# IQuest 历史任务与继续/重跑决策审计

结论：工程与预检继续；正式模型比较从当前冻结配置开始新的一轮，10 题 × 5 harness 全部新建 trial。旧 trial、patch、verifier 和 trace 保留，不续接旧 Agent 会话，不选择性导入旧成功结果。

## 范围与纠正

以本地原始 result.json、config.json、请求日志和沙箱生命周期为准；state.json 只做索引。此前的“30+10”仅覆盖状态文件和兼容烟测，遗漏了未登记原始结果、无 result 的中断目录，以及 jobs/ 下更早的烟测。早期 5 次包括 3 个自定义 integration-smoke 和 2 个官方 statemachine 题；49 个目录不能称为 49 个正式 DeepSWE 评测试次。

| 阶段 | trial 目录 | 有原始 result | reward=1 | reward=0 | 无 reward | 有请求记录 | 有原生 token 用量 |
|---|---:|---:|---:|---:|---:|---:|---:|
| pilot_recovery | 34 | 32 | 6 | 13 | 15 | 34 | 28 |
| compatibility_smoke | 10 | 10 | 0 | 7 | 3 | 8 | 7 |
| early_smoke | 5 | 5 | 0 | 1 | 4 | 2 | 2 |

共 49 个唯一任务尝试目录，其中 47 个有原始 result、44 个有请求记录。6 个 reward=1、21 个 reward=0、22 个无 reward；零分中包含烟测步数截停、传输失败和限流，不能统一归为模型解题失败。
有模型请求记录表示 harness 发起请求，包含错误/本地策略拒绝，不能等同于成功模型生成或服务端计费调用次数。无请求记录不证明服务端从未接收请求，只表示本地证据不足。

## 为什么必须新开一轮

- 旧批次的模型请求首包/间隔/整请求时限与新口径不同（常见 1s/30s/120s），Agent 总时限即使同为 10800s，也不能消除单请求被提前截停的差异。
- 当前 v2 bundle 与旧模型 trial 的实际 runtime SHA 不同或旧记录缺失；旧通过只能说明旧代码/运行包能通过该题。
- 旧批次有重跑、步数截停、模型服务错误和中断；只保留旧成功再补跑失败会产生选择偏差，并破坏成本/时延对比。
- 已有 6 条 reward=1 是实际的独立 verifier 成功证据，应保留；新协议重跑不否定旧成功。旧 reward=0 则必须和模型服务、工程故障、预算终止分别解释。

## 关键历史证据与修正

- 六条成功均有非空 patch，官方 verifier 的 F2P/P2P 全部通过：Codex 的 query、anko、igel、vulture；mini-swe-agent 的 anko；dsh-minimal 的 query。5/6 成功 trial 的请求日志同时记录过 first_chunk 超时，说明它们也受旧请求策略影响，不能只重跑失败项。
- 旧 state 少收录了 pilot-10-gap30 的 scc×Codex（reward=0、空 patch）和 helm×dsh-minimal（无 reward、TLS 异常）；另有 meriyah×mini 和 igel×mini 两个有请求日志却无 result 的中断目录。本地 running 标记是旧文件状态，不证明云端实例当前仍运行。
- mini×igel 的 state 仍写 harness 故障，当前原始证据复核为实现未通过 verifier；其自建验证脚本的 import 错误被模型后续修复，不能用那个瞬时错误判 Harness 缺依赖。
- 429 修复后的 FastAPI×Codex 提交了非空补丁但 reward=0，并记录过首包超时；Arktype×Codex 的 371 条客户端请求记录中末条是本地步数/请求限额拒绝，370 次已放行，空 patch、reward=0。二者都不能称为本次上游 429 耗尽。
- recovery-claude 的 meriyah 已通过 litellm_chat 桥接产生约 1124 万输入 token 和 13,538-byte patch；最终报 Runtime source changed; rebuild before allocating a sandbox，且 state 标记 source_changed_during_trial。另一个 scc×mini 也有运行中源码变更标记。需要冻结执行源码，不能边运行边覆盖 evaluator/bundle。
- Claude Code 直连 IQuest 的原生 Messages 405 与实际评测的 Claude Code→本地 LiteLLM 协议桥→IQuest Chat Completions 是两条路径。历史桥接确实工作过；应验收当前 SHA 的桥接闭环，而不是把直连 405 当作桥接完全不可用。
- 旧 trial-validity 台账会重复计入多处引用的 Oracle，且把负 reward=-1 标为 retain，不能作为本轮正式成绩准入表。本报告按原始 trial 路径去重；逐行规则归因仍是辅助标签，不替代 raw reward、异常和人工因果核对。

## 可复用与待验收

- 复用同一 10 题及 checksum、官方 solution/tests、现有模板、runtime 启动/清理预检、旧 trace 和人工复核结论；不重复下载或构建模板。
- 当前 10 题都有匹配 checksum 的历史 Oracle reward=1（igel 来自 oracle-venv-repair）；此前 oracle-pilot-validated 中的 igel=-1 不可选为通过证据。旧 Oracle 的网络标记为 operator-declared-external，不能据此证明新适配器的创建时禁网策略。
- 正式开始前，用当前适配器对这 10 题做无模型的离线 Oracle 验收；复用已有 solution，不把环境排错交给 IQuest。
- 五种当前 bundle 的启动/清理预检已通过。Claude Code 直连 IQuest 原生协议仍返回 405；应先验证实际正式使用的协议桥接路径，不能用 CLI --help 或旧 bundle 的 smoke 代替。
- 当前配置全部通过端到端验收并冻结后，按固定顺序每个 cell 一次尝试；后续仅在配置及执行源码未变化且没有原始实例/进程时恢复调度未启动项，不恢复已终止的 Agent 会话。
- 新协议预先声明的步数/官方总时限耗尽属于该有界协议的任务结果，需单列原因；若 verifier 为 0，不能通过剔除预算失败来抬高完成率，也不能免费反复尝试直到成功。工程/服务失败则另列，保留完整尝试和成本。

## 逐 trial 台账

| 阶段 | root | task × harness | result | reward | 当前归因 | 请求记录 | patch bytes | v2 复用 |
|---|---|---|---|---:|---|---:|---:|---|
| compatibility_smoke | compat-smoke-20261023T111548Z | [meriyah-explicit-resource-declarations × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/compat-smoke-20261023T111548Z/jobs/meriyah-explicit-resource-declarations--mini-swe-agent/meriyah-explicit-resource-declar__Efg5X7b) | True | None | harness_runtime_dependency_missing | 0 | 0 | 重跑 |
| compatibility_smoke | compat-smoke-20261023T112458Z | [meriyah-explicit-resource-declarations × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/compat-smoke-20261023T112458Z/jobs/meriyah-explicit-resource-declarations--mini-swe-agent/meriyah-explicit-resource-declar__ryiH8Yy) | True | None | unclassified_runtime_failure | 0 | None | 重跑 |
| compatibility_smoke | compat-smoke-20261023T112816Z | [meriyah-explicit-resource-declarations × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/compat-smoke-20261023T112816Z/jobs/meriyah-explicit-resource-declarations--mini-swe-agent/meriyah-explicit-resource-declar__7hPTXk8) | True | 0 | task_execution_budget_exhausted | 3 | 0 | 重跑 |
| compatibility_smoke | compat-smoke-20261023T113131Z | [meriyah-explicit-resource-declarations × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/compat-smoke-20261023T113131Z/jobs/meriyah-explicit-resource-declarations--codex/meriyah-explicit-resource-declar__UD8i8j2) | True | 0 | model_request_limit | 4 | 0 | 重跑 |
| compatibility_smoke | compat-smoke-20261023T113131Z | [meriyah-explicit-resource-declarations × dsh-minimal](/Users/lizhao/Documents/ChatGPT/swe/artifacts/compat-smoke-20261023T113131Z/jobs/meriyah-explicit-resource-declarations--dsh-minimal/meriyah-explicit-resource-declar__KmKbFxE) | True | 0 | model_request_limit | 9 | 0 | 重跑 |
| compatibility_smoke | compat-smoke-20261023T113131Z | [meriyah-explicit-resource-declarations × opencode](/Users/lizhao/Documents/ChatGPT/swe/artifacts/compat-smoke-20261023T113131Z/jobs/meriyah-explicit-resource-declarations--opencode/meriyah-explicit-resource-declar__UCXZumK) | True | 0 | model_request_limit | 9 | 0 | 重跑 |
| compatibility_smoke | compat-smoke-20261023T114914Z | [meriyah-explicit-resource-declarations × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/compat-smoke-20261023T114914Z/jobs/meriyah-explicit-resource-declarations--mini-swe-agent/meriyah-explicit-resource-declar__dk9FmCC) | True | 0 | task_execution_budget_exhausted | 3 | 0 | 重跑 |
| compatibility_smoke | compat-smoke-20261023T120709Z | [meriyah-explicit-resource-declarations × claude-code](/Users/lizhao/Documents/ChatGPT/swe/artifacts/compat-smoke-20261023T120709Z/jobs/meriyah-explicit-resource-declarations--claude-code/meriyah-explicit-resource-declar__QjwzsY9) | True | 0 | model_request_limit | 13 | 0 | 重跑 |
| compatibility_smoke | compat-smoke-20261023T235853Z | [meriyah-explicit-resource-declarations × claude-code](/Users/lizhao/Documents/ChatGPT/swe/artifacts/compat-smoke-20261023T235853Z/jobs/meriyah-explicit-resource-declarations--claude-code/meriyah-explicit-resource-declar__yHUuRvH) | True | 0 | model_request_limit | 14 | 0 | 重跑 |
| compatibility_smoke | compat-smoke-20261024T160052Z | [meriyah-explicit-resource-declarations × claude-code](/Users/lizhao/Documents/ChatGPT/swe/artifacts/compat-smoke-20261024T160052Z/jobs/meriyah-explicit-resource-declarations--claude-code/meriyah-explicit-resource-declar__Ks9JiTD) | True | None | model_request_limit | 7 | None | 重跑 |
| pilot_recovery | pilot-10-final | [anko-typed-variable-bindings × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-final/jobs/anko-typed-variable-bindings--mini-swe-agent/anko-typed-variable-bindings__vwi3FJg) | True | 0 | request_chunk_gap_timeout | 20 | 0 | 重跑 |
| pilot_recovery | pilot-10-final | [helm-unified-manifest-stream × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-final/jobs/helm-unified-manifest-stream--mini-swe-agent/helm-unified-manifest-stream__5ViYzm9) | True | 0 | request_chunk_gap_timeout | 25 | 0 | 重跑 |
| pilot_recovery | pilot-10-final | [igel-persist-feature-schema × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-final/jobs/igel-persist-feature-schema--mini-swe-agent/igel-persist-feature-schema__afSxt4r) | False | None | request_chunk_gap_timeout | 11 | None | 重跑 |
| pilot_recovery | pilot-10-final | [meriyah-explicit-resource-declarations × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-final/jobs/meriyah-explicit-resource-declarations--mini-swe-agent/meriyah-explicit-resource-declar__P38ZpJ7) | True | 0 | request_chunk_gap_timeout | 29 | 0 | 重跑 |
| pilot_recovery | pilot-10-final | [query-persist-restored-query-state × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-final/jobs/query-persist-restored-query-state--mini-swe-agent/query-persist-restored-query-sta__FQtcUov) | True | 0 | request_chunk_gap_timeout | 7 | 0 | 重跑 |
| pilot_recovery | pilot-10-gap30 | [anko-typed-variable-bindings × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-gap30/jobs/anko-typed-variable-bindings--codex/anko-typed-variable-bindings__SeWobyx) | True | 1 | passed | 76 | 107942 | 重跑 |
| pilot_recovery | pilot-10-gap30 | [anko-typed-variable-bindings × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-gap30/jobs/anko-typed-variable-bindings--mini-swe-agent/anko-typed-variable-bindings__eY7vtaC) | True | None | unclassified_runtime_failure | 76 | None | 重跑 |
| pilot_recovery | pilot-10-gap30 | [fastapi-deprecation-response-headers × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-gap30/jobs/fastapi-deprecation-response-headers--codex/fastapi-deprecation-response-hea__emqtE9A) | True | None | unclassified_runtime_failure | 111 | None | 重跑 |
| pilot_recovery | pilot-10-gap30 | [helm-unified-manifest-stream × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-gap30/jobs/helm-unified-manifest-stream--codex/helm-unified-manifest-stream__NeQ5GyN) | True | None | unclassified_runtime_failure | 201 | None | 重跑 |
| pilot_recovery | pilot-10-gap30 | [helm-unified-manifest-stream × dsh-minimal](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-gap30/jobs/helm-unified-manifest-stream--dsh-minimal/helm-unified-manifest-stream__hR6kuEm) | True | None | unclassified_runtime_failure | 213 | None | 重跑 |
| pilot_recovery | pilot-10-gap30 | [helm-unified-manifest-stream × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-gap30/jobs/helm-unified-manifest-stream--mini-swe-agent/helm-unified-manifest-stream__abSFyQz) | True | None | unclassified_runtime_failure | 197 | None | 重跑 |
| pilot_recovery | pilot-10-gap30 | [igel-persist-feature-schema × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-gap30/jobs/igel-persist-feature-schema--codex/igel-persist-feature-schema__HW6vcHu) | True | 1 | passed | 74 | 23010 | 重跑 |
| pilot_recovery | pilot-10-gap30 | [igel-persist-feature-schema × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-gap30/jobs/igel-persist-feature-schema--mini-swe-agent/igel-persist-feature-schema__E8ovpoa) | True | 0 | implementation_failed_verifier | 135 | 44698 | 重跑 |
| pilot_recovery | pilot-10-gap30 | [meriyah-explicit-resource-declarations × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-gap30/jobs/meriyah-explicit-resource-declarations--codex/meriyah-explicit-resource-declar__7mj4tNz) | True | None | provider_rate_limit | 95 | None | 重跑 |
| pilot_recovery | pilot-10-gap30 | [meriyah-explicit-resource-declarations × dsh-minimal](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-gap30/jobs/meriyah-explicit-resource-declarations--dsh-minimal/meriyah-explicit-resource-declar__tTRf7SR) | True | None | unclassified_runtime_failure | 17 | None | 重跑 |
| pilot_recovery | pilot-10-gap30 | [meriyah-explicit-resource-declarations × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-gap30/jobs/meriyah-explicit-resource-declarations--mini-swe-agent/meriyah-explicit-resource-declar__jQAbcSL) | False | None | ungraded_or_incomplete_trial | 54 | None | 重跑 |
| pilot_recovery | pilot-10-gap30 | [query-persist-restored-query-state × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-gap30/jobs/query-persist-restored-query-state--codex/query-persist-restored-query-sta__hFpPo3N) | True | 1 | passed | 126 | 22866 | 重跑 |
| pilot_recovery | pilot-10-gap30 | [query-persist-restored-query-state × dsh-minimal](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-gap30/jobs/query-persist-restored-query-state--dsh-minimal/query-persist-restored-query-sta__nzRdhEi) | True | 1 | passed | 377 | 26708 | 重跑 |
| pilot_recovery | pilot-10-gap30 | [query-persist-restored-query-state × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-gap30/jobs/query-persist-restored-query-state--mini-swe-agent/query-persist-restored-query-sta__bfVkKE8) | True | None | no_progress | 81 | 0 | 重跑 |
| pilot_recovery | pilot-10-gap30 | [scc-bounded-memory-spilling × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/pilot-10-gap30/jobs/scc-bounded-memory-spilling--codex/scc-bounded-memory-spilling__UjTW2Fv) | True | 0 | missing_submission_without_terminal_evidence | 26 | 0 | 重跑 |
| pilot_recovery | recovery-claude-gap30 | [meriyah-explicit-resource-declarations × claude-code](/Users/lizhao/Documents/ChatGPT/swe/artifacts/recovery-claude-gap30/jobs/meriyah-explicit-resource-declarations--claude-code/meriyah-explicit-resource-declar__HJscCCc) | True | None | unclassified_runtime_failure | 144 | 13538 | 重跑 |
| pilot_recovery | recovery-codex-gap30 | [arktype-json-schema-refs-dependencies × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/recovery-codex-gap30/jobs/arktype-json-schema-refs-dependencies--codex/arktype-json-schema-refs-depende__4ttt7Le) | True | 0 | provider_rate_limit | 253 | 0 | 重跑 |
| pilot_recovery | recovery-codex-gap30 | [fastapi-deprecation-response-headers × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/recovery-codex-gap30/jobs/fastapi-deprecation-response-headers--codex/fastapi-deprecation-response-hea__VCSUGA6) | True | 0 | provider_rate_limit | 88 | 0 | 重跑 |
| pilot_recovery | recovery-codex-gap30 | [helm-unified-manifest-stream × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/recovery-codex-gap30/jobs/helm-unified-manifest-stream--codex/helm-unified-manifest-stream__gcV34Kd) | True | None | unclassified_runtime_failure | 56 | None | 重跑 |
| pilot_recovery | recovery-codex-gap30 | [katex-multicolumn-array-spans × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/recovery-codex-gap30/jobs/katex-multicolumn-array-spans--codex/katex-multicolumn-array-spans__Z3qi36y) | True | 0 | missing_submission_without_terminal_evidence | 19 | 0 | 重跑 |
| pilot_recovery | recovery-codex-gap30 | [meriyah-explicit-resource-declarations × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/recovery-codex-gap30/jobs/meriyah-explicit-resource-declarations--codex/meriyah-explicit-resource-declar__CogcRrr) | True | 0 | implementation_failed_verifier | 141 | 23204 | 重跑 |
| pilot_recovery | recovery-codex-gap30 | [vulture-persistent-analysis-cache × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/recovery-codex-gap30/jobs/vulture-persistent-analysis-cache--codex/vulture-persistent-analysis-cach__fxm853K) | True | 1 | passed | 51 | 36457 | 重跑 |
| pilot_recovery | recovery-mini-gap30 | [anko-typed-variable-bindings × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/recovery-mini-gap30/jobs/anko-typed-variable-bindings--mini-swe-agent/anko-typed-variable-bindings__mMfbE2j) | True | 1 | passed | 68 | 108983 | 重跑 |
| pilot_recovery | recovery-mini-gap30 | [fastapi-deprecation-response-headers × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/recovery-mini-gap30/jobs/fastapi-deprecation-response-headers--mini-swe-agent/fastapi-deprecation-response-hea__kbva48p) | True | 0 | implementation_failed_verifier | 121 | 76373 | 重跑 |
| pilot_recovery | recovery-mini-gap30 | [helm-unified-manifest-stream × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/recovery-mini-gap30/jobs/helm-unified-manifest-stream--mini-swe-agent/helm-unified-manifest-stream__Mzorvvw) | True | None | unclassified_runtime_failure | 27 | None | 重跑 |
| pilot_recovery | recovery-mini-gap30 | [meriyah-explicit-resource-declarations × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/recovery-mini-gap30/jobs/meriyah-explicit-resource-declarations--mini-swe-agent/meriyah-explicit-resource-declar__CJmysK5) | True | None | request_first_chunk_timeout | 184 | None | 重跑 |
| pilot_recovery | recovery-mini-gap30 | [scc-bounded-memory-spilling × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/artifacts/recovery-mini-gap30/jobs/scc-bounded-memory-spilling--mini-swe-agent/scc-bounded-memory-spilling__saVi6Dx) | True | None | request_total_timeout | 14 | 0 | 重跑 |
| pilot_recovery | retry-429-codex-gap30 | [arktype-json-schema-refs-dependencies × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/retry-429-codex-gap30/jobs/arktype-json-schema-refs-dependencies--codex/arktype-json-schema-refs-depende__8pM27eZ) | True | 0 | model_request_limit | 371 | 0 | 重跑 |
| pilot_recovery | retry-429-codex-gap30 | [fastapi-deprecation-response-headers × codex](/Users/lizhao/Documents/ChatGPT/swe/artifacts/retry-429-codex-gap30/jobs/fastapi-deprecation-response-headers--codex/fastapi-deprecation-response-hea__CoLu4pz) | True | 0 | implementation_failed_verifier | 120 | 63231 | 重跑 |
| early_smoke | iquest-aliyun-20261022T153021Z | [local/aliyun-iquest-integration-smoke × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/jobs/iquest-aliyun-20261022T153021Z/smoke-httpx__SaWdZeM) | True | None | unclassified_runtime_failure | 0 | None | 重跑 |
| early_smoke | iquest-aliyun-20261022T153405Z | [local/aliyun-iquest-integration-smoke × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/jobs/iquest-aliyun-20261022T153405Z/smoke-httpx__xWiSdEL) | True | None | unclassified_runtime_failure | 0 | None | 重跑 |
| early_smoke | iquest-aliyun-20261022T153850Z | [local/aliyun-iquest-integration-smoke × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/jobs/iquest-aliyun-20261022T153850Z/smoke-httpx__DehdchD) | True | None | unclassified_runtime_failure | 0 | None | 重跑 |
| early_smoke | smoke-iquest-20261022T155225874840Z | [python-statemachine-state-data-scoping × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/jobs/smoke-iquest-20261022T155225874840Z/python-statemachine-state-data-s__DjEcaom) | True | None | unclassified_runtime_failure | 13 | None | 重跑 |
| early_smoke | smoke-iquest-20261022T155849224969Z | [python-statemachine-state-data-scoping × mini-swe-agent](/Users/lizhao/Documents/ChatGPT/swe/jobs/smoke-iquest-20261022T155849224969Z/python-statemachine-state-data-s__csZgNof) | True | 0 | task_execution_budget_exhausted | 144 | 0 | 重跑 |

逐条 SHA、原始结果路径、网络声明、配置差异、请求错误与旧状态归因差异见同名 JSON。此审计不创建评测沙箱、不发送模型请求。