[zengchuang-hw]

Refs #371

Delivery slice

Wire Apps offline assessment to a fixed Harness record revision, the frozen protocol, and the Benchmark assess, validate, and encode ports. Limit evidence reads to the selected revision and its declared artifacts. Check the primary claim before assessment, preserve the frozen claim order, archive the canonical assessment, and publish the standard result. This slice builds on the mainline deliveries in #432 and #433.

Validation

On main plus these commits, 231 focused tests passed and 2 platform-specific tests were skipped. The test set covers Apps, the module import gate, RunStore, and Harness execution. Ruff and the diff check passed.

Regression coverage includes reassessing an old SQLite revision in a new process, primary-claim exit codes, foreign evidence rejection, assessment readback, and preserving only committed references after publication failure.

Remaining integration

The production Benchmark ports and CPU-task acceptance remain tracked in #434. This PR validates the Apps boundary with injected fixture ports.

[zengchuang-hw]

@amy-why-3459 @Dong1017 @david6666666 请复审 #371 的离线评测装配。当前提交 8bee52db，CI 已通过，基于 #432 和 #433 已合入的接口。请重点核对固定修订读取、评测前的主 claim 校验及两级归档失败时保留的引用。若通过，请对当前提交正式 Approve，首个非空行单独写 LGTM 并记录 PASS。

[amy-why-3459]

复审范围：base 1b88811870d583469437c8a28ae3f8a35d54da10（merge base）…head 8bee52db8cef67091d42db965bb3c37ab0180494，全量审阅；CI 7 项均已通过。

已核对、没有问题的部分：

• 固定修订读取：record_ref 的 attempt/record_version 由 to_evidence 绑定，test_wrong_record_version_is_rejected_before_assessing 和新进程重评 SQLite 旧修订的用例都覆盖到了。
• V34(a)：未知主 claim 键在调用 assessor 前返回 InvalidRequest / primary_claim（src/inferrsi/apps/api.py L337–L344）。
• V34(b)：主 claim 缺失会在归档前被 assessment_result(...) 拦下（L391–L397）。
• V34(c)：两级发布失败只保留已提交的引用。
• check_assessment_identity 去掉 location 的 ArtifactRef 路径，与 schema 中 EvaluationResult 声明为 ArtifactRef 的 9 个字段路径一一对应，符合投影 P。

暂不 Approve，有一个问题需要确认：非主 claim 被省略时，结果仍会被归档成成功。

• src/inferrsi/apps/api.py L380–L385 只检查返回的 claim 键是冻结协议 claim_definitions 的有序子序列（[key for key in claim_keys if key in returned_set]），不要求覆盖全部键。
• 举例：主 claim 为 C 时，如果 Benchmark 少返回了 F，这一行检查能通过；validate_assessment 只拿到 assessment、拿不到协议，也无法判断是否完整；最后这份不完整的评测会被写进 assessment.json 并以 0/8/9 退出。
• 合约 docs/contracts/schema-and-semantics.md §3 写的是 “The order of claim_definitions is the order of result claims”，V34(c) 的前提也是 “a legal complete result”。我的理解是：返回的 claims 应与协议中的全部键完全一致（returned_keys == claim_keys），省略任何一个都应按 AssessorError / assessor_response_validation 处理。
• 如果这里是有意允许子集（例如完整性由别处保证），请指出在哪里保证；否则建议改成全等校验，并补一个“省略非主 claim”的用例。现在 omitted_claim 用例删的是 claims[1]，而 fixture 的主 claim 正好是 C，所以只覆盖了“主 claim 缺失”这一种情况。

非阻塞：Benchmark 抛出的 PublicError（L355–L368）生成的结果里 primary_claim_id / primary_scope 为 null，但此时请求本身已校验合法。合约允许这样做，只是带上这两个字段会更便于排查。

[zengchuang-hw]

@amy-why-3459 已在 df748487 修复。Apps 现在要求返回的 claim 键与冻结协议完整序列逐项一致。新增漏掉非主 claim 的回归，旧实现实测返回 9，修复后返回 AssessorError/2 且不写评测归档。Windows 定向回归 232 项通过、2 项平台测试跳过，Ruff 与差异检查通过。@Dong1017 @david6666666 请和 Amy 一起复审当前提交，重点核对 V34 的主 claim 校验、完整 claim 序列和失败时保留的引用。如果通过，请对当前提交正式 Approve，首个非空行单独写 LGTM，并记录 PASS。

[amy-why-3459]

LGTM

PR review verdict — PASS

增量复审：8bee52db8cef67091d42db965bb3c37ab0180494…df74848780c80dfdc26a79c3eb6dfd814536d102（上次已对 1b888118…8bee52db 全量审阅）；当前 head 上 CI 7 项均已通过。

• 上次的问题已修复：src/inferrsi/apps/api.py L383 改为 returned_keys != claim_keys 全等校验，返回的 claims 必须与冻结协议 claim_definitions 的完整有序序列一致，省略任何 claim 都会在归档前按 AssessorError / assessor_response_validation 拒绝，符合 schema-and-semantics.md §3 与 V34(c) 的“完整结果”前提。
• tests/test_apps_assess.py 新增 omitted_non_primary_claim（删除 claims[0]，主 claim C 保留），断言 exit 2、AssessorError、assessment_ref is None 且不写 assessment.json，覆盖了上次指出的缺口；原有 omitted_claim / duplicate_claim / reordered_claims 用例在全等校验下仍成立。
• 本次增量只改这两处，未引入其他行为变化。上次提到的 PublicError 结果中 primary_claim_id / primary_scope 为 null 属于非阻塞建议，不影响通过。

[zengchuang-hw]

@Dong1017 @david6666666 df748487 已通过 CI，Amy 已完成增量复审并正式通过。请复核冻结 claim 清单校验和固定修订读取，当前需要第二份有效审批。

[cursor]

LGTM

PR review verdict — PASS
Reviewed base 1b88811870d583469437c8a28ae3f8a35d54da10 and head df74848780c80dfdc26a79c3eb6dfd814536d102. Governing #371 Apps offline assessment slice. Independent via cursor[bot]; not the PR author (zengchuang-hw).

Verified

• Apps assess binds fixed record_ref → prepared/protocol verify → primary claim membership → injected Benchmark assess/validate/encode → claim-order equality (returned_keys != claim_keys) → archive → publish; foreign artifacts rejected via AssessmentArtifacts.
• tests/test_apps_assess.py covers primary exit codes, omitted/reordered/duplicate claims (incl. omitted_non_primary_claim), foreign evidence, SQLite revision reassess, and publication failure retaining only committed refs. Local: 20 passed; ruff clean on changed Apps files.
• Hosted Ready checks on this head: full-check 3.10/3.11, installed-package 3.10/3.11, declarations, pr-references, dashboard-browser, knowledge-model-policy — success. Amy APPROVE 5904832055 already on this head.
• Remaining production Benchmark ports stay in #434 as disclosed. No merge by this review.

[Dong1017]

LGTM

PR review verdict — PASS
Reviewed base 1b88811870d583469437c8a28ae3f8a35d54da10 and head df74848780c80dfdc26a79c3eb6dfd814536d102. Governing #371 Apps offline assessment assembly. Independent review as @Dong1017 (write-eligible); not the PR author (zengchuang-hw). Amy APPROVE 5904832055 and cursor[bot] APPROVE 5907086609 already on this exact head — this vote is the second ruleset-eligible write-account Approve.

Verified

• Apps assess binds fixed record_ref → prepare/protocol verify → primary membership → injected Benchmark assess/validate/encode → claim-order equality (returned_keys != claim_keys) → archive → publish; foreign artifacts rejected via AssessmentArtifacts.
• Amy’s prior blocker is fixed on this head: full claim-sequence equality plus omitted_non_primary_claim regression (AssessorError/2, no assessment.json).
• Local: tests/test_apps_assess.py 20 passed; ruff clean on changed Apps files. Hosted Ready checks on this head: full-check 3.10/3.11, installed-package 3.10/3.11, declarations, pr-references, dashboard-browser — success. mergeable_state=clean.
• Production Benchmark ports remain in #434 as disclosed. No merge by this review.

[linyueqian]

Follow-up after the merge. AssessmentArtifacts limits an assessor to the selected attempt, but only one of its attempt filters has a test. On main 5db1bc94 I removed each line below in turn, and all 180 tests in tests/test_apps_*.py still pass. Removing the tested checks filter (artifact_reader.py L73) makes one test fail, so the method does catch a removal.

1. The attempt filter on artifacts and bound_inputs at src/inferrsi/apps/artifact_reader.py L66.
2. The attempt filter on step artifacts at artifact_reader.py L70.
3. The target_binding branch at artifact_reader.py L76-L78. No test reads an environment identity.
4. The binding check at src/inferrsi/apps/api.py L378. A returned assessment with a changed record_ref.record_version, candidate or task would cover it.
5. The pre-archive selected-claim check at api.py L392.

I found no blocking defect in the merged head df748487. These five tests could go into one follow-up PR.