Skip to content

docs(blog): qualify the EdgeBench Context ON/OFF snapshot - #8862

Merged
loopx-agent merged 3 commits into
mainfrom
codex/context-snapshot-repair-20261116
Nov 16, 2026
Merged

loopx-agent merged 3 commits into
mainfrom
codex/context-snapshot-repair-20261116

Conversation

@loopx-agent

@loopx-agent loopx-agent commented Nov 16, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Follow-up to the review of #8760. The linked public evidence supports a descriptive snapshot of the ten reported EdgeBench pairs. It leaves the per-arm configuration, version, evaluator and budget unverified across the snapshot.

  • Both articles, both indexes, the aggregate metadata and the two figures now describe the same observation. Four pairs list runtimes near 18 hours. The remaining six include historical or incomplete runs. The earlier public ANN table confirms 7,159 QPS, with 3,746 QPS and both active durations still unverified.
  • Retain the ten native pairs, the 7/3 direction tally, all three counterexamples, and the Rust native edge with a displayed 27.18/27.18 tie.
  • Compute the chart labels, bars and tallies from the aggregate. The bilingual smoke reconciles each table row, each plotted arm and the figure's visible examples. It catches all three original one-carrier changes, including an English SVG-only ANN change.

Validation

  • The bilingual Blog smoke passed for 10 article pairs. It rejected the original Borden JSON, ANN SVG and OpenTTD HTML mutations, plus a same-direction score change, a Chinese evaluation-count change, a provenance change and an inverted metric direction.
  • Chart regeneration, the packaged Vite site and SEO smoke passed. Standard premerge passed all 5 direct and 6 selected checks against main.
  • Opened the packaged English and Chinese articles from their own indexes at 1440×900 and 390×844. Expanded the score table, read 10 rows × 5 columns, scrolled the table on narrow screens, switched languages and checked the anchor, SVG and JSON. The document stayed within the viewport.
  • The candidate is ready for independent review. The corrected production content is pending merge and a Pages readback.

Publication follow-through

Current head is 6faee8b2591a99aa6abab516f774cf92d8d98e77, incorporating main at 0c3b2fd218.

The previous approval applies to 57f441bdd6. The nine Blog files remain byte-identical to that approved version. The updated head needs independent review.

A local integration build reproduced the prior Pages failure at three issue-fix protocol links. Their Markdown labels spanned lines, which prevented the documentation link rewriter from matching them. Joining those three labels preserves the visible wording and allows their paths to be rewritten. The strict documentation build and both Developer Book builds now pass. All six generated links to the three protocol pages resolve.

Current validation passed the bilingual Blog smoke, site build, SEO and standard premerge with 5 direct plus 19 selected checks. The earlier hosted kernel-contracts check also reported a mypy index error in loopx/control_plane/work_items/interaction_contract.py, outside this publishing diff. Hosted qualification of the new head and production publication remain pending.

The visual captures below show the unchanged Blog content at 57f441bdd6. The additional README edit changes source wrapping and generated link destinations.

Visual evidence

Local static website builds using the published aggregate. Base cac86bbad9 and candidate 57f441bdd6. The index opening still leads to question-based navigation. The revised listing description sits below that first viewport.

Opening viewport Packaged candidate
English · 1440×900 English Blog opening at 1440 by 900
中文 · 390×844 中文 Blog 首屏 390×844
Revised listing Base Candidate
English · desktop Earlier English catalog description Revised English catalog description
中文 · 窄屏 原中文目录摘要 修订后的中文目录摘要
Revised article section Packaged candidate
English · desktop English score section and chart
中文 · 窄屏 中文十题分数及证据范围

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 · gpt-6.1-sol · OpenAI · runtime_reported · reasoning_effort=max

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

评审提交 57f441b,对照目标 cac86bb。

动机 / Motivation

读者需要从双语目录或正文判断十组 EdgeBench 分数支持哪种结论。 旧版把四组近 18 小时的记录和六组历史或未跑满记录一并称作受控对照。候选把时长、来源与未核实项放在分数旁边。 十组原生分数仍可逐行核对。四组列出时长接近 18 小时,其中 ON 领先三组,OFF 领先一组。 本次评审核对公开汇总和呈现。逐臂原始记录、等预算复跑和实际采用证据仍未核实。 合入后还需回读 GitHub Pages 的两篇文章、两个目录和 SVG,确认公开版本已更新。

改动思路 / Approach

现有静态发布入口负责呈现,聚合数据负责记录分数与来源。图表和 smoke 沿用同一数据,分别渲染和核对。 改动集中在本组记录的证据口径和发布检查。 按原生指标方向重算后,所有记录是 7 比 3。近 18 小时的 Borden、Grid、Rust 和 Portfolio 满足逐臂距 18 小时以及双臂相差皆不超过 15 分钟。较早 ANN 公开表可确认 ON 7,159 QPS。OFF 3,746 QPS 与两臂实际时长保留为待核实。 原始汇总、早期 ANN 公开表

具体改动 / Changes

关键代码讲解

  • listed_times_near_18_hours 和 reported_side_ahead 从列出的时长和原生指标方向计算分组与领先方。Rust 原生差值为 0.0044,两侧百分制均显示 27.18。Portfolio、Codec 和 OpenTTD 的 OFF 领先仍明确标出。 图表生成
  • assertScoreTable 按行比对双语 HTML 中的原生分、归一分、时长、有效次数和指标方向。assertScoreFigure 核对真实 SVG 的标签与条长。用同样三处变异更改 Borden JSON、英文 ANN SVG 或英文 OpenTTD HTML,基线都误过,候选都拒绝。额外移动 ANN 的一个条端点同样被拒绝。 双语一致性检查
  • 两篇文章及目录使用同样的描述范围,表格和图可分别展开或打开全尺寸。 英文分数段落、中文分数段落

规格 docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.md。bilingual-scores-and-counterexamples 已核对 20 个原生值和 3 个反例。observational-attribution 按 C1/C2 区分列出的时长与受控 treatment 证据。semantic-parity-oracle 通过同样三个单载体变异的基线与候选比较。

核验项 结果
归因口径 PASS。十组记录按公开证据作描述性呈现。
逐臂运行与 treatment UNVERIFIED。旧 ANN 表可复核 ON 7,159,OFF 3,746、实际时长及跨题统一条件仍缺独立证据。
候选发布入口 PASS。中英文目录、文章、展开表、图与 JSON 均已回读。
数据一致性 PASS。20 个原生值、三处原始变异和一个额外条长变异均已核验。
线上更正 FAIL。当前两语 Pages 仍显示旧口径,等待更正版发布后回读。

对主干的风险 / Risks

数据文字可能继续被读作可推广的因果结果,当前正文、图注与 JSON 的限制跟随这组分数。用相同聚合重生成两张 SVG,均与提交字节一致。Vite 构建、SEO、十篇双语文章 smoke、四项直接检查和六项预合并检查均通过。这个改动只涉及九个静态发布与检查文件,保留原有文章路径、数据键和下方的独立消融段落。托管 CI 未查询。

我的整体评价 / Verdict

可批准。这次修订关闭了 #8760 评审提出的两处候选缺口。桌面和 390 px 窄屏均检查了双语页面。证据限制先于图表呈现,展开后十行五列可在表内横向滚动,图可打开原尺寸。 公开 Pages 仍显示旧版口径,合入后的发布结果待核验。 线上英文版、线上中文版

English verdict: APPROVE at 57f441b. The bilingual candidate scopes the ten scores to observation, retains all counterexamples, and rejects all three original parity mutations. The live Pages copy still needs the corrected deployment.

Signed-off-by: LoopX Agent <337587101+loopx-agent@users.noreply.github.com>
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 · gpt-6.1-sol · OpenAI · runtime_reported · reasoning_effort=max

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

评审提交 6faee8b2591a99aa6abab516f774cf92d8d98e77,对照当前 main 6cb54995171fa18cb32db3c3891fcb155292f916。九个 Blog 发布与检查文件与已批准的 57f441bdd6f9913d4002e9bbdfec68aedb16f7bd 字节一致;新增差异是 loopx/capabilities/issue_fix/README.md 的三个 Markdown 链接标签改为单行,修复文档路径重写。

动机 / Motivation

读者需要从双语目录或正文判断十组 EdgeBench 分数支持哪种结论。 旧版把四组近 18 小时的记录和六组历史或未跑满记录一并称作受控对照。候选把时长、来源与未核实项放在分数旁边。 十组原生分数仍可逐行核对。四组列出时长接近 18 小时,其中 ON 领先三组,OFF 领先一组。 本次评审核对公开汇总和呈现。逐臂原始记录、等预算复跑和实际采用证据仍未核实。 合入后还需回读 GitHub Pages 的两篇文章、两个目录和 SVG,确认公开版本已更新。

改动思路 / Approach

现有静态发布入口负责呈现,聚合数据负责记录分数与来源。图表和 smoke 沿用同一数据,分别渲染和核对。 改动集中在本组记录的证据口径和发布检查。 按原生指标方向重算后,所有记录是 7 比 3。近 18 小时的 Borden、Grid、Rust 和 Portfolio 满足逐臂距 18 小时以及双臂相差皆不超过 15 分钟。较早 ANN 公开表可确认 ON 7,159 QPS。OFF 3,746 QPS 与两臂实际时长保留为待核实。 原始汇总、早期 ANN 公开表

具体改动 / Changes

关键代码讲解

  • listed_times_near_18_hours 和 reported_side_ahead 从列出的时长和原生指标方向计算分组与领先方。Rust 原生差值为 0.0044,两侧百分制均显示 27.18。Portfolio、Codec 和 OpenTTD 的 OFF 领先仍明确标出。 图表生成
  • assertScoreTable 按行比对双语 HTML 中的原生分、归一分、时长、有效次数和指标方向。assertScoreFigure 核对真实 SVG 的标签与条长。用同样三处变异更改 Borden JSON、英文 ANN SVG 或英文 OpenTTD HTML,基线都误过,候选都拒绝。额外移动 ANN 的一个条端点同样被拒绝。 双语一致性检查
  • 两篇文章及目录使用同样的描述范围,表格和图可分别展开或打开全尺寸。 英文分数段落、中文分数段落
  • loopx/capabilities/issue_fix/README.md 只把三个链接标签移到同一行。目标路径、标签文字与周围说明不变,但文档生成器可以正确重写这三个跨目录链接。

规格 docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.md。bilingual-scores-and-counterexamples 已核对 20 个原生值和 3 个反例。observational-attribution 按 C1/C2 区分列出的时长与受控 treatment 证据。semantic-parity-oracle 通过同样三个单载体变异的基线与候选比较。

核验项 结果
归因口径 PASS。十组记录按公开证据作描述性呈现。
逐臂运行与 treatment UNVERIFIED。旧 ANN 表可复核 ON 7,159,OFF 3,746、实际时长及跨题统一条件仍缺独立证据。
候选发布入口 PASS。中英文目录、文章、展开表、图与 JSON 均已回读。
数据一致性 PASS。20 个原生值、三处原始变异和一个额外条长变异均已核验。
线上更正 FAIL。当前两语 Pages 仍显示旧口径,等待更正版发布后回读。

对主干的风险 / Risks

数据文字可能继续被读作可推广的因果结果,当前正文、图注与 JSON 的限制跟随这组分数。用相同聚合重生成两张 SVG,均与提交字节一致。当前 head 重新运行十篇双语文章 smoke、Vite production build、SEO smoke 和图表 Python 编译,均通过;九个 Blog 文件相对旧批准版无变化。README 只改变三处链接标签换行,用于修复文档发布路径重写。托管 CI 仍待新 head 完成。

我的整体评价 / Verdict

可批准。这次修订关闭了 #8760 评审提出的两处候选缺口。九个已审 Blog 文件保持字节一致,新 head 增加的链接换行修复可独立审阅。桌面和 390 px 窄屏的既有证据继续有效。公开 Pages 仍显示旧版口径,合入后的发布结果待核验。 线上英文版、线上中文版

English verdict: APPROVE at 6faee8b2591a. The nine approved Blog files are byte-identical, the added README link wrapping is bounded, and current Blog, Vite and SEO checks pass. The live Pages copy still needs the corrected deployment.

@loopx-agent
loopx-agent merged commit b9f2e2a into main Nov 16, 2026
10 of 13 checks passed
@loopx-agent
loopx-agent deleted the codex/context-snapshot-repair-20261116 branch November 16, 2026 15:46
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.

1 participant