feat: add paired DeepSeek routing length control

src/components/DeepSeekArtifactLab.astro

@@ -5,6 +5,11 @@ import rawAbsorb from "@/data/deepseek-v2-lite-absorb.json";
 import rawAbsorbRepro from "@/data/deepseek-v2-lite-absorb-repro.json";
 import rawCorpus from "@/data/deepseek-v2-lite-routing-corpus.json";
 import rawCorpusRepro from "@/data/deepseek-v2-lite-routing-corpus-repro.json";
+import rawMatched16 from "@/data/deepseek-v2-lite-routing-matched16.json";
+import rawMatched16Repro from "@/data/deepseek-v2-lite-routing-matched16-repro.json";
+import rawMatched24 from "@/data/deepseek-v2-lite-routing-matched24.json";
+import rawMatched24Repro from "@/data/deepseek-v2-lite-routing-matched24-repro.json";
+import rawLengthSensitivity from "@/data/deepseek-v2-lite-routing-length-sensitivity.json";
 
 const trace = rawTrace as any;
 const repro = rawRepro as any;
@@ -12,8 +17,15 @@ const absorb = rawAbsorb as any;
 const absorbRepro = rawAbsorbRepro as any;
 const corpus = rawCorpus as any;
 const corpusRepro = rawCorpusRepro as any;
+const matched16 = rawMatched16 as any;
+const matched16Repro = rawMatched16Repro as any;
+const matched24 = rawMatched24 as any;
+const matched24Repro = rawMatched24Repro as any;
+const lengthSensitivity = rawLengthSensitivity as any;
 const absorbExact = JSON.stringify(absorb) === JSON.stringify(absorbRepro);
 const corpusExact = JSON.stringify(corpus) === JSON.stringify(corpusRepro);
+const matched16Exact = JSON.stringify(matched16) === JSON.stringify(matched16Repro);
+const matched24Exact = JSON.stringify(matched24) === JSON.stringify(matched24Repro);
 const bytes = (value: number) => value >= 1024
   ? `${(value / 1024).toFixed(2)} KiB`
   : `${value.toLocaleString()} B`;
@@ -49,17 +61,25 @@ const compact = {
   },
 };
 const compactJson = JSON.stringify(compact).replaceAll("<", "\\u003c");
-const corpusCompact = {
-  domains: corpus.corpus_contract.domains,
-  labels: corpus.corpus_contract.domain_labels,
-  counts: corpus.corpus_contract.counts,
-  inference: corpus.inference_contract,
-  statistics: corpus.statistical_contract,
-  layers: corpus.layers.slice(1).map((layer: any) => ({
+const compactCorpus = (input: any) => ({
+  domains: input.corpus_contract.domains,
+  labels: input.corpus_contract.domain_labels,
+  counts: input.corpus_contract.counts,
+  inference: input.inference_contract,
+  statistics: input.statistical_contract,
+  layers: input.layers.slice(1).map((layer: any) => ({
     layer: layer.layer,
     routes: layer.routes,
     modes: layer.statistics,
   })),
+});
+const corpusCompact = {
+  cohorts: {
+    natural: compactCorpus(corpus),
+    matched16: compactCorpus(matched16),
+    matched24: compactCorpus(matched24),
+  },
+  lengthSensitivity,
 };
 const corpusCompactJson = JSON.stringify(corpusCompact).replaceAll("<", "\\u003c");
 const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoint_tensor_bytes;
@@ -73,7 +93,7 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
     </div>
     <p>
       固定官方 revision、tokenizer、模型代码和 BF16 第一分片；RTX 5090 连续执行 layer 0–6，
-      从 3,240 次 token 显微轨迹扩到 304,560 次公开语料路由，并让 layer-1 权重继续走入官方吸收式 cache。
+      从 3,240 次 token 显微轨迹扩到 488,880 次公开语料路由，并让 layer-1 权重继续走入官方吸收式 cache。
       所有结论都带证据身份与停止线。
     </p>
   </header>
@@ -390,22 +410,35 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
 
   <section class="artifact-panel" data-artifact-panel="corpus" hidden>
     <div class="panel-lead">
-      <div><span>X + S / FIXED PUBLIC CORPUS</span><h4>从四条示例，走到 128 条可重建样本与区间</h4></div>
+      <div><span>X + S / FIXED PUBLIC CORPUS</span><h4>从自然长度，再走到同 prompt 的 16 / 24-token 对照</h4></div>
       <p>
-        WikiText-2、TNEWS、HumanEval、GSM8K 各取 32 条；固定哈希选样、最多 96 tokens。
-        每个区间都重采样 prompt，而不是把同一 prompt 里的 token 假装成独立样本。
+        三个 cohort 都来自 WikiText-2、TNEWS、HumanEval、GSM8K；等长两档使用完全相同的
+        128 条源 prompt 和嵌套前缀。每个区间都重采样 prompt，不把相关 token 假装成独立样本。
       </p>
     </div>
 
     <div class="corpus-ledger">
-      <article><span>PROMPTS</span><b>{corpus.inference_contract.total_prompts}</b><p>4 domains × 32</p></article>
-      <article><span>VALID TOKENS</span><b>{corpus.inference_contract.valid_tokens.toLocaleString()}</b><p>答案未使用，代码未执行</p></article>
-      <article><span>REAL ROUTES</span><b>{(corpus.inference_contract.routes_per_moe_layer * 6).toLocaleString()}</b><p>50,760 / layer × 6 MoE layers</p></article>
+      <article><span>COHORTS</span><b>3</b><p>natural ≤96 · matched 16 / 24</p></article>
+      <article><span>PROMPT RUNS</span><b>384</b><p>每档 4 domains × 32</p></article>
+      <article><span>VALID TOKENS</span><b>13,580</b><p>答案未使用，代码未执行</p></article>
+      <article><span>REAL ROUTES</span><b>488,880</b><p>三档 × 前六个 MoE 层</p></article>
       <article><span>BOOTSTRAP</span><b>{corpus.statistical_contract.replicates.toLocaleString()}</b><p>prompt-level / domain-stratified</p></article>
-      <article class="exact"><span>INDEPENDENT RERUN</span><b>{corpusExact ? "BYTE-EXACT" : "MISMATCH"}</b><p>SHA-256 4678a1d1…a09e4</p></article>
+      <article class="exact">
+        <span>INDEPENDENT RERUN</span>
+        <b>{corpusExact && matched16Exact && matched24Exact ? "3 / 3 EXACT" : "MISMATCH"}</b>
+        <p>自然长度与两个等长 cohort</p>
+      </article>
     </div>
 
     <div class="corpus-controls">
+      <div>
+        <span>COHORT</span>
+        <div class="corpus-cohort-switch" role="group" aria-label="选择公开语料长度 cohort">
+          <button type="button" data-corpus-cohort="natural" aria-pressed="true">自然 ≤96</button>
+          <button type="button" data-corpus-cohort="matched16" aria-pressed="false">同样本 16</button>
+          <button type="button" data-corpus-cohort="matched24" aria-pressed="false">同样本 24</button>
+        </div>
+      </div>
       <div>
         <span>MOE LAYER</span>
         <div class="layer-switch corpus-layer-switch" role="group" aria-label="选择公开语料路由层">
@@ -422,7 +455,7 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
         </div>
       </div>
       <p data-corpus-mode-note>
-        先把每条 prompt 的 64 维路由分布归一，再平均；短中文标题与长代码 prompt 各有一票。
+        自然长度 cohort：每条 prompt 先归一再等权；它保留来源长度差异，适合描述实际选中样本。
       </p>
     </div>
 
@@ -456,7 +489,38 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
     <div class="corpus-findings">
       <article><span>CURRENT HIGHEST CV</span><b data-corpus-highest-cv></b><p data-corpus-highest-cv-ci></p></article>
       <article><span>CURRENT LARGEST JSD</span><b data-corpus-largest-jsd></b><p data-corpus-largest-jsd-ci></p></article>
-      <article><span>WHAT CHANGED</span><b>区间替代单点印象</b><p>看到“不同”之后，继续问 prompt 换一批时波动多大。</p></article>
+      <article><span>CURRENT COHORT</span><b data-corpus-cohort-title>自然长度 ≤96</b><p data-corpus-cohort-boundary>四域 token 总量不同；不能把差异全归因于内容。</p></article>
+    </div>
+
+    <div class="length-sensitivity">
+      <div class="length-sensitivity-head">
+        <div>
+          <span>PAIRED LENGTH SENSITIVITY / SAME SOURCE PROMPTS</span>
+          <h5>同一条 prompt：16 → 24 tokens，CV 怎样变化？</h5>
+        </div>
+        <p>
+          short / long 每次 bootstrap 使用同一组 prompt indices；下方 Δ = CV24 − CV16。
+          负值表示读入后续 8 tokens 后，64-expert 分布更平。
+        </p>
+      </div>
+      <div class="length-delta-grid" data-length-delta-grid></div>
+      <div class="length-pair-summary">
+        <article>
+          <span>中文新闻 ↔ 代码 / JSD</span>
+          <b data-length-jsd></b>
+          <p data-length-jsd-ci></p>
+        </article>
+        <article>
+          <span>LARGEST |CV Δ|</span>
+          <b data-length-largest></b>
+          <p data-length-largest-ci></p>
+        </article>
+        <article>
+          <span>READING</span>
+          <b>长度影响存在，但没有抹掉域差异</b>
+          <p>六层中文↔代码 JSD 都下降，24-token 下仍保持非零经验距离。</p>
+        </article>
+      </div>
     </div>
 
     <div class="evidence-links">
@@ -469,7 +533,8 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
     <div class="artifact-boundary">
       <b>DESCRIPTIVE, NOT SEMANTICS</b>
       <p>
-        这是固定 128 条公开 prompt 上的前六个 MoE 层，不是训练分布或线上流量。
+        这是三个固定公开 cohort 上的前六个 MoE 层，不是训练分布或线上流量；matched cohort
+        只代表各域至少 24 tokens 的子群。
         bootstrap 区间描述本探针换 prompt 的稳定性，不是零差异假设检验；没有多重比较校正，
         也不能把 E29、E48 等参数索引命名成“中文专家”或“代码专家”。
       </p>
@@ -566,7 +631,8 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
     <code>experiments/deepseek/v2_lite_trace.py</code> ·
     <code>experiments/deepseek/v2_lite_absorb_probe.py</code> ·
     <code>experiments/deepseek/v2_lite_routing_corpus.py</code> ·
-    <code>research/DEEPSEEK_ROUTING_CORPUS_AUDIT.md</code>
+    <code>experiments/deepseek/compare_routing_length_control.py</code> ·
+    <code>research/DEEPSEEK_ROUTING_LENGTH_CONTROL_AUDIT.md</code>
   </figcaption>
 
   <script is:inline type="application/json" data-dsv2-trace set:html={compactJson}></script>
@@ -779,13 +845,34 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
       code: "Python 代码",
       math: "小学数学",
     };
-    const corpusLayer = (layer: number) => corpusData.layers.find((item: any) => item.layer === layer);
+    const cohortMeta: Record<string, { title: string; boundary: string }> = {
+      natural: {
+        title: "自然长度 ≤96",
+        boundary: "四域 token 总量不同；不能把差异全归因于内容。",
+      },
+      matched16: {
+        title: "同样本 · 16 tokens",
+        boundary: "只读取至少 24-token 固定 cohort 的前 16 tokens。",
+      },
+      matched24: {
+        title: "同样本 · 24 tokens",
+        boundary: "与 16-token 档源 prompt 完全相同，只增加后续 8 tokens。",
+      },
+    };
+    let corpusCohort = "natural";
     let corpusLayerNumber = 1;
     let corpusMode = "prompt_balanced";
     const formatCi = (ci: number[], digits = 3) => `[${ci[0].toFixed(digits)}, ${ci[1].toFixed(digits)}]`;
+    const signed = (value: number, digits = 3) => `${value >= 0 ? "+" : ""}${value.toFixed(digits)}`;
+    const formatSignedCi = (ci: number[], digits = 3) =>
+      `[${signed(ci[0], digits)}, ${signed(ci[1], digits)}]`;
     const renderCorpus = () => {
-      const layer = corpusLayer(corpusLayerNumber);
+      const cohort = corpusData.cohorts[corpusCohort];
+      const layer = cohort.layers.find((item: any) => item.layer === corpusLayerNumber);
       const mode = layer.modes[corpusMode];
+      all<HTMLButtonElement>("[data-corpus-cohort]").forEach((button) => {
+        button.setAttribute("aria-pressed", String(button.dataset.corpusCohort === corpusCohort));
+      });
       all<HTMLButtonElement>("[data-corpus-layer]").forEach((button) => {
         button.classList.toggle("active", Number(button.dataset.corpusLayer) === corpusLayerNumber);
       });
@@ -794,18 +881,24 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
       });
       set(
         "[data-corpus-mode-note]",
-        corpusMode === "prompt_balanced"
-          ? "先把每条 prompt 的 64 维路由分布归一，再平均；短中文标题与长代码 prompt 各有一票。"
-          : "直接汇总全部 token 的路由计数；长 prompt 权重更高，适合回答“本批 token 实际流向哪里”。",
+        corpusCohort === "natural"
+          ? corpusMode === "prompt_balanced"
+            ? "自然长度 cohort：每条 prompt 先归一再等权；它保留来源长度差异，适合描述实际选中样本。"
+            : "自然长度 cohort：直接汇总 token 路由；长 prompt 权重更高，适合回答本批 token 实际流向哪里。"
+          : corpusMode === "prompt_balanced"
+            ? "等长 cohort：每条 prompt 等权；16 / 24 两档使用同一批源样本，可以做 paired delta。"
+            : "等长 cohort 中每条 prompt 的 token 数相同，因此 token 加权与 prompt 等权理论上重合。",
       );
       set(
         "[data-corpus-heat-title]",
-        `layer ${corpusLayerNumber} · ${corpusMode === "prompt_balanced" ? "每条 prompt 等权" : "按 token 加权"}`,
+        `${cohortMeta[corpusCohort].title} · layer ${corpusLayerNumber} · ${corpusMode === "prompt_balanced" ? "prompt 等权" : "token 加权"}`,
       );
+      set("[data-corpus-cohort-title]", cohortMeta[corpusCohort].title);
+      set("[data-corpus-cohort-boundary]", cohortMeta[corpusCohort].boundary);
 
       const rows = one<HTMLElement>("[data-corpus-domain-rows]");
       if (rows) {
-        rows.replaceChildren(...corpusData.domains.map((domain: string) => {
+        rows.replaceChildren(...cohort.domains.map((domain: string) => {
           const result = mode.domains[domain];
           const metrics = result.metrics;
           const top = result.top_experts[0];
@@ -816,7 +909,7 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
           const identityLabel = document.createElement("b");
           const identityMeta = document.createElement("small");
           identityLabel.textContent = corpusLabels[domain];
-          identityMeta.textContent = `32 prompts · ${corpusData.counts[domain].valid_tokens.toLocaleString()} tokens`;
+          identityMeta.textContent = `32 prompts · ${cohort.counts[domain].valid_tokens.toLocaleString()} tokens`;
           identity.append(identityLabel, identityMeta);
 
           const cv = document.createElement("span");
@@ -846,7 +939,7 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
 
       const heatmap = one<HTMLElement>("[data-corpus-heatmap]");
       if (heatmap) {
-        heatmap.replaceChildren(...corpusData.domains.map((domain: string) => {
+        heatmap.replaceChildren(...cohort.domains.map((domain: string) => {
           const result = mode.domains[domain];
           const max = Math.max(...result.distribution);
           const row = document.createElement("div");
@@ -874,16 +967,16 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
         const corner = document.createElement("b");
         corner.textContent = "DOMAIN";
         cells.push(corner);
-        corpusData.domains.forEach((domain: string) => {
+        cohort.domains.forEach((domain: string) => {
           const header = document.createElement("b");
           header.textContent = corpusLabels[domain];
           cells.push(header);
         });
-        corpusData.domains.forEach((left: string) => {
+        cohort.domains.forEach((left: string) => {
           const header = document.createElement("b");
           header.textContent = corpusLabels[left];
           cells.push(header);
-          corpusData.domains.forEach((right: string) => {
+          cohort.domains.forEach((right: string) => {
             const cell = document.createElement("span");
             if (left === right) {
               cell.className = "diagonal";
@@ -902,7 +995,7 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
         jsd.replaceChildren(...cells);
       }
 
-      const highest = corpusData.domains
+      const highest = cohort.domains
         .map((domain: string) => ({ domain, value: mode.domains[domain].metrics.cv }))
         .sort((left: any, right: any) => right.value.point - left.value.point)[0];
       const largest = [...mode.pairs]
@@ -925,7 +1018,63 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
         "[data-corpus-largest-jsd-ci]",
         `95% CI ${formatCi(largest.js_divergence.ci95)} · nats`,
       );
+
+      const comparisonLayer = corpusData.lengthSensitivity.layers
+        .find((item: any) => item.layer === corpusLayerNumber);
+      const comparison = comparisonLayer.modes[corpusMode];
+      const deltaGrid = one<HTMLElement>("[data-length-delta-grid]");
+      if (deltaGrid) {
+        deltaGrid.replaceChildren(...cohort.domains.map((domain: string) => {
+          const result = comparison.domains[domain];
+          const cv = result.metrics.cv;
+          const card = document.createElement("article");
+          const label = document.createElement("span");
+          const values = document.createElement("b");
+          const delta = document.createElement("strong");
+          const detail = document.createElement("p");
+          label.textContent = corpusLabels[domain];
+          values.textContent = `${cv.short.toFixed(3)} → ${cv.long.toFixed(3)}`;
+          delta.textContent = `Δ ${signed(cv.delta_long_minus_short)}`;
+          delta.className = cv.delta_long_minus_short <= 0 ? "down" : "up";
+          detail.textContent = `paired 95% ${formatSignedCi(cv.delta_ci95)} · TV ${result.total_variation.point.toFixed(3)}`;
+          card.append(label, values, delta, detail);
+          return card;
+        }));
+      }
+      const chineseCode = comparison.pairs.find((pair: any) =>
+        pair.left === "chinese" && pair.right === "code"
+      ).js_divergence;
+      set(
+        "[data-length-jsd]",
+        `${chineseCode.short.toFixed(3)} → ${chineseCode.long.toFixed(3)} · Δ ${signed(chineseCode.delta_long_minus_short)}`,
+      );
+      set(
+        "[data-length-jsd-ci]",
+        `paired 95% ${formatSignedCi(chineseCode.delta_ci95)} · L${corpusLayerNumber}`,
+      );
+      const largestDelta = cohort.domains
+        .map((domain: string) => ({
+          domain,
+          value: comparison.domains[domain].metrics.cv,
+        }))
+        .sort((left: any, right: any) =>
+          Math.abs(right.value.delta_long_minus_short) - Math.abs(left.value.delta_long_minus_short)
+        )[0];
+      set(
+        "[data-length-largest]",
+        `${corpusLabels[largestDelta.domain]} · Δ ${signed(largestDelta.value.delta_long_minus_short)}`,
+      );
+      set(
+        "[data-length-largest-ci]",
+        `paired 95% ${formatSignedCi(largestDelta.value.delta_ci95)} · CV24 − CV16`,
+      );
     };
+    all<HTMLButtonElement>("[data-corpus-cohort]").forEach((button) => {
+      button.addEventListener("click", () => {
+        corpusCohort = button.dataset.corpusCohort ?? "natural";
+        renderCorpus();
+      });
+    });
     all<HTMLButtonElement>("[data-corpus-layer]").forEach((button) => {
       button.addEventListener("click", () => {
         corpusLayerNumber = Number(button.dataset.corpusLayer);
@@ -987,7 +1136,8 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
   .corpus-controls > div > span,
   .corpus-heat-head span,
   .corpus-comparison span,
-  .corpus-findings span {
+  .corpus-findings span,
+  .length-sensitivity span {
     margin: 0;
     color: var(--blue);
     font: 700 .69rem/1.3 var(--font-mono);
@@ -1339,7 +1489,7 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
   .load-lessons b { display: block; margin-top: .35rem; font-size: .77rem; }
   .corpus-ledger {
     display: grid;
-    grid-template-columns: repeat(5, 1fr);
+    grid-template-columns: repeat(6, 1fr);
     border: 1px solid rgba(32,32,39,.14);
   }
   .corpus-ledger article {
@@ -1361,7 +1511,7 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
   }
   .corpus-controls {
     display: grid;
-    grid-template-columns: auto 1.1fr minmax(15rem, 1.3fr);
+    grid-template-columns: 1.2fr auto 1fr;
     gap: 1rem;
     align-items: end;
     margin-top: .8rem;
@@ -1371,8 +1521,10 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
   }
   .corpus-controls > div { display: grid; gap: .45rem; }
   .corpus-controls .layer-switch { margin: 0; }
-  .corpus-mode-switch { display: flex; }
-  .corpus-mode-switch button {
+  .corpus-mode-switch,
+  .corpus-cohort-switch { display: flex; }
+  .corpus-mode-switch button,
+  .corpus-cohort-switch button {
     padding: .58rem .75rem;
     border: 1px solid rgba(32,32,39,.22);
     background: #fffdf8;
@@ -1380,14 +1532,19 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
     font: 650 .67rem/1 var(--font-mono);
     cursor: pointer;
   }
-  .corpus-mode-switch button + button { border-left: 0; }
-  .corpus-mode-switch button[aria-pressed="true"] {
+  .corpus-mode-switch button + button,
+  .corpus-cohort-switch button + button { border-left: 0; }
+  .corpus-mode-switch button[aria-pressed="true"],
+  .corpus-cohort-switch button[aria-pressed="true"] {
     border-color: var(--blue);
     background: var(--blue);
     color: white;
   }
   .corpus-controls > p {
+    grid-column: 1 / -1;
     margin: 0;
+    padding-top: .75rem;
+    border-top: 1px solid rgba(32,32,39,.12);
     color: rgba(32,32,39,.62);
     font-size: .69rem;
     line-height: 1.5;
@@ -1543,6 +1700,93 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
     font-size: .65rem;
     line-height: 1.45;
   }
+  .length-sensitivity {
+    margin-top: .8rem;
+    padding: 1rem;
+    border: 1px solid rgba(32,32,39,.15);
+    background:
+      linear-gradient(120deg, rgba(57,120,110,.11), transparent 42%),
+      #e8e2d7;
+  }
+  .length-sensitivity-head {
+    display: grid;
+    grid-template-columns: 1fr 1.25fr;
+    gap: 1.2rem;
+    align-items: end;
+  }
+  .length-sensitivity h5 {
+    margin: .4rem 0 0;
+    font: 720 1.08rem/1.15 var(--font-display);
+  }
+  .length-sensitivity-head > p {
+    margin: 0;
+    color: rgba(32,32,39,.62);
+    font-size: .69rem;
+    line-height: 1.55;
+  }
+  .length-delta-grid {
+    display: grid;
+    grid-template-columns: repeat(4, 1fr);
+    margin-top: .8rem;
+    border: 1px solid rgba(32,32,39,.13);
+    background: #fffdf8;
+  }
+  .length-delta-grid > :global(article) {
+    padding: .8rem;
+    border-right: 1px solid rgba(32,32,39,.11);
+  }
+  .length-delta-grid > :global(article:last-child) { border-right: 0; }
+  .length-delta-grid > :global(article > span) {
+    display: block;
+    color: var(--blue);
+    font: 700 .62rem/1.2 var(--font-mono);
+  }
+  .length-delta-grid > :global(article > b) {
+    display: block;
+    margin-top: .45rem;
+    font: 730 .88rem/1.1 var(--font-mono);
+  }
+  .length-delta-grid > :global(article > strong) {
+    display: inline-block;
+    margin-top: .35rem;
+    padding: .22rem .35rem;
+    font: 750 .7rem/1 var(--font-mono);
+  }
+  .length-delta-grid > :global(article > strong.down) {
+    background: rgba(57,120,110,.12);
+    color: var(--teal);
+  }
+  .length-delta-grid > :global(article > strong.up) {
+    background: rgba(186,118,44,.12);
+    color: var(--amber);
+  }
+  .length-delta-grid > :global(article > p) {
+    margin: .4rem 0 0;
+    color: rgba(32,32,39,.54);
+    font: .59rem/1.4 var(--font-mono);
+  }
+  .length-pair-summary {
+    display: grid;
+    grid-template-columns: repeat(3, 1fr);
+    margin-top: .65rem;
+    border: 1px solid rgba(32,32,39,.13);
+  }
+  .length-pair-summary article {
+    padding: .8rem;
+    border-right: 1px solid rgba(32,32,39,.11);
+  }
+  .length-pair-summary article:last-child { border-right: 0; }
+  .length-pair-summary b {
+    display: block;
+    margin-top: .4rem;
+    font-size: .77rem;
+  }
+  .length-pair-summary p {
+    margin: .3rem 0 0;
+    color: rgba(32,32,39,.55);
+    font-size: .63rem;
+    line-height: 1.45;
+  }
   .observed-cache {
     display: grid;
     grid-template-columns: 1fr auto 1.25fr;
@@ -1777,7 +2021,8 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
     .kernel-contract,
     .corpus-controls,
     .corpus-heat-head,
-    .corpus-comparison { grid-template-columns: 1fr; }
+    .corpus-comparison,
+    .length-sensitivity-head { grid-template-columns: 1fr; }
     .artifact-status { grid-template-columns: 1fr 1fr; }
     .artifact-tabs { grid-template-columns: 1fr 1fr; }
     .route-controls { grid-template-columns: 1fr 1fr; }
@@ -1789,6 +2034,7 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
     .checksum-grid,
     .absorb-metrics { grid-template-columns: 1fr 1fr; }
     .corpus-ledger { grid-template-columns: repeat(3, 1fr); }
+    .length-delta-grid { grid-template-columns: 1fr 1fr; }
     .corpus-heat-head p { text-align: left; }
     .layer-evidence { grid-template-columns: repeat(9, 1fr); }
     .repro-gate { grid-template-columns: 1fr 1fr; }
@@ -1817,7 +2063,9 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
     .checksum-grid,
     .cache-ledger,
     .corpus-ledger,
-    .corpus-findings { grid-template-columns: 1fr; }
+    .corpus-findings,
+    .length-delta-grid,
+    .length-pair-summary { grid-template-columns: 1fr; }
     .route-metrics article,
     .cache-ratio article,
     .load-lessons article,
@@ -1826,8 +2074,12 @@ const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoi
     .checksum-grid article { border-right: 0; border-bottom: 1px solid rgba(32,32,39,.12); }
     .corpus-ledger article,
     .corpus-findings article { border-right: 0; border-bottom: 1px solid rgba(32,32,39,.12); }
-    .corpus-mode-switch { display: grid; grid-template-columns: 1fr; }
-    .corpus-mode-switch button + button { border-left: 1px solid rgba(32,32,39,.22); border-top: 0; }
+    .corpus-mode-switch,
+    .corpus-cohort-switch { display: grid; grid-template-columns: 1fr; }
+    .corpus-mode-switch button + button,
+    .corpus-cohort-switch button + button { border-left: 1px solid rgba(32,32,39,.22); border-top: 0; }
+    .length-delta-grid > :global(article),
+    .length-pair-summary article { border-right: 0; border-bottom: 1px solid rgba(32,32,39,.11); }
     .artifact-boundary { grid-template-columns: 1fr; }
     .load-dials { grid-template-columns: 1fr; }
     .observed-cache,

src/pages/deepseek/index.astro

@@ -770,13 +770,13 @@ const toc = [
         <p class="lede">
           前面的四联实验负责建立公式与角色合同；下面的六联工件实验固定官方 revision、tokenizer、
           模型代码和 checkpoint 第一分片，在 RTX 5090 上连续执行 layer 0–6。它把真实观测、shape 推导、
-          吸收式 latent cache、128 条公开语料的路由区间、实现差距和未覆盖范围放在同一张证据图里。
+          吸收式 latent cache、自然长度与同 prompt 等长对照、实现差距和未覆盖范围放在同一张证据图里。
         </p>
         <div class="artifact-callout">
           <article><span>X / FORWARD</span><b>7 / 27 layers</b><p>1 个 dense 层 + 6 个 MoE 层；layer 7 因跨分片停止。</p></article>
-          <article><span>X / ROUTES</span><b>304,560</b><p>128 条公开 prompt、8,460 token、6 个 MoE 层的真实 top-6 选择。</p></article>
+          <article><span>X / ROUTES</span><b>488,880</b><p>自然 ≤96 与同样本 16 / 24-token 三档、六个 MoE 层的真实 top-6 选择。</p></article>
           <article><span>X / ABSORB CACHE</span><b>266,240 → 29,952 B</b><p>同一真实 layer-1 权重的 naive / absorb active buffers。</p></article>
-          <article><span>X / RERUN</span><b>BYTE-EXACT</b><p>固定选样、逐 prompt loads 与 2,000 次 bootstrap 摘要完整复跑。</p></article>
+          <article><span>X / RERUN</span><b>3 / 3 EXACT</b><p>三档 trace byte-exact；16→24 delta 使用同 prompt paired bootstrap。</p></article>
         </div>
         <DeepSeekArtifactLab />
       </section>

src/pages/progress/index.astro

@@ -15,7 +15,7 @@ const workstreams = [
   { label: "表示、位置与残差高速公路", value: 81, next: "加入真实 hidden-state / norm traces、长上下文位置外推复现与更多深层稳定性消融" },
   { label: "Scaling Laws", value: 74, next: "加入真实拟合复现、置信区间与更多模型族对照" },
   { label: "数据工程与预训练配方", value: 73, next: "逐图精读 FineWeb / DCLM，加入真实去重与 mixture traces" },
-  { label: "DeepSeek 专题", value: 91, next: "SM90 FlashMLA kernel、完整 27 层、长度匹配对照、FP8/pipeline 与 R1-like RL 复现" },
+  { label: "DeepSeek 专题", value: 92, next: "SM90 FlashMLA kernel、完整 27 层、tokenization 扰动、FP8/pipeline 与 R1-like RL 复现" },
   { label: "指令微调与人类偏好", value: 75, next: "加入真实偏好分歧样本、RM 长度偏置与 PPO/DPO 小模型复现" },
   { label: "推理与测试时扩展", value: 76, next: "真实模型采样曲线、PRM 案例与逐篇图表精读" },
   { label: "工具使用与长程 Agent", value: 74, next: "补真实环境 traces、cross-harness 对照、Agent RL 训练曲线与安全案例" },
@@ -50,7 +50,7 @@ const workstreams = [
         <div><dt>OVERALL</dt><dd>专题平均 {average}%</dd></div>
         <div><dt>READABLE</dt><dd>{published} 个首版可读专题</dd></div>
         <div><dt>ACTIVE</dt><dd>{researching} 个研究/写作中</dd></div>
-        <div><dt>UPDATED</dt><dd>2026-07-29 15:25 CST</dd></div>
+        <div><dt>UPDATED</dt><dd>2026-07-29 16:03 CST</dd></div>
         <div><dt>MODE</dt><dd>持续迭代，不锁死版本</dd></div>
       </dl>
     </div>
@@ -102,7 +102,7 @@ const workstreams = [
       <article><span>✓</span><h3>语言模型前史深度专题</h3><p>八张独立问题账、33 个正式节点、20 段长文与概率—向量—记忆—对齐四联实验。</p></article>
       <article><span>✓</span><h3>Transformer 深度专题</h3><p>十张独立问题账、40 个正式节点、21 段正文与 QKV—Mask—多头位置—Block 成本四联实验。</p></article>
       <article><span>✓</span><h3>表示、位置与残差高速公路深度专题</h3><p>二十张问题账、66 个一手节点、DeepSeek/Kimi 双谱系，以及 Token—位置—Norm—Residual/FFN 四联实验。</p></article>
-      <article><span>✓</span><h3>DeepSeek 三轮真实权重里程碑</h3><p>在二十四张问题账、十次转向与四联公式实验上，新增 V2-Lite 7/27 层连续 forward、官方 V3 absorb 的 576 元素真实缓存，以及四域 128 条固定公开 prompt、304,560 次真实路由、2,000 次 prompt bootstrap 与 byte-exact 独立重跑的六联实验。</p></article>
+      <article><span>✓</span><h3>DeepSeek 三轮真实权重里程碑</h3><p>在二十四张问题账、十次转向与四联公式实验上，新增 V2-Lite 7/27 层连续 forward、官方 V3 absorb 的 576 元素真实缓存，以及自然长度 / 同源 16 / 同源 24 三个 cohort、488,880 次真实路由、2,000 次成对 prompt bootstrap 与 3/3 byte-exact 独立重跑的六联实验。</p></article>
       <article><span>✓</span><h3>Kimi K3 技术报告二轮深读</h3><p>三十二张问题账、Figure 1–16 / Table 1–5 审计、100 节点阅读链，以及 Delta—Decay—AttnRes—LatentMoE—SiTU—QB—MOPD—Cache 八联实验。</p></article>
       <article><span>✓</span><h3>Kimi K3 三轮开放工件里程碑</h3><p>固定官方 revisions，审计 96 个 shards、497,220 个 tensor entries 与真实 KDA / MLA / MoE / MoonViT shapes；四联实验分开显示层型、tensor anatomy、参数范围和复现边界。</p></article>
       <article><span>✓</span><h3>FlashKDA RTX 5090 执行闸门</h3><p>隔离 CUDA 13.0 / glibc 2.39 编译 sm_120a wheel；6/6 官方参考逐元素相等，并完成 fixed / varlen、三种 state mode 的 1,800 个 CUDA Event samples。</p></article>
@@ -134,7 +134,7 @@ const workstreams = [
     <div class="queue-table">
       <div class="head"><b>优先级</b><b>专题</b><b>本轮交付</b><b>完成闸门</b></div>
       <div><span>P0</span><strong>K3 三轮</strong><p>开放权重 traces → FlashKDA / AttnRes / MoE 真实行为 → Figure 1–16 数值重绘与独立复现</p><em>运行证据 + 逐图复现</em></div>
-      <div><span>P0</span><strong>DeepSeek 三轮</strong><p>SM90 FlashMLA kernel / 完整 27 层 / 长度匹配对照 → FP8 / pipeline traces → R1-like RL 小模型复现</p><em>运行证据 + 独立复现</em></div>
+      <div><span>P0</span><strong>DeepSeek 三轮</strong><p>SM90 FlashMLA kernel / 完整 27 层 / tokenizer 与 prompt-template 扰动 → FP8 / pipeline traces → R1-like RL 小模型复现</p><em>运行证据 + 独立复现</em></div>
       <div><span>P0</span><strong>Transformer 二轮</strong><p>多头电路逐图 → Pre/Post-LN 真实 traces → Flash/KV 配置与 kernel 对照</p><em>逐图笔记 + 实测边界</em></div>
       <div><span>P0</span><strong>表示、位置与残差二轮</strong><p>真实 hidden-state / norm traces → 长上下文位置外推 → mHC / AttnRes 深层稳定性消融</p><em>可复现实验 + 逐图笔记</em></div>
       <div><span>P0</span><strong>语言模型前史二轮</strong><p>Kneser–Ney / LSTM / Bahdanau 逐图 → 真实小语料复现 → tokenizer 公平性</p><em>可复现实验 + 逐图笔记</em></div>
@@ -215,6 +215,9 @@ const workstreams = [
       <div><time>2026-07-29</time><b>A_log 形状冲突保持未决</b><p>checkpoint 的 [128] 与 config / remote code / FlashKDA API 期待的 [96] 并列展示；不宣布权重损坏，也不把 channel-wise 假设写成真实 forward。</p></div>
       <div><time>2026-07-29</time><b>FlashKDA 编译与执行永久分两道闸门</b><p>容器产出 sm_120a wheel 只证明可编译；RTX 5090 的 6/6 official-reference exact suite 通过后，才把证据升级为本机执行 X。</p></div>
       <div><time>2026-07-29</time><b>作者表与 RTX 5090 表永久分账</b><p>H20 / GB200 保持 O；本站只报告独立环境、协议、300 samples/mode 和延迟分布，未跑本机 FLA 就不写本机 speedup。</p></div>
+      <div><time>2026-07-29</time><b>32-token 对照改为同源 16→24</b><p>TNEWS 只有 105/10,000 条达到 32 tokens，强行统一会落入约 1% 极端长尾；24-token eligibility 仍保留 1,609 条中文候选。</p></div>
+      <div><time>2026-07-29</time><b>长度敏感性必须成对重采样</b><p>16-token 输入严格是 24-token 输入前缀，2,000 次 bootstrap 共用 prompt indices；结果只描述固定 cohort 的长度敏感性。</p></div>
+      <div><time>2026-07-29</time><b>三类 cohort 永久分身份</b><p>自然长度回答本批样本如何路由；matched-16 / 24 回答同一 prompt 多看 8 tokens 后如何变化，不把二者混成内容因果。</p></div>
     </div>
   </section>