Skip to content

docs(research): define episode evaluation and evidence-based progress reading - #6573

Merged
huangruiteng merged 1 commit into
mainfrom
codex/research-progress-method-20261016
Oct 16, 2026
Merged

huangruiteng merged 1 commit into
mainfrom
codex/research-progress-method-20261016

Conversation

@loopx-agent

Copy link
Copy Markdown
Collaborator

Flat native scores do not explain whether later work tested a different mechanism, reduced a decision-relevant uncertainty, or used an earlier result. This proposal adds research-episode analysis to the existing benchmark measurement RFC and connects it to a bilingual progress-reading design: one selection links the original outcome window, branch, versioned evidence, check and observed adoption.

The method separates native outcomes, information, discrimination, actual adoption, repetition and costs/coverage. It specifies denominators and missing-data bounds, baseline observability, prospective late/common windows, continuation costs and independent calibration. Existing product metrics keep their definitions; the stricter proposed adoption measure is distinguished from communication/reuse signals. Roadmap S5/S11, Explore and workspace owners link to the same method and synthetic example, preserving existing scoring, evidence and presentation authority.

Validation:

  • RFC generated status index check and docs-governance smoke passed.
  • All local links in the 14 changed/new documents checked: 522 targets and 87 Markdown anchors; fenced blocks balanced.
  • Source-checkout public-boundary scan: 0 errors, 0 warnings. Staged diff and DCO checked.
  • Manual English/Chinese semantic and evidence-boundary review completed. The existing measurement owner centralizes the method; no new graph, ledger or runtime vocabulary is introduced.

This is a design proposal. Mermaid rendering, two independent scientific coders, held-out model-judge calibration and the packaged frontend/real-backend reading journey are not qualified here. The synthetic diagram is illustrative. Design validation establishes neither installed behavior nor causal benefit.

… reading

Signed-off-by: LoopX Agent <337587101+loopx-agent@users.noreply.github.com>

@loopx-agent loopx-agent left a comment

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

Approval conclusion (author-owned PR; GitHub blocks formal self-approval). Exact head 9e6b53f320411237b0f4a3249d72ceefb1bb2187;独立核对的 merge-base 为 a99cb07addfc163fda9435c6455cc9462b657855。

动机

需要判断长任务是否有效推进、是否值得继续的研究者,以及阅读既有 Goal 和团队进度的所有者。

原规范已有原生结果、成本与四臂对照,但没有完整的 episode 读数口径;本提案补齐分母、未知、重叠时间和实际采用的区分,并给出打开已有 Goal 后先看变化、再看分支、按需查证据的设计。

交付是可独立评审的研究读数与阅读设计:原生结果不被信息增益补偿,实际采用必须有独立证据,缺失与历史版本不显示为零或当前通过。

本 PR 只有文档,不实现 sidecar、judge、progress 页面或新状态,不执行实验,不证明产品可用性、科学显著性、实际采用或更大路线图完成。

改动思路

既有 benchmark-toolkit 保持实验与证据 owner,既有 Goal/团队投影保持产品 owner;episode 仅为只读分析单元,避免新建一套运行图、评分器或 closure。 14 个 Markdown 文件统一研究口径、英文/中文阅读设计和既有 RFC 索引;无运行代码、配置开关、调度规则或权限变化。 先固定实验问题、分段与分母,再观察结果,才能避免按结果挑选“有效回合”。产品侧沿用已经打开的 Goal、团队分支和版本证据,把全局变化、分支解释与必要证据放在一个阅读层级中;没有新增运行图、评分器或自动执行权限。

具体改动

规格依据是改动前的 docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.md,spec_revision a99cb07addfc163fda9435c6455cc9462b657855,重点读了原 claim、四臂对照、测量、证据访问与里程碑条款。以下判断针对本次文档结果,不用作者新写的规范反过来证明运行实现。

  • native-outcome implemented:原生任务结果、冻结 scoring/task/budget 与 C0–C4 资格保留。信息与采用的向量读数不能抵消 native failure,挑选例子不能建立因果提升。
  • fair-evidence implemented:episode 可跨 Turn;分段在 arm/outcome 解盲前固定,未完成、恢复和未分配成本仍保留。已知分母的缺失数据区间明确不是置信区间;分母未知不显示比例。实际采用须有后续使用证据,引用、叙述、剪枝或没有采用机会都不能直接算采用。
  • evaluator-boundary implemented:保留原 §6.5 的反馈模式,区别允许的 native feedback 与受保护 evaluator 材料。后者必须同时满足 solver 已结束且 scoring 已完成,只能作只读分析,不能回流 solver;没有悄悄禁止原任务允许的反馈,也没有放松隐藏证据边界。
  • staged-reading implemented:新中英设计明确为 proposal,单一焦点曲线、小分支与上下文证据按需展开,拒绝卡片墙和虚假统一百分比。缺失、受限、版本变化、分歧、刷新返回及无事件都有可读出口;窄屏、键盘和 reduced motion 是后续验收义务。

两语言完整对照后,原生结果/信息/实际采用、缺失与版本、权限以及阶段声明一致。合成 Mermaid 只解释负证据如何剪枝、后续校验如何关联当前版本,没有把虚线缺边表示为采用。RFC 索引提到的部分团队 inspection 已在基线的既有实现和 #4828 记录中存在;当前文档没有把新 progress 设计宣称为已发货,也没有让 L1/R2 等待本研究。

对主干的风险

独立运行 RFC status index 检查、docs governance smoke、diff whitespace 检查均通过。额外逐一解析 14 个改动文档的 525 个本地目标(包括同页引用)和 88 个标题锚点,无失效链接,围栏闭合;候选公开文件的本机路径、私有状态路径和明显凭据扫描无命中,单提交具备 DCO。数量与作者范围可能不同,以上只报告本次独立检查口径。

最强反例不是文档能否生成,而是 native result 变差时用信息增益粉饰进展。本次独立核对了该反例、未知 N、已知 N 的未知子集、没有采用机会、并行 wall-time 区间、晚到反馈与不匹配 baseline:条款均保留保守结论。历史产物校验不能把改变后的字节显示为当前通过;证据受限不能推断无进展,也不授予新访问权。

这些检查不证明独立标注员一致性、judge 校准、实验效果或真实用户理解;未渲染 Mermaid,也没有运行打包 UI/backend、模型或 benchmark。原文已把这些列为后续实现与科学资格,因此这里不是用 skipped test 代替本次必须的运行验收。未查询、轮询、等待或触发 hosted CI。

我的整体评价

APPROVE,限于本次研究与阅读设计阶段。长期推进维度 improved:固定分段、分母、反例和继续观察条件,使负证据与未完成成本可见;用户体验维度 improved:提出从既有 Goal 一次进入、上下文查证和返回的具体短旅程,尚不认证实现可用性。既有 benchmark-toolkit 保持实验与证据 owner,既有 Goal/团队投影保持产品 owner;episode 仅为只读分析单元,避免新建一套运行图、评分器或 closure。 14 个 Markdown 文件统一研究口径、英文/中文阅读设计和既有 RFC 索引;无运行代码、配置开关、调度规则或权限变化。

有界前瞻整理已体现在复用 sidecar 与原产品投影,未发现需要在本 PR 添加代码抽象的相邻边界。下一实现应证明真实入口、失败恢复和阅读理解;科学资格仍依原研究协议独立验证。本评审不授予合并权限,也不结算任何父目标。

English verdict: APPROVE — exact head 9e6b53f320411237b0f4a3249d72ceefb1bb2187. This is a coherent documentation stage: preserve native outcomes, frozen comparisons and protected evaluator access; distinguish information, actual adoption, unknown coverage and versioned evidence; reuse existing research and Goal/team owners. Independent old/new and bilingual counterexamples, RFC/docs governance, 525 local targets and 88 anchors pass. Runtime/UI implementation, rendering, calibration and scientific effects remain unproven; no hosted CI or merge authority is implied.

@huangruiteng
huangruiteng merged commit cd9a756 into main Oct 16, 2026
23 of 24 checks passed
@huangruiteng
huangruiteng deleted the codex/research-progress-method-20261016 branch October 16, 2026 05:17
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants