[songoow]

The native client, local HTTP bridge, Docker transport and verifier wrapper behind a terminal batch_local_counting pilot were retained only in the experiment workspace. This PR preserves their exact five-file source snapshot and provides a read-only identity check, dependencies and known limitations so the published artifacts can be inspected.

Scope and provenance

• Register benchmark/frontiercs2-single-host-v1/ under the existing immutable archive policy. The four Python files and original image manifest keep their original bytes and non-executable Git modes; LF checkout attributes preserve those identities.
• Add the companion archive note and a checker for hashes, modes, Python/JSON syntax and the named no-new-privileges bootstrap. The checker imports no archived module and performs no task, Docker, network or model execution.
• The archive is an inert historical reference. External controller, credential provisioning, native data and runtime setup are documented prerequisites. It is not a newly installed capability or a current runner/portability qualification.

The separate HF release contains 63 train evaluations, four final evaluations, 63 distinct arm/archive pairs, all 112 purpose-scoped stored-payload records and 686 numeric usage observations. All 148 files were downloaded at revision 7adf7b7e6828895dee4393506af80afecc843d37 and matched against local SHA-256 digests; the downloaded relationship validator also passed. Task/patch attribution and licenses are included there. Raw experiment records and candidate outputs are kept out of this Git PR.

The evaluated LoopX revision is e9eda527afd5fcb262ee7c199e1822b1a936cc24; the task export pins d247b3f15b73805e455ff93c6d043d84603ec341. This single-host, one-task/one-run-per-arm pilot used a five-hour cap and earlier final selections. The original board's integrity and score-countability flags remain false, and LoopX treatment fidelity remains pending. Publication does not qualify a matched benefit, official Driver/Management deployment or leaderboard result.

Validation

• Snapshot check: five source files, four Python modules, one JSON manifest and one embedded bootstrap passed; source hashes match the collected original files and HF snapshot.
• Independent mutations: changed bytes, extra source, missing source, symlink substitution and executable-mode drift all reject. A bytecode-only cache remains valid.
• Ruff on the new checker, diff whitespace, bilingual RFC gate and public/private scan passed. Standard premerge executed five direct checks and 19 selected checks; its overall result remains failed because issue-fix-pr-lifecycle-smoke.py raises KeyError: evidence. The same command on immutable main baseline 71fe25b809fb9ea1ef72a02f5be3c1b833ff1b73 fails at the same line with byte-identical stderr after replacing only the checkout prefix. The maintainability diagnostic also reproduces with identical output on that baseline; runtime, application, package and original smoke source bytes are unchanged. The full gate remains failed, with the separate benchmark-sensitive maintainer hold retained.
• The diff-scoped semantic advisory refuses explicit paths outside loopx/; it therefore does not cover this benchmark/ archive. The inherited Frontier client schema/status vocabulary remains local to the frozen compatibility source, with no new LoopX vocabulary registration. The standard full-tree semantic check passed.

Bounded refactoring is unnecessary inside an immutable historical snapshot: changing the original adapters would invalidate provenance. Later execution repairs require a separately versioned implementation and real-path qualification. No graphical UI changes or model/scorer rerun is part of this PR. Maintainer review and merge authority remain separate.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-03T02:22:08.358292Z	f3aee45	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[loopx-agent]

Reviewer: model_agent; gpt-6.1-sol; OpenAI; runtime_reported; reasoning_effort=max

没有发现阻塞该归档切片的缺陷。APPROVE；完整 premerge 的既有失败和 benchmark-sensitive merge hold 仍保留。

动机 / Motivation

检查历史实验的维护者，需要看到当时实际使用的客户端、桥接与裁判包装源码，才能理解已经发布的观察来自什么执行边界。 此前这五个文件只在实验工作区和独立发布中，仓库读者缺少固定版本、依赖说明与可运行的身份检查。 本 PR 把原始源码作为不可变历史快照保留，并提供只读校验；本轮验证了五份文件与固定 HF 发布的逐字节一致性和异常文件拒绝。 本次只解决来源可检查性，不交付当前 runner、可移植沙箱、正式 Driver/Management 部署或实验收益资格。

改动思路 / Approach

复用既有 benchmark archive placement，不增加 capability、产品导入或活跃 CI benchmark。当前边界合理：冻结树只保存原始字节，说明与校验器放在外部，后续运行修复另立版本。这样既保留历史失败，又避免为当前产品长期维护第二套 runner。不保存源码会使历史检查继续依赖外部工作区；把它直接接入产品则会引入已知 controller、token 恢复及沙箱边界问题。

因此选择保留这个显式历史兼容边界，而不是重构冻结源码。只读检查器是局部 helper，负责文件集合、内容身份、执行位和语法；它不拥有评分、运行、权限或发布决策。哈希一致只证明与已发布快照一致，不证明实验当时的运行、分数或效果可信。需要重跑的人仍必须在新的隔离工作区准备外部 controller、数据与授权，不能把归档检查当成执行许可。

具体改动 / Changes

固定 head f3aee4586d87b678581a22d23b61fba7e49d4672，实际 merge base 71fe25b809fb9ea1ef72a02f5be3c1b833ff1b73。完整变更 9 个文件、+1085/-0：五个冻结文件、77 行检查器、外部版本说明、README 登记和仅针对冻结树的 LF 属性。loopx/、apps/、packages/、工作流及打包配置未改；在产品、测试、CI 中没有归档导入或自动执行引用。

关键代码讲解

• check_archive 对固定五文件集合检查 hash、非执行模式和语法；拒绝文件级 symlink、额外源码和缺失源码。它只 compile 四个 Python 模块和命名 bootstrap，不 import 或执行归档程序。
• main 将错误作为 JSON/exit 1 返回，成功也明确保留 benchmark_executed=false、standalone_runnable=false、runtime_validation=not_performed。调用者能区分身份检查和运行验证。
• Trial.request 与 NNP_BOOTSTRAP 属于被冻结的历史实现：bridge 的内存 token/idempotency 不能靠 persist 恢复，bootstrap 依赖 Linux/Docker 的特定安全状态。版本说明没有把这些未提供的 controller 边界冒充为当前可用能力。

规格 spec_ref benchmark/README.md#archive-placement，spec_revision 71fe25b809fb9ea1ef72a02f5be3c1b833ff1b73，按变更前规范的四项条件映射：条件 1 public-safe 源码 implemented，本轮逐文件审查及边界扫描通过；条件 2 外部 provenance、依赖与已知缺陷说明 implemented；条件 3 内容和模式只读检查 implemented，有独立 mutation 对照；条件 4 惰性归档与另版本修复 implemented，无产品/活跃 CI 消费者。规范明确排除当前运行资格，不另加整场实验或父级 roadmap 验收。

对主干的风险 / Risks

独立下载 HF 固定版本的 protocol_snapshot 后，五份文件的全部 42085 字节与当前 Git 文件一致，Git 模式均为 100644。未扩称已复核外部发布中的全部 148 文件、评分、处理忠实度或原始运行。外部发布与 Git 相同仍不是独立历史执行证明。

本轮实际检查	结果
五文件身份、四模块语法、一个嵌入 bootstrap	PASS；未执行归档模块
改字节、额外/缺失源码、文件 symlink、执行位变化、非字节码 cache	PASS：六类 mutation 均 exit 1
原始副本、纯 pyc cache、显式 root locator symlink	PASS；root locator 可由调用者选择，内部文件约束仍执行
每次检查前后文件内容对照	PASS：校验器不写源码
base/head capability list 与 issue-fix help	PASS：完整输出逐字节一致
Ruff、diff/编译/双语 RFC/public-boundary、完整 premerge 的其他选择项	PASS；full gate 仍 FAIL，见下文

standard premerge 的 5 个 direct 检查通过，19 项选中检查中 issue-fix-pr-lifecycle-smoke.py 在读取 gate_item["evidence"] 时失败，benchmark-sensitive hold 另存。相同命令在固定 base/head 独立重跑，完整 stderr 仅替换 checkout 前缀后 511 字节完全相同；整个产品源码与该失败路径无差异，归档身份、拒绝路径和产品入口另有独立通过证据。它是本轮验证的 pre_existing_unrelated，仍为 failed，不能称完整 gate 全绿，也不要求本 PR 修复无关产品代码。

没有加载归档为新 runner，因而没有新增自动启用、quota、Todo、凭据或 actor 生命周期语义。冻结协议值属于历史外部兼容输入，未注册为通用 LoopX 义务。最强残余风险是读者把可检查源码误认作可复现实验；版本说明保留 integrity=false、fidelity pending、countability=false 和 controller 缺失，检查输出也明确不证明 runtime。后续修复须新版本和独立真实路径验证。

我的整体评价 / Verdict

APPROVE，problem_context 为 goal_achieved，仅指这次不可变归档的可检查性；long_horizon 和 user_experience 均 preserved，当前工作路径没有被新门禁或人工步骤阻断。 架构选择 retain：既有归档规范已给出精确兼容边界，冻结树内重构会破坏本次 provenance；bounded refactor 在该树中 unnecessary，新的只读校验保持局部职责。 已检查的是 CLI、Git 字节/模式、目录及文档，没有渲染界面变更，visual_review 不适用。 原生重跑、模型执行、controller 安全和历史效果资格未验证，也不属于本次接受范围。

本次批准不撤销失败观察，不提供合并或 benchmark 执行权限。 精确范围 CQR 也保留 required premerge failure，decision 为 fail；未把它改成 optional 来通过 strict receipt。完整 premerge 与 benchmark-sensitive 人工 hold 继续阻止把这条 review 当作 ready-to-merge 证明；后续由维护者处理现有门禁，不改写历史源码或实验结果。

English verdict: APPROVE — the five-file immutable archive matches the pinned published snapshot, rejects identity/mode mutations, and stays outside product execution. The reproduced pre-existing premerge failure and benchmark merge hold remain separate and unresolved.