diff --git a/AGENTS.md b/AGENTS.md
index 082ae40..acc4025 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -149,7 +149,7 @@ cloneRepo → discoverFiles → countCodeStats → tryBuild → tryBrowse/trySta
 
 - **排名用多次聚合**：`aggregateScores`/`aggregateEntryScores`（standard-utils.ts）——最近 N 次（默认3）`review_snapshots.score` 中位数；N=2 平均、N=1 单次。聚合前提=各快照 `standard_snapshot` 一致。`<3 次标"初评（未达聚合样本）"`，排名区分正式/初评
 - **快照分数列**：`review_snapshots.score` 存含迟交扣分的 final_score（写快照时一并存，不解析 ai_report——它只有 totalScore）。历史快照 backfill 从 totalScore best-effort
-- **可验证能力三档**（`classifyVerifiability`，效果/提效类维度，校准之前判档）：A=基准证据（`entries.benchmark_json` status=done）；B=测试通过或覆盖率非 null；C=数据缺位→封顶 maxScore*0.3 + note"数据缺位（未证明），非无效"。**构建成功≠效果可验证**。C 档 note 渲染到前端维度表 + PDF
+- **可验证能力三档**（`classifyVerifiability`，效果/提效类维度，校准之前判档）：A=基准证据（`entries.benchmark_json` status=done）；B=测试通过或覆盖率非 null；C=数据缺位→封顶 maxScore*0.3 + note"数据缺位（未证明），非无效"。**构建成功≠效果可验证**。C 档 note 渲染到前端维度表 + PDF。**纯增益维度严格制（2026-08-26）**：`提效幅度/效果对比/效率提升` 类维度（isPureGainDim）测试通过**不构成**提效证据——无基线对比数据一律 C 档封顶（实测案例：六边形战队提效幅度无量化数据 raw=6，修复后同类场景 C≤3）
 - **基准证据按 entry 落库**（`entries.benchmark_json`，非 env 变量——MAX_CONCURRENT=3 并发会串数据）
 - **确定性 L1**（`detectStructuralContradictions`）：仅证据性矛盾触发（有测试/基准证据但效果≈0 → under；效果高分+实现全低 → over）。**效果维度 under 一律丢弃**（computeCalibration 内过滤）——效果维度只降不升，诚实由三档封顶负责。禁止"实现高分+无数据"当 under
 - **overall 中性边界**（`neutralizeTestEvidence`）：测试 summary 含"中性"→ 标 `[中性证据]`；不含（"未检测到测试框架配置"=真缺测试）→ 真实弱点。synthesizeOverall prompt 禁止把 `[中性证据]` 列为不足
diff --git a/server/src/__tests__/standard-utils.test.ts b/server/src/__tests__/standard-utils.test.ts
index 7a3de2a..a122d0a 100644
--- a/server/src/__tests__/standard-utils.test.ts
+++ b/server/src/__tests__/standard-utils.test.ts
@@ -330,6 +330,28 @@ describe('TC-VERIFY · 可验证能力三档（2026-08-19）', () => {
     expect(isEffectDim('效果与数据')).toBe(true);
     expect(isEffectDim('开发范式')).toBe(false);
   });
+
+  it('纯增益维度（提效幅度）：测试通过不构成提效证据 → C 档封顶（2026-08-26）', () => {
+    const d = { name: '提效幅度', score: 6, maxScore: 10 };
+    const r = classifyVerifiability(d, { hasEffectEvidence: true });
+    expect(r.tier).toBe('C');
+    expect(r.capped).toBe(true);
+    expect(r.effectiveScore).toBeLessThanOrEqual(3);
+    expect(r.note).toContain('基线对比');
+  });
+
+  it('纯增益维度：有基准对比证据 → A 档不封顶', () => {
+    const d = { name: '提效幅度', score: 9, maxScore: 10 };
+    const r = classifyVerifiability(d, { hasBenchmarkEvidence: true, hasEffectEvidence: true });
+    expect(r.tier).toBe('A');
+    expect(r.capped).toBe(false);
+  });
+
+  it('非纯增益的效果维度仍可用测试证据豁免（B 档）——赛道一行为不变', () => {
+    const r = classifyVerifiability(dim, { hasEffectEvidence: true });
+    expect(r.tier).toBe('B');
+    expect(r.capped).toBe(false);
+  });
 });
 
 describe('TC-CONTRAD · 确定性 L1 证据性矛盾（2026-08-19）', () => {
diff --git a/server/src/services/review.service.ts b/server/src/services/review.service.ts
index 162172f..887f9c1 100644
--- a/server/src/services/review.service.ts
+++ b/server/src/services/review.service.ts
@@ -1273,6 +1273,7 @@ ${overviewFileBlock}
       };
       for (const d of dimensions) {
         const v = classifyVerifiability(d, verifEvidence);
+        pipeLog(entryId, '  VERIF', `${d.name} raw=${d.score} tier=${v.tier} capped=${v.capped} eff=${v.effectiveScore}`);
         if (v.capped && d.score > v.effectiveScore) {
           d.score = v.effectiveScore;
           (d as any).verifiability = v;
diff --git a/server/src/services/standard-utils.ts b/server/src/services/standard-utils.ts
index 9ac7a8f..2aff576 100644
--- a/server/src/services/standard-utils.ts
+++ b/server/src/services/standard-utils.ts
@@ -268,14 +268,27 @@ export function isEffectDim(name: string): boolean {
   return EFFECT_EVIDENCE_KEYS.some(k => name.includes(k));
 }
 
+// 纯增益对比类维度（2026-08-26）：此类维度考察的是"基线 vs 改造后"的量化对比，
+// 工具自身的单元测试通过不构成提效证据——只有基准/对比数据（A档）才能豁免 C 档封顶。
+const PURE_GAIN_KEYS = ['提效幅度', '效果对比', '效率提升'];
+
+export function isPureGainDim(name: string): boolean {
+  return PURE_GAIN_KEYS.some(k => (name || '').includes(k));
+}
+
 export function classifyVerifiability(
   dim: { name: string; score: number; maxScore: number },
   evidence: { hasBenchmarkEvidence?: boolean; hasEffectEvidence?: boolean; hasBuildEvidence?: boolean } = {}
 ) {
   if (evidence.hasBenchmarkEvidence) return { tier: 'A' as const, capped: false, effectiveScore: dim.score, note: '有确定性基准证据（seed-defect benchmark）' };
-  if (!isEffectDim(dim.name) || evidence.hasEffectEvidence) return { tier: 'B' as const, capped: false, effectiveScore: dim.score, note: '' };
+  const strict = isPureGainDim(dim.name);
+  // 非效果维度直接 B；纯增益维度即使有测试通过证据也不豁免（严格基准制）
+  if (!isEffectDim(dim.name) || (evidence.hasEffectEvidence && !strict)) return { tier: 'B' as const, capped: false, effectiveScore: dim.score, note: '' };
   const cap = Math.floor(dim.maxScore * 0.3);
-  return { tier: 'C' as const, capped: true, effectiveScore: Math.min(dim.score, cap), note: `数据缺位（未证明），非无效；C档封顶 ${cap}/${dim.maxScore}` };
+  const note = strict
+    ? `提效类维度缺基线对比数据（未证明），非无效；测试通过不构成提效证据；C档封顶 ${cap}/${dim.maxScore}`
+    : `数据缺位（未证明），非无效；C档封顶 ${cap}/${dim.maxScore}`;
+  return { tier: 'C' as const, capped: true, effectiveScore: Math.min(dim.score, cap), note };
 }
 
 // ================= 确定性 L1 证据性矛盾（2026-08-19）=================