Skip to content

docs(blog): add empirical Context/Stride ablation and Astra 18h breakthrough evidence - #7968

Merged
huangruiteng merged 3 commits into
mainfrom
codex/blog-empirical-refinement-20261103
Nov 3, 2026
Merged

huangruiteng merged 3 commits into
mainfrom
codex/blog-empirical-refinement-20261103

Conversation

@loopx-agent

@loopx-agent loopx-agent commented Nov 3, 2026 •

Copy link
Copy Markdown
Collaborator

Goal And Delivered Outcome

Following v2.2.0 release verification and the late-October / early-November EdgeBench multi-arm empirical studies (#7311, #7759, #7800, #7834, #7959, #7960), the bilingual engineering blog (apps/presentation/site/public/blog/) needed three concrete updates:

  1. Context Delivery (blog/context-delivery/ & blog/zh/context-delivery/):
    • Remove the redundant context-native-path-v10 architecture diagram (which duplicated the decision-<id>/ file tree and clipped dense CLI lines at article width) and retain the clean four-file path tree (projection.json, required-context.json, delivery.json, *.reader.json).
    • Add <h3 id="ablation-evidence"> (Real-Model and Cross-Task Ablations: Value Chain and Boundary Counterexamples) with paired bilingual SVG charts (context-delivery-ablation-zh.svg / context-delivery-ablation-en.svg) and a 9-row empirical table covering:
      • EdgeBench ANN Vector Search (gpt-6.1-sol / xhigh, 18h, Recall@10 >= 0.95): 6-arm progression (Official 5,753 QPS @ 1.05h, Single 5,706 QPS, Native-Goal 4,372 QPS, Explore · Context off 5,766 QPS, Explore · Context on 6,020 QPS, Explore · Context on + Stride 7,159 QPS (+24.4% vs Official) via QT_8bit_direct (HNSW-SQ8)).
      • Layered Byte & Host Input-Token Accounting: -74.25% delta selected bytes (3.09 MB -> 0.80 MB), -38.54% 46-event paired selected bytes (3,089,841 B -> 1,898,889 B), -25.75% Stride selected bytes, -12.07% full-18h uncached host input tokens (8.57M -> 7.53M), -7.91% 16h common-window uncached input tokens, and -7.46% full-18h total input tokens (272.03M -> 251.74M, -6.46% turns).
      • Three Explicit Boundary Counterexamples: (1) project+read process stdout is 1.57x–1.83x selected file bytes due to navigation/diff envelope metadata; (2) first-2h early-window uncached input is 4.05x (1.00M vs 0.25M tokens) before long-horizon turnover amortizes initial full reads; (3) cross-task Historical Grid (lower is better) exhibits Official 1967 < Explore Context off 2054 < Explore Context on+Stride 2088 < Explore Context on 2242 under different runner/resource budgets (#7959, #7960), kept separate without cross-arm causal extrapolation.
  2. EdgeBench Feedback & Memory (blog/edgebench-feedback-and-memory/ & blog/zh/edgebench-feedback-and-memory/):
    • Extend the Section 1 (#research-history) timeline through late October / early November (#7311, #7759, #7800, #7959, #7960).
    • Add <h3 id="astra-late-breakthrough"> (New Foundations Consistency (Astra max, 18h): Breaking the Late Plateau After Four Arms Stall) in Section 6 (#later-common-windows) with paired bilingual SVG charts (astra-late-breakthrough-zh.svg / astra-late-breakthrough-en.svg) including a 6h–18h (91.5–92.5) zoomed inset showing how official (91.9412, 2065/2246 @ 5.72h, followed by 11 zero-gain submissions over 11.5h) and heartbeat-resume (91.9412) stall while heartbeat-explore steps through 91.99 -> 92.03 -> 92.07 -> 92.25 (9.65h–13.98h) to 92.3419 (2074/2246, +9 net repaired declarations, 0 regressions) at 17.57h.
    • Add <h3 id="ablation-followup"> in Section 8 (#new-study) linking the ANN 18h layered ablation and the newly merged cross-task five-arm causal attribution studies (#7959, #7960).
  3. TypeScript Control-Plane Refactor (blog/typescript-control-plane-refactor/ & blog/zh/typescript-control-plane-refactor/) & Blog Indexes:
    • Note #7834 (sqlite_v0 transactional todo_archive paging beyond N = 5 completed items) in Section 6 (#authority-storage) and #7510 / #7793 (v2.2.0 atomic CLI self-claim under hard_lease) in Section 9 (#semantic-repair).
    • Update the Chinese and English blog index summaries and November 3 empirical update labels.

Outcome basis: maintainer-requested bilingual blog refinement integrating verified EdgeBench and v2.2.0 evidence. Base: origin/main at 0098a0918efba50606a32f6190033abadef0fe87.

Author Declaration

  • Written by: model_agent — Google Gemini / Antigravity.
  • Specification and evidence owners: released v2.2.0 context-delivery contract, merged EdgeBench studies (#7311, #7759, #7800, #7834, #7959, #7960), and @shangzh0 archived 5-arm/6-arm trajectory ledgers.
Criterion Disposition Path / Evidence
Replace redundant context-native-path-v10 SVG with clean path tree and add empirical Context × Stride ablation chart + boundary counterexamples implemented apps/presentation/site/public/blog/{,zh/}context-delivery/index.html, context-delivery-ablation-{zh,en}.svg
Add Astra max 18h late-stage plateau breakthrough chart (with 6h–18h zoomed inset) and cross-task 5-arm trajectory-attribution summary (#7959/#7960) implemented apps/presentation/site/public/blog/{,zh/}edgebench-feedback-and-memory/index.html, astra-late-breakthrough-{zh,en}.svg
Record #7834 SQLite todo_archive transaction cutover and #7510/#7793 atomic CLI self-claim implemented apps/presentation/site/public/blog/{,zh/}typescript-control-plane-refactor/index.html
Preserve bilingual parity, existing anchor IDs, historical tables, and experimental caveats implemented node examples/blog-bilingual-index-smoke.mjs (10 paired articles), all existing IDs retained

Scope And Continuation

Complete within this static bilingual editorial slice. No runtime, control-plane, benchmark runner, or scoring behavior changes. No new abstraction or tracked binary capture pack is introduced into the PR branch.

Review Repairs

  • Kept the source discrepancy visible instead of forcing a false common denominator: the taxonomy records 86 classified replan events and 27 valid jumps, while managed-arm totals are 44 + 36 + 24 = 104 and 5 + 11 + 15 = 31; the article publishes only the arm-local ratios.
  • Recomputed the fixed-component expression as 0.4 × 0.7088 + 0.3 × 0.7251 + 0.2 × 0.7112 + 0.1 = 0.74329; 0.73903 is recorded only as the source value, and neither number is presented as an all-model upper bound.
  • Rebuilt the early-window chart on one linear 0–300% axis and recomputed the +304.7% endpoint from the axis ticks.
  • Scoped the five-arm evidence as descriptive arm-local observation rather than a matched repeated C2 comparison, without claiming a production-default change.
  • Added a bilingual source-level regression oracle for the reconciled claims, formula and SVG geometry.

Validation

  • Tested revision: 8aac5afbd5d1a543ad7f6ac989b086c0246774fe (rebased onto origin/main 0098a0918efba50606a32f6190033abadef0fe87).
  • Run state: finished. Input classes: public_fixture, archived public-safe experiment ledgers.
Check kind Result Evidence / limitation
static passed npm run build --prefix apps/presentation/site — TypeScript and Vite client/SSR build
static passed npm run smoke:seo --prefix apps/presentation/site — five prerendered routes, 28 canonical sitemap targets, assets and metadata
static passed node examples/blog-bilingual-index-smoke.mjs -> blog-bilingual-index-smoke: ok (10 paired articles)
static passed uv run --extra test loopx canary premerge --from-git-diff --git-diff-base origin/main -> status: passed, merge_gate_passed: true, self_merge_validation_passed: true, manual_holds: 0 (4/4 direct checks, 9/9 catalog canaries, and 1/1 public-boundary check passed)
static passed loopx check public-boundary scan -> 0 violations; zero remaining references to removed context-native-path-v10 SVGs; all four revised SVG files pass xmllint --noout
manual passed Rendered the bilingual pages in Chromium (ego-browser) at desktop (1440x1000) and narrow (390x844) viewports; zero page-level horizontal overflow. The static editorial changes since that capture preserve the reviewed layout.

Frontend / Visual Evidence

  • UI impact: changed (static bilingual blog articles and SVG empirical figures under apps/presentation/site/public/blog/).
  • Source data: public-safe article HTML and SVG charts derived from verified archive summaries; no private operating state or live account data.
  • Boundary: locally served static site comparing base origin/main against head 8aac5afbd5d1a543ad7f6ac989b086c0246774fe at desktop (1440×1000) and narrow (390×844) viewports. Binary PNG captures are hosted on isolated evidence commit 95fb46a88cd925565161b41fd16077107b1c54e1 outside the PR branch; the current PR head adds text-only C2 qualification after those captures.

1. Context Delivery — Native Path Section (#native-path, Desktop 1440×1000 Before vs After)

Removes the redundant context-native-path-v10 diagram and keeps the clean, readable decision-<id>/ file tree:

Context Delivery #native-path Before vs After (Desktop 1440x1000)

2. Context Delivery — Real-Model & Cross-Task Ablation Section (#ablation-evidence, Desktop 1440×1000 & Narrow 390×844 Before vs After)

Adds the ANN 18h 6-arm QPS progression + layered byte/token audit figure (context-delivery-ablation-{zh,en}.svg), 9-row empirical table, and 3 explicit boundary counterexamples:

Context Delivery #evidence Before vs After (Desktop 1440x1000)

Context Delivery — Narrow Viewport (390×844 Before vs After) & Standalone Bilingual SVGs

Context Delivery #evidence Before vs After (Narrow 390x844)

Chinese SVG (context-delivery-ablation-zh.svg):

context-delivery-ablation-zh.svg

English SVG (context-delivery-ablation-en.svg):

context-delivery-ablation-en.svg

3. EdgeBench Feedback & Memory — First Viewport (1440×1000 Before vs After)

Updates the revision header and Section 1 timeline through November 3 (#7311, #7759, #7800, #7959, #7960):

EdgeBench First Viewport Before vs After (Desktop 1440x1000)

4. EdgeBench Feedback & Memory — Astra max 18h Late-Stage Breakthrough & 6h–18h Zoomed Inset (#later-common-windows, Desktop 1440×1000 & Narrow 390×844 Before vs After)

Adds the 5-arm New Foundations Consistency (Astra max, 18h) trajectory and 6h–18h (91.5–92.5) zoomed inset showing heartbeat-explore breaking the 91.9412 4-arm plateau to 92.3419 (+9 net repaired declarations, 0 regressions):

EdgeBench #later-common-windows Before vs After (Desktop 1440x1000)

EdgeBench — Narrow Viewport (390×844 Before vs After) & Standalone Bilingual SVGs

EdgeBench #later-common-windows Before vs After (Narrow 390x844)

Chinese SVG (astra-late-breakthrough-zh.svg):

astra-late-breakthrough-zh.svg

English SVG (astra-late-breakthrough-en.svg):

astra-late-breakthrough-en.svg

Type of Change / LoopX Area

  • Documentation update (apps/presentation/site/public/blog/**)
  • Public docs or presentation surface

Boundary Checklist

  • Diff, body, and attached screenshots contain no private operating context, credentials, raw research traces, or local machine paths.
  • No benchmark run, scoring change, or duplicate research work launched.
  • Both Chinese and English editions updated in lockstep (10 paired articles smoke passed).
  • Actual Before/After desktop (1440×1000) and narrow (390×844) comparison screenshots are embedded directly above; binary PNGs remain outside the PR branch.
  • Commit includes DCO sign-off.

@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

Request changes conclusion (author-owned PR; GitHub blocks formal self-review)

精确评审 1cbc9a65b3b57d128fd4228638863bb8f5b979e8,不可变基点 feea5a6e56066eef87889d470d7991922f9f00ec。有三项 P2,均在本 PR 新增的公开结论或图中:

  1. [P2] 先核对重规划总体,再发布转换率结论。 新增“86 次重规划”的叙述,与引用的 replan_causal_breakdown.json 自相矛盾:分臂为 44 + 36 + 24 = 104,另一个 summary_metrics 的 replan_count 合计也是 104;taxonomy 的有效跃升 27 次与分臂合计 31 次也不一致。11.4%、30.6%、62.5% 分别使用 5/44、11/36、15/24,不能直接写成已核对的同一 86 事件总体,更不能仅凭这些聚合数断言替换机制导致提升。请在原证据 owner 核对总体、排除规则和分母,正文两种语言同步改成有据的描述性观察或保留未验证;回归断言总体、分组及分子守恒。位置:中文第124行,英文同段。
  2. [P2] 不要把固定观测分量的估算写成线性模型理论上限。 同段将 0.73903 作为 Graph 的理论天花板,但引用报告给出的算式 0.4×0.7088 + 0.3×0.7251 + 0.2×0.7112 + 0.1 实际为 0.74329。即便修正算术,固定一份候选的分类指标、只优化校准和延迟,也不能上界整个线性模型族可达到的分类表现;由此推断后面全部用量必然无效不成立。请复核原指标定义,将其明确限定为固定观测分量的条件估算,或删除这一上限推论;增加独立公式复算与口径断言。位置:英文第124行,中文同段。
  3. [P2] 按真实比例绘制早窗开销。 两张新增 ablation SVG 的 patch_23 将 +304.7% 画成 +72%:零点 x=514.451821,+50 刻度 x=598.143761,红柱终点 x=634.968215,换算为约72%。图上没有断轴或截断标记,红柱于是看起来比 −74.25% 的节省还短,弱化了正文要并列披露的早窗反例。请使用真实线性比例,或把不同量级拆图并明确尺度;两种语言及图内几何/标注一致性回归都要更新。位置:英文 SVG 第559–562行,中文 SVG 相同坐标。

动机 / Motivation

需要了解 LoopX 上下文机制和长程研究效果、判断是否值得使用的博客读者。

旧文把合成字节测量与真实模型结果分开;本次补入 ANN 消融、Astra 后程轨迹和跨题观察,意图让读者同时看到收益、成本与反例。

静态页面、双语导航与图像已可打开,但新增总体数字不守恒、条件上限被扩大、早窗柱形比例失真,当前还不能把这些材料作为已核验的实证更新发布。

本评审不启动新 benchmark 或模型,不要求公开原始轨迹、隐藏评测或私有日志;需要的是复用现有证据完成口径核对和公开安全投影。

改动思路 / Approach

沿现有静态博客、双语路径、共用 CSS 与 SVG 资产补充材料是合适的边界;删除重复路径图也减少了重复解释。没有理由为这次编辑新增运行时、数据账本或研究界面。

本次应交付的是可追溯且尺度真实的双语研究解释;文章消费实验 owner 的证据,不能靠图表、作者声明或构建通过自行生成 countability 或因果资格。当前最小修复是在同一编辑切片核对分母、限定推论并更正图形,不需要先完成整个研究路线或重跑付费实验。

特别检查了上下文传输字节、主会话 tokens、完整成本与原生成果的区分。新增文字保留早窗和跨题反例是有价值的,但“公开来源也这么写”不能消除来源内部矛盾。来自既有评审经验的建议在此用于检查读者实际得到什么;它不是记忆效用或实验收益的证明。

具体改动 / Changes

接受依据为 docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.md,spec_revision feea5a6e56066eef87889d470d7991922f9f00ec。5.4 Result separation:字节、模型用量与 native outcome 在表面上分层,完整费用仍未知;6.1 Native outcomes:新增 best/净修复/零丢失缺少本轮独立可核对的归档绑定,保留 unverified;6.2 Efficiency:新增总体、条件上限和图形比例存在上述不一致,not_met。C2 的重复匹配因果资格不由一批描述性轨迹、报告标题或哈希替代;本 PR 可按真实限定范围修订,不必扩展成新的实验项目。

完整差异 14 文件 +3402/-326:8 个 HTML 文件 +168/-22,4 张新 SVG +3234,删除2张旧 SVG −304。SVG 是有文章消费者的图形资产,不是运行时模块;本次也没有修改模型、评分器、权限、scheduler 或持久状态。

关键代码讲解

  • #ablation-evidence 位于两种语言的 context-delivery 页面:新增六臂 QPS、九行对照与字节/token/早窗反例。两个 SVG 由文章直接加载;当前几何错误会随正常页面路径对外呈现。
  • #astra-late-breakthrough 和 #ablation-followup 位于两种语言的 EdgeBench 页面:加入 Astra 主图与放大窗,并把另外两份公开研究的数值带进正文。当前归档账号名及缩略摘要不能单独认证完整运行的 token 覆盖或“零退化”集合关系。
  • 两篇 TypeScript 重构文章的存储、认领段落补充 #7834、#7510、#7793;两个目录页更新简介。原锚点仍在,旧路径图删除后没有残留引用。没有新增自动加载的 Agent 指令或配置开关。

对主干的风险 / Risks

最强反例正是“构建和双语 smoke 全绿,但读者得到错误科学结论”。本轮从精确头公开文件复算:总体与分组断言失败、条件算式失败、两种语言 SVG 的比例断言均失败;浏览器也实际显示了被缩短的红柱。引用数据 blob 与本头中的原文件一致,未用当前可变 main 的不同数据解释本 PR。

PASS:两版站点 client/SSR 构建;当前 SEO 及10篇双语文章 smoke;12个现存变更文件公开边界扫描、两项删除及引用检查;完整 SVG XML 与可见文字检查,无脚本、foreignObject 或外部资产引用。FAIL:四项数值一致性断言;精确范围质量记录 cqr_59534379be4f5268c842;原 standard premerge。

premerge 的4项直接检查通过,9项 catalog 中7通过、hot-path预算失败、ratchet为继承 advisory。后续相同 hot-path 单独命令在 base/head 均通过,但没有解释原失败,故不重标整门通过;ratchet两端完整输出与因果源码相同,原失败保留。这些维护状态与上述新编辑缺陷分别记录;未获取、轮询或等待 CI。

语义与 CI 对齐

这里未新增共享 vocabulary、默认激活、执行义务或更广的 actor 权限,领域词汇属于明确的研究文章。merge-base advisory 没有支持的语义候选,不代表 HTML/SVG 内所有语义已被自动核对。图片比例和推论正是需要人工及独立数字 oracle 补足的部分。没有用扫描通过消除数值缺陷,也没有把传输裁剪率写成已验证的完整费用下降。

我的整体评价 / Verdict

REQUEST_CHANGES。这是体量相称、能在现有文章内完成的编辑任务,架构归属可保留;当前可验证的信息错误阻止交付资格。long_horizon=preserved 仅指静态编辑不改变运行规则,不表示研究收益已证实;user_experience=regression,因为正常读图路径显著缩小了主要反例。

实际整页观察:默认标题与正文层次清楚,删除旧路径图后文件树更直接;英文桌面 Astra 放大窗可辨,390px 两种图成为缩略图,需要原图或旁边表格阅读详细数值。正文和表格可访问、无整页横向溢出,不能据此称统计或图形正确。视觉未验证范围:真实移动设备缩放体验、归档数据真实性与完整用量/集合差异;本轮为源码构建静态站点,没有安装或重新发布。

请修复三项已验证问题,并提供已有数据的公开安全口径/来源绑定后重验两种语言、比例与数值。ANN/Astra 完整归档和总费用仍未独立资格化,不能从本次构建、发布截图或其他已合并 PR 推断通过。

English verdict: REQUEST_CHANGES — 1cbc9a6. Reconcile the 86/104 event population, correct and qualify the Graph ceiling, and plot the 304.7% overhead at its actual scale. Builds, SEO and bilingual navigation pass; numerical consistency fails, and the original premerge failure remains recorded.

@huangruiteng
huangruiteng force-pushed the codex/blog-empirical-refinement-20261103 branch from 1cbc9a6 to 58c1f8b Compare November 3, 2026 16:22
loopx-agent and others added 3 commits November 4, 2026 00:28
…through evidence

Signed-off-by: LoopX Agent <337587101+loopx-agent@users.noreply.github.com>
Signed-off-by: huangruiteng <huangrt01@163.com>
Signed-off-by: huangruiteng <huangrt01@163.com>
@huangruiteng
huangruiteng force-pushed the codex/blog-empirical-refinement-20261103 branch from 2d8d74e to 8aac5af Compare November 3, 2026 16:30

@huangruiteng huangruiteng left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

Reviewer: model_agent · GPT-5 · OpenAI · self_reported

Approval conclusion

精确评审 8aac5afbd5d1a543ad7f6ac989b086c0246774fe,基点 0098a0918efba50606a32f6190033abadef0fe87。旧 head 的三个 P2 已在当前 head 修复,新增回归 oracle 能独立检查总体口径、公式与 SVG 几何;当前没有剩余 blocker。批准的是“公开、双语、带边界的编辑性实证说明”,不是对未重放 benchmark 归档或原生 evaluator 的再次认证。

动机 / Motivation

读者需要从 LoopX 工程博客判断 Context Delivery、Stride 与长程复盘到底改善了什么、在哪里失效,以及现有证据能支持多强的结论。旧 head 已经补入了有价值的 ANN 18h、Astra 18h 与跨任务材料,但普通页面路径仍可能把 86 个分类记录和 104 个受管臂记录混成一个总体,把 0.73903 扩写成所有线性模型的上界,并将标注为 +304.7% 的早窗反例画成约 +72%。这些问题即使构建全绿,也会让读者高估收益与因果确定性。

当前改动保留了原有文章和图表结构,但把差异明确展示出来:86/104、27/31 不再被强行对账,11.4%/30.6%/62.5% 只作为各臂自身比例;公式按实际输入复算为 0.74329,并明确只是固定分量向量的条件估算;早窗柱形使用统一的 0–300% 线性轴。五臂材料同时被限定为描述性单臂观察,而不是匹配重复的 C2 比较,也不据以改变生产默认。读者由此得到一个完整、可复算且不会自动升格因果资格的实证切片。它没有改变运行时、调度或权限,也没有重放 benchmark 归档或原生 evaluator。剩余验证缺口是重放外部研究归档与原生 evaluator;本 PR 明确把相关数字保持为来源绑定的描述性投影。

改动思路 / Approach

入口仍是现有静态站点构建到双语 HTML 与 SVG 的路径,权威输入是仓库内文章、图形资产,以及已合并研究目录中的公开 JSON/Markdown;没有新增运行时 reader、实验执行器或第二套数据源。改动复用现有 blog/context-delivery/、blog/edgebench-feedback-and-memory/、两个目录页和既有的 examples/blog-bilingual-index-smoke.mjs。删除重复的 context-native-path-v10 图是同一展示边界内的机械收敛,不改变读者任务。

正向路径是:研究 owner 发布带来源与限制的数值,文章只投影可公开解释的部分,双语 smoke 从 HTML/SVG 直接检查关键断言,站点 build/SEO 再验证公开入口。负向路径是:旧总体、错误公式、失真柱形或缺失的 C2 限定被重新引入时,source-level oracle 失败,不能把“页面能构建”误当作“结论已核验”。这比新增数据服务或重新跑付费 benchmark 更贴近当前问题;重构没有必要,纠正同一证据投影才是最小而完整的切片。

具体改动 / Changes

依据是 docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.md。在当前 head 上,5.4 Result separation 已实现:原生结果、控制观察、成本与限制分层展示;6.1 Native outcomes 保持 deferred,因为本 PR 没有重放外部归档或 evaluator;6.2 Efficiency 已实现:字节、token 与总成本没有被混成同一结论;C2 已实现为负向边界:文章明确五臂观察不是匹配重复比较,因此不声称因果资格。

完整差异为 15 个文件、+3466/-326:8 个双语文章或索引 HTML、4 个新 SVG、2 个删除的重复路径 SVG,以及现有双语 smoke 的 64 行增强。没有修改 runtime、权限、调度、评分器、persistent state 或生产默认。

关键代码讲解

  • #ablation-evidence 位于两种语言的 Context Delivery 文章,直接消费新的 ANN 18h 图与九行对照表。patch_23 的 +304.7% 柱不再是视觉压缩后的条形:源 smoke 从 0 与 300 两个刻度反算线性位置,要求终点与真实值相差不超过 0.11px。这样读者看到的早窗开销与正文披露 4.05x 反例一致。
  • #ablation-followup 位于两种语言的 EdgeBench 文章。当前文本同时展示 taxonomy 的 86/27 与受管臂小计的 104/31,拒绝制造一个不存在的公共分母;它进一步说明这组比例只是 arm-local、描述性、非 C2,不授权生产配置变化。Graph 段落把来源记录值 0.73903 与实际复算 0.74329 并列,避免把固定分量的条件估算写成模型族理论上限。
  • #astra-late-breakthrough 保留 Astra 18h 五臂轨迹和 6h–18h 放大窗,并明确这是单批次描述性证据;Schema 反例紧随其后,避免把“后程仍可突破”读成普遍保证。该段没有把基础模型/Official 的改善归功于 LoopX。
  • validateEmpiricalEvidenceClaims 是当前行为的独立 oracle:它重算固定分量公式,要求中英文文章保留 86/104、27/31、0.74329 与非 C2 限定,拒绝旧 “86 replans” 叙述,并从 SVG 刻度验证 patch_23 的几何。检查依据是数值与坐标,不是文章中出现几个漂亮关键词。
  • 删除 context-native-path-v10 两版图后,仓库搜索已无残留引用;现有四文件路径树继续解释同一机制,减少重复图和窄屏裁剪风险。

对主干的风险 / Risks

最强回归不是运行时故障,而是“页面做得更漂亮,却让错误研究结论更可信”。具体链路是:错误分母、未复算公式、被压缩的柱形或缺失的 C2 限定进入公开文章;普通 build/SEO 不会理解这些语义,读者因此高估效果。当前 head 通过源码复核和 smoke 直接阻断这条链路,blast radius 限于公开双语研究解释,回滚只需还原静态文本或 SVG 并重跑同一检查,没有数据迁移。

当前 head 的 PASS 证据包括:node examples/blog-bilingual-index-smoke.mjs(10 篇双语文章)、npm run build、npm run smoke:seo(5 个预渲染路由、28 个 sitemap 目标)、四张修订 SVG 的 xmllint --noout,以及最新 rebase 后的 loopx canary premerge(4/4 direct、9/9 catalog、1/1 public boundary,0 failure,0 manual hold)。此前 Chromium 在桌面与 390px 宽度检查过中英文页面,均无页面级横向溢出;这次新增文字没有改变布局结构。

明确未验证的范围是:没有重放链接的研究归档与原生 evaluator,没有重新执行 benchmark,没有在文字修正后重复物理移动设备测试,也没有等待或轮询 CI。PR 不改变权限、typed runtime state、scheduler 或默认 feature,故 default-off isolation 和 runtime authority 为 not_applicable。这些缺口限制 6.1 的资格,不影响本 PR 作为经过来源绑定的编辑性投影交付。

语义与 CI 对齐

这里没有新增共享 vocabulary、协议字段或执行义务;C2 是被复用的已登记研究契约,并在文章中以负向边界出现。smoke 的字符串检查只保护公开文章为这篇特定证据保留关键口径,几何和公式另由数值 oracle 验证,不冒充对 benchmark 真值或因果关系的自动认证。public-boundary scan、双语配对、build 与 premerge 各自证明公开可构建性、边界和仓库卫生;它们共同支持“当前修订可以发布”,而不支持“所有引用实验已经完成独立复验”。

我的整体评价 / Verdict

APPROVE。改动范围与问题相称:它复用现有文章、SVG 和 smoke,没有把一次证据纠错扩张成新服务或新实验;同时删除了重复图,并用 64 行针对性 oracle 把最容易再次漂移的数值、限定与几何绑定起来。long_horizon=preserved,因为运行与治理规则不变,后续研究讨论获得的却是更诚实的证据表达;user_experience=improved,因为读者现在能看到真实尺度的反例、可复算的条件估算和明确的非 C2 边界。

残余风险是外部归档没有在本 PR 内重放、原生结果仍依赖已合并研究 owner 的公开材料,以及文字修正后没有再做物理设备截图。文章已把这些结果限定在来源和观察范围内,没有把运输字节、模型 token、原生成果或完整成本拼成单一结论。若后续要升级 6.1 或提出生产默认变化,需要新的归档复验或匹配重复研究,而不是继续改这段公开说明。

English verdict: APPROVE — 8aac5af. The 86/104 and 27/31 discrepancies are explicit, arm-local ratios remain descriptive, the fixed-component formula is recomputed as 0.74329, the +304.7% bar uses a true linear scale, and the bilingual text excludes C2 and production-default claims. Bilingual smoke, site build, SEO, SVG validation, and rebased premerge pass; archive replay and native-evaluator requalification remain explicitly deferred.

@huangruiteng
huangruiteng merged commit be2a5c3 into main Nov 3, 2026
21 of 26 checks passed
@huangruiteng
huangruiteng deleted the codex/blog-empirical-refinement-20261103 branch November 3, 2026 16:36
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