与测试确认控制效果","limitations":"方法不是模型成绩；组织控制和评估深度需任务定义。"},{"id":"WU-C-PAPER-CHECKLIST","title":"NeurIPS Paper Checklist","url":"https://neurips.cc/public/guides/PaperChecklist","publishedAt":"单页未标日期","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"Claims/Limitations/Reproducibility/Code/Data/licenses","scope":"研究陈述与复算材料要求","limitations":"投稿/自查方法，不是AI科研传播成绩。"},{"id":"WU-A-EDA-METHOD","title":"NIST工程统计手册：What is EDA","url":"https://www.itl.nist.gov/div898/handbook/eda/section1/eda11.htm","publishedAt":"单页未标日期","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"1.1.1 EDA定义与目的","scope":"分布、异常、结构、关系与假设探索","limitations":"统计方法，不是AI实测，也不提供因果识别。"},{"id":"WU-C-K8S-AUTH","title":"kubectl auth can-i","url":"https://kubernetes.io/docs/reference/kubectl/generated/kubectl_auth/kubectl_auth_can-i/","publishedAt":"持续更新文档，单页日期以访问版为准","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"命令说明与权限检查例","scope":"可执行授权检查接口","limitations":"工具接口不是AI成绩；目标产品、角色、许可与版本需实际核实，本站未使用。"},{"id":"WU-C-EDR-ACTIONS","title":"Defender for Endpoint设备响应操作","url":"https://learn.microsoft.com/en-us/defender-endpoint/respond-machine-alerts","publishedAt":"持续更新文档，单页日期以访问版为准","checkedAt":"2026-10-20","nextReview":"2027-01-20","status":"active","locator":"response actions、isolation、investigation package与权限","scope":"商业端点响应/证据收集接口","limitations":"工具接口不是AI成绩；目标产品、角色、许可与版本需实际核实，本站未使用。"},{"id":"WU-FIN-FINCH-V5","title":"Finch v5：完整财务工作流与产物验收","url":"https://arxiv.org/html/2512.13168v5","publishedAt":"2026-04-15","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"§2.3.1—2.3.2、§3.1、§3.2、Appendix D/F","scope":"172个复合工作流的作者报告与整包验收方法。","limitations":"2026-04-15版本与其中模型配置；非当前全部模型排行、非本站试验。"},{"id":"WU-FIN-BEAN-VALIDATION","title":"Beancount固定代码：账户、币种与交易平衡校验","url":"https://github.com/beancount/beancount/blob/9747213802acb733ce53acc6e963f5961f47d6cd/beancount/ops/validation.py#L194-L410","publishedAt":"未列统一发布日期；固定版本见URL","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"validate_active_accounts、validate_currency_constraints、validate_check_transaction_balances","scope":"可检查确定性账务校验实现。","limitations":"账务规则工具，不证明原始交易真实、会计政策正确或AI入账成功率。"},{"id":"WU-FIN-FINROBOT-DCF","title":"FinRobot固定DCF计算与边界处理","url":"https://github.com/AI4Finance-Foundation/FinRobot/blob/2717499b8e30f242640af08c4ad9afd1113c2d45/finrobot_desktop/finrobot/engine/compute/operators/dcf.py#L48-L177","publishedAt":"未列统一发布日期；固定版本见URL","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"calculate_dcf；terminal growth/WACC与terminal FCF边界","scope":"确定性DCF实现可审阅，可与Agent的资料和假设生成分开。","limitations":"仅代码抽查，未运行；不证明假设合理、正式估值或投资收益。"},{"id":"WU-FIN-FINROBOT-REPORT","title":"FinRobot已有NVDA示例研究报告","url":"https://github.com/AI4Finance-Foundation/FinRobot/blob/2717499b8e30f242640af08c4ad9afd1113c2d45/finrobot_autogen/report/NVDA_report.pdf","publishedAt":"未列统一发布日期；固定版本见URL","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"report/NVDA_report.pdf；同commit README的V0安装入口","scope":"公开可取的报告产物入口。","limitations":"只核对产物存在及来源身份，没有复算全报告，不作为质量或收益独立验证。"},{"id":"WU-FIN-QLIB-BACKTEST","title":"Qlib固定文档：成本前后回测","url":"https://github.com/microsoft/qlib/blob/54355232463878d2eebb91fe0ee5fa7fa1f5976c/docs/component/strategy.rst#L137-L260","publishedAt":"未列统一发布日期；固定版本见URL","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Running backtest；exchange_kwargs与cost分别输出","scope":"可检查回测配置、仓位和成本分析路径。","limitations":"方法和示例；不证明Agent策略有效，回测不等于实盘成交。"},{"id":"WU-FIN-QLIB-EXECUTION","title":"Qlib固定执行仿真配置","url":"https://github.com/microsoft/qlib/blob/54355232463878d2eebb91fe0ee5fa7fa1f5976c/docs/component/rl/quickstart.rst","publishedAt":"未列统一发布日期；固定版本见URL","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"backtesting配置、exchange、volume_threshold与simulator","scope":"可配置执行仿真环境。","limitations":"不是交易所或券商真实接口，成交和市场冲击模型须单独核对。"},{"id":"WU-FIN-QUANTLIB-PRICING","title":"QuantLib固定BermudanSwaption示例","url":"https://github.com/lballabio/QuantLib/blob/eaeac565149ced7c1e0cd515c0fa3ed0b89f2b39/Examples/BermudanSwaption/BermudanSwaption.cpp#L157-L290","publishedAt":"未列统一发布日期；固定版本见URL","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"模型校准及tree/fdm NPV输出","scope":"可检查数值定价与不同求解路线对照。","limitations":"不是任何模型均正确的证明；市场、合约与日期约定需冻结。"},{"id":"WU-FIN-ORE-STRESS","title":"ORE固定XVA压力与损益归因算例","url":"https://github.com/OpenSourceRisk/Engine/blob/3d75a69087911a7e2cf1e6882da7bc27135c0762/Examples/XvaRisk/Readme.md","publishedAt":"未列统一发布日期；固定版本见URL","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"XVA Stress Testing / XVA P&L Explain；run_stress.py与配置文件","scope":"可检查的定价风险情景输入和运行入口。","limitations":"示例风险组合及市场模型；不涵盖操作、反欺诈和全部机构风险。"},{"id":"WU-FIN-CHAIN-MACK","title":"chainladder固定Mack准备金实现","url":"https://github.com/casact/chainladder-python/blob/a090275cfc74deb93ece455cc576f5d20235adcf/chainladder/methods/mack.py#L8-L162","publishedAt":"未列统一发布日期；固定版本见URL","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"MackChainladder、ukmotor示例、ibnr_及预测误差","scope":"产险赔款发展三角和准备金方法可检查。","limitations":"不证明AI会选对方法、尾部假设或真实准备金充分性。"},{"id":"WU-FIN-LIFELIB","title":"lifelib固定寿险模型入口","url":"https://github.com/lifelib-dev/lifelib/blob/b706d8bbab530cc9ef94e06600e9c9d1468d8cb1/README.rst#L16-L39","publishedAt":"未列统一发布日期；固定版本见URL","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"What is lifelib / What for；basiclife模型库","scope":"寿险现金流、定价和模型测试的公开实现入口。","limitations":"README用途声明与代码入口，不是AI实测或任何法域监管适用性结论。"},{"id":"WU-FIN-TAXCALC","title":"Tax-Calculator固定模型与可选Agent入口","url":"https://github.com/PSLmodels/Tax-Calculator/blob/94c7f5f85a6d5a0c5f703179a8c4f2566d27b0cc/agents/README.md","publishedAt":"未列统一发布日期；固定版本见URL","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Model-Development Agent / Model-Use Agent；同commit README定义美国联邦个税与工资税","scope":"固定版本的税务微观模拟与自然语言模型使用路径。","limitations":"不是电子报税系统；公开说明无独立Agent成功率，不能覆盖中国税种或企业税务。"},{"id":"WU-FIN-IRS1040-2024","title":"美国2024税年Form1040固定说明","url":"https://www.irs.gov/pub/irs-prior/i1040gi--2024.pdf","publishedAt":"2024税年说明","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"封面税年、printed pp.31—37与64起税表","scope":"只作美国联邦个人所得税2024年示范的规则和表单核对入口。","limitations":"不提供本网站用户当前税率/申报结论，不外推州税、企业税或中国规则。"},{"id":"WU-FIN-AS1105","title":"PCAOB AS1105职业证据约束","url":"https://pcaobus.org/oversight/standards/auditing-standards/details/AS1105","publishedAt":"页面采用Release2010-004及列示修订","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"§.02—.10、.11、.29；美国PCAOB适用审计语境","scope":"证据相关可靠性、主张与冲突证据的专业方法。","limitations":"职业方法来源，非AI能力证明；具体审计期间与适用修订应按委托另核。"},{"id":"WU-FIN-PCAOB-GENAI-2024","title":"PCAOB人员2024年GenAI实际调研","url":"https://pcaobus.org/documents/generative-ai-spotlight.pdf","publishedAt":"2024-07","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"printed pp.4—5：Current Use / Risks；封面非规则声明","scope":"有限访谈中的资料研究、草稿和部分对账使用观察。","limitations":"历史有限访谈，非2026能力上限、非正式规则、非审计全流程成绩。"},{"id":"WU-FIN-AMLSIM","title":"IBM AMLSim固定合成交易与已知模式","url":"https://github.com/IBM/AMLSim/blob/7338a4bcb1af9bcfea2201ad7daccfe2a4d569ca/README.md","publishedAt":"未列统一发布日期；固定版本见URL","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"目标、运行和scripts/validation/validate_alerts.py","scope":"能提供带已知模式的合成数据以设计案例正负例。","limitations":"合成器不是AI调查器；合成标签不能证明真实洗钱、可疑报告法律结论或现实漏报率。"},{"id":"WU-FIN-CLOSE-TEMPLATE","title":"官方固定月结Agent模板","url":"https://github.com/anthropics/financial-services/blob/574ed3624aebd0418c7e96cd101262f30210ab26/plugins/agent-plugins/month-end-closer/agents/month-end-closer.md","publishedAt":"固定commit见URL；不推断首次功能发布日期","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"What you produce / Guardrails","scope":"计提、滚动表和差异说明组成月结候选包。","limitations":"该模板不直接过账；是可配置参考而非可靠性试验，不能当永久人机边界。"},{"id":"WU-FIN-KYC-TEMPLATE","title":"官方固定KYC案例包模板","url":"https://github.com/anthropics/financial-services/blob/574ed3624aebd0418c7e96cd101262f30210ab26/plugins/agent-plugins/kyc-screener/agents/kyc-screener.md","publishedAt":"固定commit见URL；不推断首次功能发布日期","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"What you produce / new-client onboarding not transaction monitoring","scope":"客户实体、规则结果、命中和升级包参考。","limitations":"不涵盖真实交易监测或全部合规；需内部规则、数据许可和筛查接口。"},{"id":"WU-FIN-RETIREMENT-REFERENCE","title":"cFIREsim旧开源退休模拟参考","url":"https://github.com/boknows/cFIREsim-open/blob/c7c4ba441830c1da457b663d2e602567306e8459/README.md","publishedAt":"固定commit见URL；不推断首次功能发布日期","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"README历史参考声明；同commit js/cFIREsimOpen.js","scope":"退休情景模拟有可检查历史实现。","limitations":"仓库明确只是旧参考，当前同名服务闭源；未复放，不将旧代码与当前服务成绩互换。"},{"id":"WU-FIN-FINFIRST","title":"FinFIRST原文：答案与证据分项评测","url":"https://arxiv.org/html/2609.25192v1","publishedAt":"2026-09-21","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"§3.1—3.4、§4.1 Table3、Table2例子","scope":"123题统一搜索/网页/Python配置的外部Agent研究。","limitations":"作者单次指定运行，自动裁判；Table2分子含美国加拿大而分母是美国，地域口径需独立审查，不能把高分当研究绝对正确。"},{"id":"WU-FIN-FINANCEAGENT","title":"Finance Agent原文：SEC研究任务与工具","url":"https://arxiv.org/html/2508.00828v1","publishedAt":"HTML页标2025-05-20；arXiv编号不同，不据编号推事件日","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"§2.2 Dataset split、§3、§4与AppendixC","scope":"专家问题、原始披露与工具配置的研究任务证据。","limitations":"旧配置；537题与公测50/私有150/测试337分母不能互换，不支持整套投资工作或收益。"},{"id":"WU-FIN-FINGAIA","title":"FinGAIA v1：具体信用、风险、配置任务","url":"https://arxiv.org/html/2507.17186v1","publishedAt":"2025-07-23","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Table2、Agents/Evaluation Methods、Table3及AppendixB","scope":"多模态、多工具的金融资料取材与计算评估；专业角色标签须对照Appendix A实际任务。","limitations":"Fig6为财报事实/实体核对，Fig7为新闻音频识别，Fig8为历史回撤计算，Fig9含图像计数、基金排序与持仓检索；这些样例不直接测授信、欺诈检测或配置优化。正文与附表口径也有差异，不使用统一专业成功率。"},{"id":"WU-FIN-TAXCALCBENCH","title":"TaxCalcBench固定TY25评估产物","url":"https://github.com/column-tax/tax-calc-bench/blob/f9f8b9a517cc39300b2eff57fa332dea70703c0c/README.md","publishedAt":"2026-10-19固定快照；v2于2026-06发布，未列日","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"TY2025表格、结果口径与Known issues；固定2026-10-19 commit","scope":"PDF输入、州及联邦计算和保存输出可审阅。","limitations":"通常每配置单次保存结果，按最佳thinking设置展示，部分模型覆盖不足；生成成本不等于端到端报税成本。"},{"id":"WU-FIN-TAX-SAVED-CASE","title":"TaxCalcBench TY25一个可检查成功产物","url":"https://github.com/column-tax/tax-calc-bench/blob/f9f8b9a517cc39300b2eff57fa332dea70703c0c/tax_calc_bench/ty25/results/ty25-ca-001/openai/gpt-6.1-sol/evaluation_result_high_1.md","publishedAt":"固定2026-10-19快照","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"逐行expected/actual及Strictly correct return、API Usage and Cost","scope":"一个保存的模型评估文件，可检查逐行比较及成本口径。","limitations":"单个作者自评产物、未本站重放；选中成功例不证明新案例稳定性或提交成功。"},{"id":"WU-FIN-INS-ACTBENCH","title":"INS-ActBench：精算案例、表格与R实作","url":"https://arxiv.org/html/2607.24273v1","publishedAt":"2026-07-27","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"§3.2—3.3、§4.1 Evaluation、Table3、§4.4","scope":"专业精算材料改编的直接LLM任务评估。","limitations":"知识/案例经选择题改编；表格按答案格平均，不能换成整表通过率、真实准备金充足或长期精算完成率。"},{"id":"WU-FIN-INVESTORBENCH","title":"InvestorBench：单资产历史决策Agent","url":"https://aclanthology.org/2025.acl-long.126.pdf","publishedAt":"ACL2025会议论文；页面未列统一日","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"§3—4、AppendixF/G与Limitations","scope":"在固定历史市场环境比较Agent决策与基线。","limitations":"单资产历史模拟，早期模型与有限市场时段；不支持组合、真实成交、当前模型alpha或利润保证。"},{"id":"WU-FIN-INTUIT-AGENT","title":"Intuit官方企业Agent产品发布","url":"https://investors.intuit.com/news-events/press-releases/detail/1260/intuit-launches-new-agentic-ai-experiences-and-financial-management-capabilities-for-intuit-enterprise-suite-to-drive-mid-market-business-growth","publishedAt":"2025-07-22","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Accounting Agent、Finance Agent与产品可用性段落","scope":"厂商介绍分类、对账及财务洞察的产品交付。","limitations":"产品自报非独立结果；没有本站账套、错误分母或人工介入完整日志，不使用客户节时百分比。"},{"id":"WU-FIN-ALLIANZ-NEMO","title":"Allianz实际Agent理赔采用说明","url":"https://www.allianz.com/en/mediacenter/news/articles/251103-when-the-storm-clears-so-should-the-claim-queue.html","publishedAt":"2025-11-03","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Nemo works like this / Human-in-the-loop；澳洲低额食品变质险","scope":"自报Agent把覆盖天气反欺诈及金额形成可交最终支付决策的处理包。","limitations":"自报部署，保险公司保留最终支付决定；未公开模型、预算、逐案正确率或可重放产物，不扩大到人身或争议案件。"},{"id":"WU-FIN-ALLIANZ-2026","title":"Allianz 2026采用回顾：区分自动化与Agent","url":"https://www.allianz.com/en/mediacenter/news/articles/260318-responsible-use-of-ai-at-allianz.html","publishedAt":"2026-03-18","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Responsible AI in action：German pet insurance / Australia food spoilage","scope":"厂商分别描述自动宠物理赔与Agent食品理赔，不宜统称同系统。","limitations":"自动宠物处理数字不证明LLM Agent；国家险种、拒赔升级与人工环节不同，无独立产物重放。"},{"id":"WU-FIN-GW-AGENT","title":"Guidewire Qusar Agent发布与准入范围","url":"https://www.guidewire.com/en/about/press-center/press-releases/20260803/guidewire-introduces-qusar-release-to-help-insurers-build-and-control-ai-agents","publishedAt":"2026-08-03","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Policy Change for ProNavigator及*、** availability footnotes","scope":"厂商提供接入保单/理赔/账单工作流的框架和专门Agent。","limitations":"Policy Change为Early Access，Claim Summary/FNOL为Restricted Availability；不是公众无条件可用、完整核保决定或真实承保AI成绩。"},{"id":"WU-FIN-GW-BIND","title":"PolicyCenter 2026.03承保绑定状态方法","url":"https://docs.guidewire.com/cloud/is/202603/cloudapibf/cloudAPI/PolicyCenter/job-types/submissions/c_binding_and_issuing_a_submission_at_the_same_time.html","publishedAt":"Guidewire Cloud 2026.03","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"POST bind-and-issue及前置引用；Palisades2026.03","scope":"产品业务API为绑定与出单的状态操作提供接口。","limitations":"方法文档，不证明Agent已选对承保条件；需商业实例、权限与实际产品配置。"},{"id":"WU-FIN-GW-RESERVE","title":"ClaimCenter2026.03准备金创建、追加和批准","url":"https://docs.guidewire.com/cloud/is/202603/cloudapibf/cloudAPI/ClaimCenter/financials/reserves/c_creating-reserves.html","publishedAt":"Guidewire Cloud 2026.03","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Approval status、Example of adding to reserves与币种","scope":"准备金交易、批准及变动而非仅计算报告的公开方法。","limitations":"估计合理性及持续维护需专业证据；未批准金额不进入准备金线，方法本身非AI试验。"},{"id":"WU-FIN-GW-PAYMENT","title":"ClaimCenter2026.03付款状态创建方法","url":"https://docs.guidewire.com/cloud/is/202603/cloudapibf/cloudAPI/ClaimCenter/financials/check-creating/c_POSTing-checks.html","publishedAt":"Guidewire Cloud 2026.03","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Validation levels、payment types与POST check-sets","scope":"付款请求的前置验证与状态对象可检查。","limitations":"创建支票/支付对象不等于银行实际付讫，需外部回报及实例规则；不是AI成功证据。"},{"id":"WU-FIN-GW-CLOSE","title":"ClaimCenter2026.03关闭与重开方法","url":"https://docs.guidewire.com/cloud/is/202603/cloudapibf/cloudAPI/ClaimCenter/fnol/claims/c_closing_claims.html","publishedAt":"Guidewire Cloud 2026.03","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Closing claims / Reopening a closed claim","scope":"状态关闭受业务校验，存在未关闭exposure等阻断。","limitations":"基础配置示例，不能替代实际机构结案标准或债务履行检查。"},{"id":"WU-FIN-CLAIM-VARIABLES","title":"Leveraging LLMs for Unstructured Claims Data Analysis","url":"https://arxiv.org/pdf/2606.06089","publishedAt":"2026-06-04","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Executive Summary、§2、Appendix B/D","scope":"合成赔案文本到结构化变量与分层分析的概念验证；两位临床评阅者按量表审阅核心变量。","limitations":"合成数据和专家量表不是实赔准确率；分层准备金示范不证明真实准备金充分或寿险适用。"},{"id":"WU-M-AUTOFORM","title":"研究与本科数学陈述自动形式化的可靠评价","url":"https://aclanthology.org/2025.emnlp-main.907/","publishedAt":"2025-11","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Abstract；论文评测设置与BEq+、ProofNetVerif/RLM25","scope":"原作者研究：语义对齐与可编译分开，提供标注集和基准。","limitations":"2025配置不是当下通用成功率；度量近似专家判断，不替代原题核对。"},{"id":"WU-M-SHADOW","title":"ShadowBench v1：语义对齐与编译率分离","url":"https://arxiv.org/html/2608.29270v1","publishedAt":"2026-08-29","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"§3 SA-Pass；§5.1配置；§5.2编译与语义；§6专家一致性","scope":"原作者自动形式化评估及shadow辅助命题法。","limitations":"shadow集合与专家判断仍有误差，非任意研究陈述的完整语义判定；未重跑。"},{"id":"WU-M-ORAGENT","title":"ORAgentBench v1：业务材料到决策产物","url":"https://arxiv.org/html/2606.19787v1","publishedAt":"2026-06-18","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"§3.4 scoring；§4.1 setup、Table2；附录F案例；代码c9eb952435a4352f33daa2a35efe0f8c76d31b28","scope":"107个有人工审查的可执行OR任务，原作者区分约束可行、归一化质量与通过。","limitations":"14个模型/agent配置的任务集结果；模型不同runner，一行三次其余单次；不等于真实运营效果或本任务成功率。"},{"id":"WU-M-SCICODE","title":"SciCode：科学研究编码原始基准","url":"https://github.com/scicode-bench/SciCode/blob/e3158ea011d4235245a547460d3688d7ccbf9900/README.md","publishedAt":"2024（论文）；固定代码2025-10-07","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Introduction / Dataset Creation / Usage；NeurIPS2024原论文","scope":"科学问题到可运行数值代码的原始任务与评估入口。","limitations":"README旧模型低分不能作2026能力上限；后有独立评价缺陷审计，需连读修订。"},{"id":"WU-M-SCICODE-VERIFIED","title":"SciCode-Verified v1：科学验收缺陷审计与匹配重评（含协议例外）","url":"https://arxiv.org/html/2608.04975v1","publishedAt":"2026-08-05","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"§3 matched reevaluation；§4缺陷分类；§6限制；代码ddab4a92f8d80a7113ab946628e994b52354d838","scope":"独立研究者审计科学代码基准，固定比较数据和测量协议。","limitations":"作者自报审计并非本站独立裁定；v1与仓库统计有263/262口径，应避免混用；主任务/子题、带背景/不带背景各有分母。 Table1中两条修订版DeepSeek V4 Preview行为较早harness的三次均值，多数其它行为单次；不能对12行笼统称只变数据。"},{"id":"WU-M-STATQA","title":"StatQA：统计方法选择与适用性","url":"https://github.com/HKUSTDial/StatQA/blob/3d5f6b5b6926600dc952f39d79dfdef82999aeeb/README.md","publishedAt":"2024-05-26","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Overview；Benchmark/Evaluation；NeurIPS2024论文","scope":"原作者以统计任务、列和方法适用性评估LLM；提供数据和脚本。","limitations":"合成构建/元数据先验与有限方法集；历史模型成绩不等于新模型统计全流程能力。"},{"id":"WU-M-SAPAI","title":"SAPAI：临床方案到统计分析计划的验证研究","url":"https://www.medrxiv.org/content/10.64898/2026.03.19.26348626v1","publishedAt":"2026-03-19","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Abstract Methods/Results；Data/Code","scope":"预印本，9个试验方案、3模型生成27份SAP，统计师双评分；描述与统计推理项分开。","limitations":"未同行接收；清单准确性不是整份SAP临床有效或合规裁决；工具github入口本轮未核实可取。"},{"id":"WU-M-ICH-E9","title":"ICH E9与E9(R1)：估计目标和敏感性方法","url":"https://www.ema.europa.eu/en/scientific-guidelines/ich-e9-statistical-principles-clinical-trials","publishedAt":"1998-09-01；R1页面2020-02-18","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Current version；E9(R1)估计目标/敏感性框架","scope":"官方方法指南，支持设计、分析、解释对齐及同估计目标敏感性核验。","limitations":"方法与工作约束，不是AI能力实测；适用法规/试验需另按辖区核实。"},{"id":"WU-M-BIX3","title":"BixBench3 v1：研究尺度生物分析产物","url":"https://arxiv.org/html/2608.25286v1","publishedAt":"2026-08-26","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"§2结果；§3.2限制；§4.1/4.2/4.4；代码d0e0bbb41222335b1b8878f533a58466a3a782dc","scope":"原作者以20研究、138结构化产物评估原始数据到指定方法分析的执行。","limitations":"是按给定方法复现生物研究，不是临床试验设计/疗效判断；任务平均分和产物通过率不混用，原论文也可能有错。"},{"id":"WU-M-DAFNY","title":"DafnyBench v1：程序证明注解补全","url":"https://arxiv.org/html/2406.08467v1","publishedAt":"2024-06-12","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Benchmark construction / evaluation；官方sun-wendy/DafnyBench","scope":"原作者以现有Dafny程序移除提示，再让模型补全并验证。","limitations":"受限程序/固定规格，不等于自动发现业务规格或验证整个生产系统；历史模型不代表今天。"},{"id":"WU-M-AUTOROCQ","title":"AutoRocq v1：程序验证证明义务","url":"https://arxiv.org/html/2511.17330v1","publishedAt":"2025-11-21","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"§4、§5、§6.1、§7；代码24b92683d049aaa9ecf0cd19e247535e762f43b9","scope":"原作者将Frama-C程序验证义务送Rocq证明代理，并给Linux模块案例。","limitations":"641条SV-COMP引理分母不是641程序/完整内核；论文Rocq8.18，当前仓库9.0.0，不能静默继承成绩。"},{"id":"WU-M-CRYPFORM","title":"CrypFormBench v1：密码形式模型能力","url":"https://arxiv.org/html/2606.25561v1","publishedAt":"2026-06-24","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"§3五类任务；§5设置/结果；§6错误与语义；代码b5540654533dbb7a491ece75e50c8da67defc604","scope":"原作者覆盖符号及计算安全模型的解释、生成、补全、转换和修正。","limitations":"可编译和目标一致不等于部署安全；论文700实例/677方案等统计不混成每种任务统一分母。"},{"id":"WU-M-SHANNON-PAPER","title":"ShannonProver v1：密码引理级形式证明","url":"https://arxiv.org/html/2607.02847v1","publishedAt":"2026-07-03","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Abstract；§V evaluation/case studies；AppendixC","scope":"原作者在给定安全模型和引理分解下自动写EasyCrypt证明，含ChaChaPoly1305等案例。","limitations":"部分私有证明用于污染控制；不代表独立设计安全模型、完整部署或任意协议安全。"},{"id":"WU-M-SHANNON-CODE","title":"Shannon Prover固定源码与项目级入口","url":"https://github.com/SkyShannonProver/shannon-prover/blob/60701dbe37ee72075e60e655a076e282efe7485e/README.md","publishedAt":"2026-10-01（固定提交）","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"What does Shannon automate / Install / Project-level proving / Benchmark guide","scope":"源码声明支持游戏/辅助引理分解与证明搜索，提供安装、隔离benchmark及完整文件再检查入口。","limitations":"比v1论文任务更宽的源码能力声明须另验证；Linux严格评估需bubblewrap，未安装或调用。"},{"id":"WU-M-HACL","title":"HACL*固定高保证密码库方法","url":"https://github.com/hacl-star/hacl-star/blob/504c2987452f87fe44bce9b9f12e19d6e051761f/README.md","publishedAt":"2026-04-10（固定提交）","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"HACL* / ValeCrypt / Status","scope":"官方库对限定代码提供内存安全、功能正确与secret independence保证，并指向生产packages。","limitations":"不是AI实测；secret independence只抵抗某些定时侧信道，不能证明任意硬件无泄漏或集成正确。"},{"id":"WU-M-TEACH-RCT","title":"数学辅导随机试验：辅助作答与独立学习分离","url":"https://www.pnas.org/doi/10.1073/pnas.2422633122","publishedAt":"2025-06-25","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Experimental design / Main Results / Limitations；作者稿https://hamsabastani.github.io/education_llm.pdf；2025-08-20更正为作者单位","scope":"原作者高中数学现场试验，GPT-4普通界面与有教学约束界面对比，无工具考试单列。","limitations":"高中、特定课程和短期考试，不是大学数学或长期保持效果；使用正文指定版本而非销售效果。"},{"id":"WU-M-TUTORCOPILOT","title":"Tutor CoPilot v1：真人辅导中的AI建议","url":"https://arxiv.org/html/2410.03017v1","publishedAt":"2024-10-03","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Design / RCT / Results / Limitations","scope":"预注册现场试验考察向导师提供实时AI建议的教学结果及行为。","limitations":"真人使用建议，不是无人辅导；课题掌握不等于多年学习、大学高等数学或生成材料自动正确。"},{"id":"WU-M-KHAN2Y","title":"Khanmigo两学年实验原始工作论文","url":"https://www.nber.org/papers/w35620","publishedAt":"2026-08-17","checkedAt":"2026-10-20","nextReview":"2026-11-20","status":"active","locator":"Abstract；Working Paper35620","scope":"原作者两年18校集群随机试验，报告数学成绩、实际AI使用和参与问题。","limitations":"中学补救练习，效果类似无AI Khan练习；不能把组合方案效果单归因AI，也不能外推大学或停止使用后的保持。"},{"id":"WU-M-MATH-OCT-FIXED","title":"OpenAI十月数学公开产物固定版本与修订","url":"https://github.com/openai/math/tree/1581d4cbdd69a0faf34b2089614fe2a814316a6a","publishedAt":"2026-10-06；固定提交2026-10-12","checkedAt":"2026-10-20","nextReview":"2026-10-27","status":"active","locator":"README L12-L20/L40-L44；history.md L5-L31；lean/ComparatorChallenges/README.md L12-L28","scope":"原发布者公开手稿、部分Lean产物、内部模型及人类整理说明，以及撤回/修订的可追溯版本。","limitations":"各稿验证/新颖状态不同，部分formalization仅支撑引理；未逐稿重放/独立裁定；内部模型不等同公开服务。"},{"id":"WU-M-KHAN-ENTRY","title":"Khanmigo官方产品与角色访问入口","url":"https://www.khanmigo.ai/","publishedAt":"未标独立发布日期；2026-10-20读取","checkedAt":"2026-10-20","nextReview":"2026-11-03","status":"active","locator":"For teachers / For learners；FAQ Can anyone use、classroom access","scope":"官方教师材料/辅导产品及角色、地区/学校访问说明。","limitations":"产品能力声明与可注册入口不证明大学数学或长期学习效果；未注册/调用，实际使用时须核验账户和地区资格。"},{"id":"WU-DEEPSWE-V1-YJS-TRIAL","title":"DeepSWE v1：Yjs冲突检测公开试次","url":"https://deepswe.datacurve.ai/data/v1/trials/yjs-map-conflict-detection__ESgfZ7p","publishedAt":"试次页面未提取可靠首次公开日","checkedAt":"2026-10-21","nextReview":"2026-11-21","status":"active","locator":"Trial ESgfZ7p；claude-opus-4.7 high；基线与新增测试日志；model.patch","scope":"v1 Yjs冲突检测单个外部基准试次；发布者公开补丁和验收日志","limitations":"仅取证抽查，未独立重放；不是已合并PR、生产部署或本站成功率。"},{"id":"WU-FIN-CLAIMS-IMPLEMENTATION","title":"赔案文本分析：作者公开实现与运行说明","url":"https://github.com/mdsight/llm-claims-analysis/tree/a699ccab010698fa0fdb0018981a3ac9e53f40af","publishedAt":"固定Git提交2026-04-26；论文附录另列","checkedAt":"2026-10-21","nextReview":"2026-11-21","status":"active","locator":"README Quick Start；requirements；分析脚本与配置","scope":"论文作者公开分析代码与调用入口的只读核查","limitations":"未安装、调用模型或运行真实赔案；开放代码不等于验证提取质量或部署可用。"}],"toolSignals":[{"id":"TS-OPENHANDS","title":"OpenHands：把不同编程Agent接入持续工作环境","domainIds":["computing"],"workUnitIds":["W-C01","W-C02","W-C03"],"sourceIds":["WU-OPENHANDS"],"claim":"项目提供本地/远程Agent工作台、后端选择和自动化连接。","checked":"固定commit的README、许可证和自动化接口测试已读；旧resolver路径404。主仓库当前负责Canvas，SDK和automation分别承接执行与触发。","support":"编程工作已有组织Agent运行、工作区和自动化的实现路径，可以作为搭建交付环境的候选。","limit":"接口测试存在不证明任务成功；没有在本站启动该工具或使用其云服务。","nextCheck":"固定SDK/Canvas匹配版本，在隔离仓库用一个有失败回归的缺陷和一项新功能检查完整产物、重试、权限与人工接收。","level":"implementation_inspected","checkedAt":"2026-10-19","version":"f3caa2778553662c22715bfb52b5eb6edf975d5b","evidenceStatus":"active"},{"id":"TS-MATHLIB","title":"Lean/Mathlib：证明交付有可执行的检查基础","domainIds":["math"],"workUnitIds":["W-M01"],"sourceIds":["WU-MATHLIB"],"claim":"数学库提供定理、策略、工具链与构建流程。","checked":"已核对固定commit README、lean-toolchain、依赖清单及CI配置；最新主分支标为Lean 4.36.0-rc1。本站代表契约另固定Lean 4.19.0并已编译校准。","support":"形式证明可以把接收条件落实为固定命题的内核检查和公理核对。","limit":"最新主分支检查与本站旧固定版本重放不是同一次运行；库规模不等于AI证明能力。","nextCheck":"工具链升级后用相同锁定命题、参考解、额外公理和命题替换反例重跑；再增加新题型，保留语义人工核对。","level":"implementation_inspected","checkedAt":"2026-10-19","version":"e3e75e86143b43849829fdcd44a39a6b0fe35234","evidenceStatus":"active"},{"id":"TS-RDAGENT","title":"RD-Agent：研究、实现与反馈可以形成循环","domainIds":["finance","analysis"],"workUnitIds":["W-F03","W-A01"],"sourceIds":["WU-RDAGENT"],"claim":"项目声明自动开展量化因子/模型研究，并提供数据科学场景。","checked":"固定commit的README、许可证、factor.py和model.py已读；可见研发循环、执行结果和失败反馈的实现入口。","support":"对金融研究备忘录和可重放分析包提供相邻运行方法：把假设、代码、结果与反馈一并保存。","limit":"量化因子回测不等于财务对账或投资收益；本轮未复现其模型/回测，不采用README收益数字作本站能力结论。","nextCheck":"选择截止时间锁定的公开数据，保留训练/验证/测试分隔、交易成本和全部候选，独立复算结果；与固定脚本比较人工复核时间。","level":"implementation_inspected","checkedAt":"2026-10-19","version":"484776c211e4fbbeef03e0ec00d6bbee7362a4f4","evidenceStatus":"active"},{"id":"TS-COGNITION","title":"Cognition：高比例AI代码提交是采用线索","domainIds":["computing"],"workUnitIds":["W-C01","W-C02"],"sourceIds":["WU-COGNITION"],"claim":"2026-05-27官方文章称，内部工程师提交的代码89%由Devin提交。","checked":"实际读取原始What’s Next段；没有把融资、收入或转述数量视为独立证据。","support":"提示工程师工作可能从逐行编写转向问题定义、并行执行与接收，可纳入未来职责情景的观察清单。","limit":"原文比例不等于89%工程工作自动完成，也未提供本站可复算的分母、窗口及人工复核成本。","nextCheck":"取得明确时间窗与可审计PR样本，区分生成/提交/合并/接收，纳入删除、回滚、人工重写及完整人时，再判断净效果。","level":"claim_checked","checkedAt":"2026-10-19","version":"official-post-2026-05-27","evidenceStatus":"active"},{"id":"TS-ARISTOTLE","title":"Aristotle：形式证明从研究成果走向工具接口","domainIds":["math"],"workUnitIds":["W-M01","W-M03"],"sourceIds":["WU-ARISTOTLE"],"claim":"产品页声明从英文问题或现有Lean项目开始证明与形式化，并可产生库可用代码。","checked":"已读官方产品页能力声明；未调用服务。","support":"证明工作已经有面向使用者的产品候选，可以用本站锁命题与公理验收方案核查其输出。","limit":"产品承诺、推荐语和“数学超级智能”措辞不替代任务结果；自然语言与形式命题可能不一致。","nextCheck":"在已获服务授权和预算后提交锁定命题与已知假命题，下载实际证明，离线内核检查并核对语义、版本、失败和人工修改。","level":"claim_checked","checkedAt":"2026-10-19","version":"product-page-observed-2026-10-19","evidenceStatus":"active"}]}