getaskclaw · askclaw-vesper · Sep 18, 2026 · adaptcom · Sep 18, 2026 · adaptcom
diff --git a/assets/completion-matrix-7way.en.png b/assets/completion-matrix-7way.en.png
diff --git a/assets/completion-matrix-7way.png b/assets/completion-matrix-7way.png
diff --git a/assets/effort-curves-20260911.en.png b/assets/effort-curves-20260911.en.png
diff --git a/assets/effort-curves-20260911.png b/assets/effort-curves-20260911.png
diff --git a/assets/score-vs-tokens-2026-w37.en.png b/assets/score-vs-tokens-2026-w37.en.png
diff --git a/assets/score-vs-tokens-2026-w37.png b/assets/score-vs-tokens-2026-w37.png
diff --git a/assets/top5-2026-w38.en.png b/assets/top5-2026-w38.en.png
diff --git a/assets/top5-2026-w38.png b/assets/top5-2026-w38.png
diff --git a/assets/trust-chain.en.png b/assets/trust-chain.en.png
diff --git a/assets/trust-chain.png b/assets/trust-chain.png
diff --git a/assets/wallclock-strip-2026-w37.en.png b/assets/wallclock-strip-2026-w37.en.png
diff --git a/assets/wallclock-strip-2026-w37.png b/assets/wallclock-strip-2026-w37.png
diff --git a/assets/what-is-amber.en.png b/assets/what-is-amber.en.png
diff --git a/assets/what-is-amber.png b/assets/what-is-amber.png
diff --git a/en.html b/en.html
@@ -0,0 +1,150 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="UTF-8">
+<meta name="viewport" content="width=device-width, initial-scale=1.0">
+<title>AskClaw — AMBER: sealed-in-amber replay evaluation</title>
+<meta name="description" content="AMBER replays real, auditable historical events — frozen the moment before the answer was revealed. Private case library, public scores, independently checkable.">
+<style>
+  :root { --ink:#17242b; --sub:#5a6b73; --teal:#00a5b5; --teal-dark:#0b7c88; --bg:#fafdfe; --card:#fff; --line:#e3ecef; }
+  * { box-sizing: border-box; margin: 0; padding: 0; }
+  body { font-family: system-ui, -apple-system, "Segoe UI", Roboto, sans-serif; background: var(--bg); color: var(--ink); line-height: 1.7; }
+  a { color: var(--teal-dark); text-decoration: none; }
+  a:hover { text-decoration: underline; }
+  .wrap { max-width: 980px; margin: 0 auto; padding: 0 20px; }
+  header { border-bottom: 1px solid var(--line); padding: 18px 0; }
+  header .wrap { display: flex; justify-content: space-between; align-items: center; }
+  .logo { font-weight: 800; font-size: 20px; color: var(--ink); }
+  .logo span { color: var(--teal); }
+  nav a { margin-left: 18px; font-size: 14px; color: var(--sub); }
+  .hero { padding: 64px 0 40px; }
+  .hero h1 { font-size: 40px; line-height: 1.3; letter-spacing: .5px; }
+  .hero h1 em { font-style: normal; color: var(--teal); }
+  .hero p.lead { margin-top: 16px; font-size: 18px; color: var(--sub); max-width: 720px; }
+  .badges { display: flex; gap: 14px; margin-top: 28px; flex-wrap: wrap; }
+  .badge { background: var(--card); border: 1px solid var(--line); border-radius: 10px; padding: 12px 18px; }
+  .badge b { font-size: 22px; color: var(--teal-dark); }
+  .badge small { display: block; color: var(--sub); font-size: 12.5px; }
+  section { padding: 36px 0; }
+  h2 { font-size: 24px; margin-bottom: 6px; }
+  .hint { color: var(--sub); font-size: 14.5px; margin-bottom: 18px; }
+  figure { margin: 10px 0 6px; }
+  figure img { width: 100%; border: 1px solid var(--line); border-radius: 10px; background: #fff; }
+  figcaption { color: var(--sub); font-size: 13.5px; margin-top: 6px; }
+  .grid2 { display: grid; grid-template-columns: 1fr 1fr; gap: 18px; }
+  .repos { display: grid; grid-template-columns: repeat(auto-fill, minmax(280px, 1fr)); gap: 14px; }
+  .repo { background: var(--card); border: 1px solid var(--line); border-radius: 10px; padding: 14px 16px; display: block; color: var(--ink); }
+  .repo:hover { border-color: var(--teal); text-decoration: none; }
+  .repo b { color: var(--teal-dark); }
+  .repo .score { float: right; font-weight: 700; color: var(--ink); }
+  .repo p { font-size: 13.5px; color: var(--sub); margin-top: 4px; }
+  .rule { border-left: 3px solid var(--teal); padding: 4px 14px; margin: 10px 0; color: var(--sub); }
+  footer { border-top: 1px solid var(--line); margin-top: 40px; padding: 26px 0 40px; color: var(--sub); font-size: 13.5px; }
+  @media (max-width: 720px) { .grid2 { grid-template-columns: 1fr; } .hero h1 { font-size: 30px; } }
+</style>
+</head>
+<body>
+<header>
+  <div class="wrap">
+    <div class="logo">Ask<span>Claw</span></div>
+    <nav>
+      <a href="https://github.com/getaskclaw">GitHub</a>
+      <a href="index.html">中文</a>
+    </nav>
+  </div>
+</header>
+
+<div class="wrap">
+
+<section class="hero">
+  <h1><em>AMBER</em> — real history,<br>sealed in amber, replayed</h1>
+  <p class="lead">We take real, auditable incidents and freeze them at the moment before the answer was revealed. Models must redo the work with only what was knowable then, against a rubric frozen before any output existed. <b>The cases are never published. The scores always are.</b></p>
+  <div class="badges">
+    <div class="badge"><b>23 cases / 26 papers</b><small>private library, hash-sealed</small></div>
+    <div class="badge"><b>11 result repos</b><small>one per vendor × endpoint</small></div>
+    <div class="badge"><b>Updated weekly</b><small>same library, same band, public hash index</small></div>
+  </div>
+</section>
+
+<section>
+  <h2>The board right now</h2>
+  <p class="hint">Same band (high), same 23 cases, same public hash index. Five rank tiers; ties are labeled as ties.</p>
+  <figure>
+    <img src="assets/top5-2026-w38.en.png" alt="AMBER board: five rank tiers (2026-W38)">
+    <figcaption>2026-W38 · leader swe-2-max @ Devin at 18/23; a five-way tie at 17/23.</figcaption>
+  </figure>
+</section>
+
+<section>
+  <h2>Same score, different shape</h2>
+  <p class="hint">The nine-axis completion matrix: each cell is a face's pin-level completion. Tied totals hide completely different profiles — the matrix shows what the sum can't.</p>
+  <figure>
+    <img src="assets/completion-matrix-7way.en.png" alt="Nine-axis completion matrix">
+    <figcaption>Five lanes tied at 17/23, plus doubao 16/23 and gp27b 14/23.</figcaption>
+  </figure>
+</section>
+
+<section>
+  <h2>Why the scores can be trusted</h2>
+  <p class="hint">Every weekly issue runs like an exam: cases sealed in advance, a clean-room sitting, per-paper wire audits, anonymized publication, and a public index anyone can re-check.</p>
+  <figure>
+    <img src="assets/trust-chain.en.png" alt="How a result is produced">
+  </figure>
+</section>
+
+<section>
+  <h2>Three counterintuitive findings</h2>
+  <p class="hint">Each comes with named boundaries and counterexamples in the original writeups.</p>
+  <figure>
+    <img src="assets/effort-curves-20260911.en.png" alt="Effort curves">
+    <figcaption>"Think longer ≠ score better": most families backfire at top bands; swe-2 is monotone to the top.</figcaption>
+  </figure>
+  <div>
+    <figure>
+      <img src="assets/score-vs-tokens-2026-w37.en.png" alt="Score vs thinking budget">
+      <figcaption>Same band, same library: output-token bills span 17× while scores differ by one case.</figcaption>
+    </figure>
+    <figure>
+      <img src="assets/wallclock-strip-2026-w37.en.png" alt="Per-case wall clock">
+      <figcaption>A vendor's high TPS only holds on easy problems; hard ones slow the token stream down.</figcaption>
+    </figure>
+  </div>
+</section>
+
+<section>
+  <h2>Result repos</h2>
+  <p class="hint">One repo per lane (model × endpoint × band), one matrix per week, wire audits and hash cross-checks included.</p>
+  <div class="repos">
+    <a class="repo" href="https://github.com/getaskclaw/amber"><b>amber</b><span class="score">spec</span><p>Method spec + public hash index + cross-repo weekly board</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-devin"><b>amber-devin</b><span class="score">18/23</span><p>swe-2-max @ Devin — current leader</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-kimi"><b>amber-kimi</b><span class="score">17/23</span><p>k3 @ Kimi official coding endpoint + the K2.8 duel</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-ollama"><b>amber-ollama</b><span class="score">17/23</span><p>glm-5.3-flash / d41f @ Ollama Cloud</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-commandcode"><b>amber-commandcode</b><span class="score">17/23</span><p>d41f @ CommandCode — five-band ladder</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-workbuddy"><b>amber-workbuddy</b><span class="score">17/23</span><p>hy4-preview-f @ WorkBuddy ACP</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-gpt"><b>amber-gpt</b><span class="score">16/23</span><p>gpt-6-astra / gpt-5.6 family @ OpenAI Codex</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-crof"><b>amber-crof</b><span class="score">16/23</span><p>qwen3.8-27b @ CrofAI — the $1.15 value king</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-deepseek"><b>amber-deepseek</b><span class="score">16/23</span><p>deepseek-flash @ DeepSeek official</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-opencode"><b>amber-opencode</b><span class="score">16/23</span><p>deepseek-flash @ OpenCode Go</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-doubao"><b>amber-doubao</b><span class="score">16/23</span><p>doubao-seed-evolving @ Volcengine Ark Agent Plan</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-goldenpotato"><b>amber-goldenpotato</b><span class="score">14/23</span><p>Qwen3.8-27B @ community self-hosted (3× V100)</p></a>
+  </div>
+</section>
+
+<section>
+  <h2>The rules</h2>
+  <div class="rule">Private side, never public: the case library, oracles, transcripts, internal case IDs and variant names.</div>
+  <div class="rule">Public side, always checkable: scores, aggregates, verdicts, the hash index — anyone can verify the library didn't change and the scores add up.</div>
+  <div class="rule">Historical outcomes are evidence, not the only correct answer; runtime sealing ≠ clean training data. Both warnings are printed on every issue.</div>
+  <p class="hint" style="margin-top:14px">Read the method: <a href="https://github.com/getaskclaw/amber/blob/main/AMBER-Core-Specification.md">AMBER Core Specification</a> · <a href="https://github.com/getaskclaw/amber/blob/main/hash-index/v2026-09.md">public hash index</a> · <a href="https://github.com/getaskclaw/amber/blob/main/PLAN.md">roadmap</a></p>
+</section>
+
+</div>
+
+<footer>
+  <div class="wrap">
+    AskClaw · AMBER results published weekly · <a href="https://github.com/getaskclaw">github.com/getaskclaw</a>
+  </div>
+</footer>
+
+</body>
+</html>
diff --git a/index.html b/index.html
@@ -0,0 +1,150 @@
+<!DOCTYPE html>
+<html lang="zh-CN">
+<head>
+<meta charset="UTF-8">
+<meta name="viewport" content="width=device-width, initial-scale=1.0">
+<title>AskClaw — AMBER 琥珀式封存回放评测</title>
+<meta name="description" content="AMBER：把真实历史事件封进琥珀，让 AI 重做当时的题。私有题库、公开成绩、人人可核对。">
+<style>
+  :root { --ink:#17242b; --sub:#5a6b73; --teal:#00a5b5; --teal-dark:#0b7c88; --bg:#fafdfe; --card:#fff; --line:#e3ecef; }
+  * { box-sizing: border-box; margin: 0; padding: 0; }
+  body { font-family: "Noto Sans SC", "PingFang SC", "Microsoft YaHei", system-ui, sans-serif; background: var(--bg); color: var(--ink); line-height: 1.7; }
+  a { color: var(--teal-dark); text-decoration: none; }
+  a:hover { text-decoration: underline; }
+  .wrap { max-width: 980px; margin: 0 auto; padding: 0 20px; }
+  header { border-bottom: 1px solid var(--line); padding: 18px 0; }
+  header .wrap { display: flex; justify-content: space-between; align-items: center; }
+  .logo { font-weight: 800; font-size: 20px; color: var(--ink); }
+  .logo span { color: var(--teal); }
+  nav a { margin-left: 18px; font-size: 14px; color: var(--sub); }
+  .hero { padding: 64px 0 40px; }
+  .hero h1 { font-size: 40px; line-height: 1.3; letter-spacing: .5px; }
+  .hero h1 em { font-style: normal; color: var(--teal); }
+  .hero p.lead { margin-top: 16px; font-size: 18px; color: var(--sub); max-width: 720px; }
+  .badges { display: flex; gap: 14px; margin-top: 28px; flex-wrap: wrap; }
+  .badge { background: var(--card); border: 1px solid var(--line); border-radius: 10px; padding: 12px 18px; }
+  .badge b { font-size: 22px; color: var(--teal-dark); }
+  .badge small { display: block; color: var(--sub); font-size: 12.5px; }
+  section { padding: 36px 0; }
+  h2 { font-size: 24px; margin-bottom: 6px; }
+  .hint { color: var(--sub); font-size: 14.5px; margin-bottom: 18px; }
+  figure { margin: 10px 0 6px; }
+  figure img { width: 100%; border: 1px solid var(--line); border-radius: 10px; background: #fff; }
+  figcaption { color: var(--sub); font-size: 13.5px; margin-top: 6px; }
+  .grid2 { display: grid; grid-template-columns: 1fr 1fr; gap: 18px; }
+  .repos { display: grid; grid-template-columns: repeat(auto-fill, minmax(280px, 1fr)); gap: 14px; }
+  .repo { background: var(--card); border: 1px solid var(--line); border-radius: 10px; padding: 14px 16px; display: block; color: var(--ink); }
+  .repo:hover { border-color: var(--teal); text-decoration: none; }
+  .repo b { color: var(--teal-dark); }
+  .repo .score { float: right; font-weight: 700; color: var(--ink); }
+  .repo p { font-size: 13.5px; color: var(--sub); margin-top: 4px; }
+  .rule { border-left: 3px solid var(--teal); padding: 4px 14px; margin: 10px 0; color: var(--sub); }
+  footer { border-top: 1px solid var(--line); margin-top: 40px; padding: 26px 0 40px; color: var(--sub); font-size: 13.5px; }
+  @media (max-width: 720px) { .grid2 { grid-template-columns: 1fr; } .hero h1 { font-size: 30px; } }
+</style>
+</head>
+<body>
+<header>
+  <div class="wrap">
+    <div class="logo">Ask<span>Claw</span></div>
+    <nav>
+      <a href="https://github.com/getaskclaw">GitHub</a>
+      <a href="en.html">English</a>
+    </nav>
+  </div>
+</header>
+
+<div class="wrap">
+
+<section class="hero">
+  <h1><em>AMBER</em> — 琥珀式封存的<br>历史回放评测</h1>
+  <p class="lead">把真实、可审计的历史事件封进琥珀，精确还原到「答案揭晓之前」，让 AI 只用当时可得的信息重做一遍。评分标准在看到任何输出之前就已冻结。<b>题目永不公开，分数永远可查。</b></p>
+  <div class="badges">
+    <div class="badge"><b>23 案 / 26 卷</b><small>私有题库，哈希封存</small></div>
+    <div class="badge"><b>11 个成绩仓</b><small>按厂商 × 端点分仓</small></div>
+    <div class="badge"><b>每周更新</b><small>同库、同档、公开哈希索引</small></div>
+  </div>
+</section>
+
+<section>
+  <h2>当前总榜</h2>
+  <p class="hint">同档（high）、同 23 案、同公开哈希索引。五个名次档，并列如实标注。</p>
+  <figure>
+    <img src="assets/top5-2026-w38.png" alt="AMBER 当前总榜：五个名次档（2026-W38，全库 23 案）">
+    <figcaption>2026-W38 · 榜首 swe-2-max @ Devin 18/23；#2 并列五席 17/23。</figcaption>
+  </figure>
+</section>
+
+<section>
+  <h2>同分 ≠ 同款</h2>
+  <p class="hint">九轴完成度矩阵：每格 = 该面全部案子的钉级完成度。同分的几家，形状完全不同——总分看不见的东西，矩阵看得见。</p>
+  <figure>
+    <img src="assets/completion-matrix-7way.png" alt="九轴完成度矩阵">
+    <figcaption>五家 17/23 + doubao 16/23 + gp27b 14/23，九轴同场。</figcaption>
+  </figure>
+</section>
+
+<section>
+  <h2>分数凭什么可信</h2>
+  <p class="hint">每期成绩像一场考试：出卷封存、净室应考、逐卷验脑、脱敏发布、人人可核对。</p>
+  <figure>
+    <img src="assets/trust-chain.png" alt="一期成绩怎么炼成">
+  </figure>
+</section>
+
+<section>
+  <h2>三个反直觉发现</h2>
+  <p class="hint">都有边界，都只代表已测家族——原文里每条都标着反例。</p>
+  <figure>
+    <img src="assets/effort-curves-20260911.png" alt="effort 曲线">
+    <figcaption>「想更久 ≠ 考更好」：多数家族高档反噬，swe-2 单调到底。</figcaption>
+  </figure>
+  <div>
+    <figure>
+      <img src="assets/score-vs-tokens-2026-w37.png" alt="分数 × 思考预算">
+      <figcaption>同档同库，输出 token 账单跨 17 倍，分差不过一案。</figcaption>
+    </figure>
+    <figure>
+      <img src="assets/wallclock-strip-2026-w37.png" alt="逐题耗时">
+      <figcaption>厂商的高 TPS 只在简单题上成立；难题想得多，吐字当场变慢。</figcaption>
+    </figure>
+  </div>
+</section>
+
+<section>
+  <h2>成绩仓</h2>
+  <p class="hint">每条车道（模型 × 端点 × 档位）一个仓，每期一篇矩阵，验脑记录与哈希对照齐全。</p>
+  <div class="repos">
+    <a class="repo" href="https://github.com/getaskclaw/amber"><b>amber</b><span class="score">规范</span><p>方法规范 + 公开哈希索引 + 跨仓周榜</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-devin"><b>amber-devin</b><span class="score">18/23</span><p>swe-2-max @ Devin — 当前榜首</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-kimi"><b>amber-kimi</b><span class="score">17/23</span><p>k3 @ Kimi 官方 coding 端点 + K2.8 对决</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-ollama"><b>amber-ollama</b><span class="score">17/23</span><p>glm-5.3-flash / d41f @ Ollama Cloud</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-commandcode"><b>amber-commandcode</b><span class="score">17/23</span><p>d41f @ CommandCode — 五档天梯</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-workbuddy"><b>amber-workbuddy</b><span class="score">17/23</span><p>hy4-preview-f @ WorkBuddy ACP</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-gpt"><b>amber-gpt</b><span class="score">16/23</span><p>gpt-6-astra / gpt-5.6 系 @ OpenAI Codex</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-crof"><b>amber-crof</b><span class="score">16/23</span><p>qwen3.8-27b @ CrofAI — $1.15 性价比之王</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-deepseek"><b>amber-deepseek</b><span class="score">16/23</span><p>deepseek-flash @ DeepSeek 官方</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-opencode"><b>amber-opencode</b><span class="score">16/23</span><p>deepseek-flash @ OpenCode Go</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-doubao"><b>amber-doubao</b><span class="score">16/23</span><p>doubao-seed-evolving @ 火山方舟 Agent Plan</p></a>
+    <a class="repo" href="https://github.com/getaskclaw/amber-goldenpotato"><b>amber-goldenpotato</b><span class="score">14/23</span><p>Qwen3.8-27B @ 社区自部署（3×V100）</p></a>
+  </div>
+</section>
+
+<section>
+  <h2>规则</h2>
+  <div class="rule">私有面永不公开：题库、oracle、transcript、内部案号与变体名。</div>
+  <div class="rule">公开面永远可查：分数、聚合、裁决、哈希索引——任何人都能核对题集没换、分数可对。</div>
+  <div class="rule">历史结果是证据，不是唯一正解；运行时密封 ≠ 训练数据纯净——这两条警告印在每期成绩上。</div>
+  <p class="hint" style="margin-top:14px">方法全文：<a href="https://github.com/getaskclaw/amber/blob/main/AMBER-Core-Specification.md">AMBER Core Specification</a> · <a href="https://github.com/getaskclaw/amber/blob/main/hash-index/v2026-09.md">公开哈希索引</a> · <a href="https://github.com/getaskclaw/amber/blob/main/PLAN.md">路线图</a></p>
+</section>
+
+</div>
+
+<footer>
+  <div class="wrap">
+    AskClaw · AMBER 成绩每周更新 · <a href="https://github.com/getaskclaw">github.com/getaskclaw</a>
+  </div>
+</footer>
+
+</body>
+</html>