Repository navigation
docs(blog): add empirical Context/Stride ablation and Astra 18h breakthrough evidence - #7968
Conversation
loopx-agent
left a comment
There was a problem hiding this comment.
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:
- [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。 - [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。 - [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.
1cbc9a6 to
58c1f8b
Compare
…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>
2d8d74e to
8aac5af
Compare
huangruiteng
left a comment
There was a problem hiding this comment.
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.
Goal And Delivered Outcome
Following
v2.2.0release 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:blog/context-delivery/&blog/zh/context-delivery/):context-native-path-v10architecture diagram (which duplicated thedecision-<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).<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:gpt-6.1-sol / xhigh,18h,Recall@10 >= 0.95): 6-arm progression (Official5,753 QPS@1.05h,Single5,706 QPS,Native-Goal4,372 QPS,Explore · Context off5,766 QPS,Explore · Context on6,020 QPS,Explore · Context on + Stride7,159 QPS(+24.4%vsOfficial) viaQT_8bit_direct (HNSW-SQ8)).-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%Strideselected 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).project+readprocessstdoutis1.57x–1.83xselected file bytes due to navigation/diff envelope metadata; (2) first-2hearly-window uncached input is4.05x(1.00Mvs0.25Mtokens) before long-horizon turnover amortizes initial full reads; (3) cross-taskHistorical Grid(lower is better) exhibitsOfficial 1967 < Explore Context off 2054 < Explore Context on+Stride 2088 < Explore Context on 2242under different runner/resource budgets (#7959,#7960), kept separate without cross-arm causal extrapolation.blog/edgebench-feedback-and-memory/&blog/zh/edgebench-feedback-and-memory/):#research-history) timeline through late October / early November (#7311,#7759,#7800,#7959,#7960).<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 a6h–18h(91.5–92.5) zoomed inset showing howofficial(91.9412,2065/2246@5.72h, followed by 11 zero-gain submissions over11.5h) andheartbeat-resume(91.9412) stall whileheartbeat-exploresteps through91.99 -> 92.03 -> 92.07 -> 92.25(9.65h–13.98h) to92.3419(2074/2246,+9net repaired declarations,0regressions) at17.57h.<h3 id="ablation-followup">in Section 8 (#new-study) linking theANN 18hlayered ablation and the newly merged cross-task five-arm causal attribution studies (#7959,#7960).blog/typescript-control-plane-refactor/&blog/zh/typescript-control-plane-refactor/) & Blog Indexes:#7834(sqlite_v0transactionaltodo_archivepaging beyondN = 5completed items) in Section 6 (#authority-storage) and#7510/#7793(v2.2.0atomic CLI self-claim underhard_lease) in Section 9 (#semantic-repair).Outcome basis: maintainer-requested bilingual blog refinement integrating verified EdgeBench and
v2.2.0evidence. Base:origin/mainat0098a0918efba50606a32f6190033abadef0fe87.Author Declaration
model_agent— Google Gemini / Antigravity.v2.2.0context-delivery contract, merged EdgeBench studies (#7311,#7759,#7800,#7834,#7959,#7960), and@shangzh0archived 5-arm/6-arm trajectory ledgers.context-native-path-v10SVG with clean path tree and add empirical Context × Stride ablation chart + boundary counterexamplesapps/presentation/site/public/blog/{,zh/}context-delivery/index.html,context-delivery-ablation-{zh,en}.svgAstra max 18hlate-stage plateau breakthrough chart (with6h–18hzoomed inset) and cross-task 5-arm trajectory-attribution summary (#7959/#7960)apps/presentation/site/public/blog/{,zh/}edgebench-feedback-and-memory/index.html,astra-late-breakthrough-{zh,en}.svg#7834SQLitetodo_archivetransaction cutover and#7510/#7793atomic CLI self-claimapps/presentation/site/public/blog/{,zh/}typescript-control-plane-refactor/index.htmlnode examples/blog-bilingual-index-smoke.mjs(10 paired articles), all existing IDs retainedScope 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
86classified replan events and27valid jumps, while managed-arm totals are44 + 36 + 24 = 104and5 + 11 + 15 = 31; the article publishes only the arm-local ratios.0.4 × 0.7088 + 0.3 × 0.7251 + 0.2 × 0.7112 + 0.1 = 0.74329;0.73903is recorded only as the source value, and neither number is presented as an all-model upper bound.0–300%axis and recomputed the+304.7%endpoint from the axis ticks.Validation
8aac5afbd5d1a543ad7f6ac989b086c0246774fe(rebased ontoorigin/main0098a0918efba50606a32f6190033abadef0fe87).public_fixture, archived public-safe experiment ledgers.staticpassednpm run build --prefix apps/presentation/site— TypeScript and Vite client/SSR buildstaticpassednpm run smoke:seo --prefix apps/presentation/site— five prerendered routes, 28 canonical sitemap targets, assets and metadatastaticpassednode examples/blog-bilingual-index-smoke.mjs->blog-bilingual-index-smoke: ok (10 paired articles)staticpasseduv 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(4/4direct checks,9/9catalog canaries, and1/1public-boundary check passed)staticpassedloopx checkpublic-boundary scan ->0violations; zero remaining references to removedcontext-native-path-v10SVGs; all four revised SVG files passxmllint --nooutmanualpassedego-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
apps/presentation/site/public/blog/).origin/mainagainst head8aac5afbd5d1a543ad7f6ac989b086c0246774feat desktop (1440×1000) and narrow (390×844) viewports. Binary PNG captures are hosted on isolated evidence commit95fb46a88cd925565161b41fd16077107b1c54e1outside the PR branch; the current PR head adds text-only C2 qualification after those captures.1. Context Delivery — Native Path Section (
#native-path, Desktop1440×1000Before vs After)Removes the redundant
context-native-path-v10diagram and keeps the clean, readabledecision-<id>/file tree:2. Context Delivery — Real-Model & Cross-Task Ablation Section (
#ablation-evidence, Desktop1440×1000& Narrow390×844Before vs After)Adds the
ANN 18h6-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 — Narrow Viewport (
390×844Before vs After) & Standalone Bilingual SVGsChinese SVG (
context-delivery-ablation-zh.svg):English SVG (
context-delivery-ablation-en.svg):3. EdgeBench Feedback & Memory — First Viewport (
1440×1000Before vs After)Updates the revision header and Section 1 timeline through November 3 (
#7311,#7759,#7800,#7959,#7960):4. EdgeBench Feedback & Memory —
Astra max 18hLate-Stage Breakthrough &6h–18hZoomed Inset (#later-common-windows, Desktop1440×1000& Narrow390×844Before vs After)Adds the 5-arm
New Foundations Consistency (Astra max, 18h)trajectory and6h–18h(91.5–92.5) zoomed inset showingheartbeat-explorebreaking the91.94124-arm plateau to92.3419(+9net repaired declarations,0regressions):EdgeBench — Narrow Viewport (
390×844Before vs After) & Standalone Bilingual SVGsChinese SVG (
astra-late-breakthrough-zh.svg):English SVG (
astra-late-breakthrough-en.svg):Type of Change / LoopX Area
apps/presentation/site/public/blog/**)Boundary Checklist
10 paired articlessmoke passed).1440×1000) and narrow (390×844) comparison screenshots are embedded directly above; binary PNGs remain outside the PR branch.