tzhouam · Oct 3, 2026 · Oct 3, 2026 · Oct 3, 2026 · Oct 3, 2026 · Oct 3, 2026
diff --git a/eval/jiuwenswarm_ab_README.md b/eval/jiuwenswarm_ab_README.md
@@ -118,3 +118,44 @@ started/completed/failed 检查点及身份摘要约束恢复；Codex 基准与
 ```
 
 上述测试不调用模型；覆盖固定输入、真实 merge-base、工具隔离、预算、引用校验、恢复与计数。完整原生日志保存在外部归档，提交到 Git 的报告使用精简指标和摘要。
+
+## 协议修复后复测
+
+原始 72 次结果及其失败保持不变。复测通过 [准备工具](jiuwenswarm_ab_retest.py)核对已发布报告的哈希绑定、12 个 PR 的固定源码、文档快照及模型预算，再创建独立外部目录。复用原来评审前冻结的 Codex 问题清单，包含 #7656 的未知状态；只重绑定实验元数据，问题正文、数量和原始记录另存哈希，不重新抽样问题基准。
+
+```bash
+AB_PREVIOUS=/absolute/path/outside/repository/jiuwenswarm-docs-ab-original/evaluation-v2
+AB_RETEST=/absolute/path/outside/repository/jiuwenswarm-docs-ab-retest
+AB_STUDY="$AB_RETEST/evaluation-v2"
+
+"$AB_PYTHON" eval/jiuwenswarm_ab_retest.py prepare \
+  --previous-study "$AB_PREVIOUS" --run-root "$AB_STUDY"
+
+"$AB_PYTHON" eval/jiuwenswarm_ab_judge.py preflight \
+  --campaign "$AB_STUDY/campaign.json" --preflight-tag protocol-fix
+
+"$AB_PYTHON" eval/jiuwenswarm_pr_review_ab.py preflight \
+  --campaign "$AB_STUDY/campaign.json"
+
+"$AB_PYTHON" eval/jiuwenswarm_pr_review_ab.py run \
+  --campaign "$AB_STUDY/campaign.json" --workers 13 \
+  --truth-manifest "$AB_STUDY/private-codex/truth-manifest.json"
+
+"$AB_PYTHON" eval/jiuwenswarm_pr_review_ab.py collect \
+  --campaign "$AB_STUDY/campaign.json"
+
+"$AB_PYTHON" eval/jiuwenswarm_ab_judge.py score \
+  --campaign "$AB_STUDY/campaign.json" --workers 4
+
+"$AB_PYTHON" eval/jiuwenswarm_ab_retest_report.py \
+  --previous-study "$AB_PREVIOUS" --study "$AB_STUDY" \
+  --output-dir "$AB_RETEST/report"
+```
+
+此次修复统一评估提示词与严格输出校验：所有评审条目都有字符串锚点，无法提供精确锚点时为空字符串；重复候选合并在 findings 中。目录检索允许安全的 `.` 根目录别名与一个尾部 `/`，文件读取和路径边界保持严格。提示词工具使用连续游标，完整读完后才开放源码与文档工具。原生预检实际分页读取大型 PR 的上下文和完整 diff，验证修复后的目录写法及规范空评审格式。
+
+所有 MCP 工具关闭 `structured_output`，防止字符串结果额外渲染同一份结构内容。分页同时限制序列化结果的字符数和 UTF‑8 字节数为 24,000；每次评审直接审计哈希绑定的原生 ToolResult，拒绝超限、适配器截断或尺寸记录不一致的结果。源码及计算请求按原生调用 ID 跨传输尝试累计，包含执行前被拒绝的请求。原生记录不包含最终 provider 请求体，该层序列化保持未知。
+
+发现协议问题后中止的批次单独归档，正式复测使用新目录和冻结身份；不得把中止后未派发的槽位当作模型失败，也不得混入新批次的 72 次分母。原始报告保留，但旧原生输出未通过本次渲染规则的认证。
+
+预检不计入正式 72 次；预检失败则停止正式派发。代码和输入冻结后不修改协议；无效内容和格式不补采样、不事后晋升。新报告同时列前后完成率、有效评审的条件精确率、有限基准召回和速度，记录提示词协议变化与服务排期差异，不能将前后差额单独归因于知识库。
diff --git a/eval/jiuwenswarm_ab_retest.py b/eval/jiuwenswarm_ab_retest.py
diff --git a/eval/jiuwenswarm_ab_retest_report.py b/eval/jiuwenswarm_ab_retest_report.py
diff --git a/eval/jiuwenswarm_pr_review_ab.py b/eval/jiuwenswarm_pr_review_ab.py
diff --git a/eval/knowledge-depth/README.md b/eval/knowledge-depth/README.md
@@ -10,6 +10,8 @@ All 490 previously active blocks retain exact bytes and native approval bindings
 
 Current delivery reports:
 
+- [Original author-document comparison](jiuwenswarm-original-docs-comparison-cn-20261003.md), with [compact data](jiuwenswarm-original-docs-comparison-cn-20261003.json), preserves the content/coverage baseline and historical A/B results unchanged.
+- [GLM-5.3 protocol-fix retest](jiuwenswarm-ab-retest-cn-20261003.md), with [compact data](jiuwenswarm-ab-retest-cn-20261003.json), reports the new evaluation batch separately from that immutable baseline.
 - [Chinese comparison and PR-review impact](jiuwenswarm-repair63-comparison-cn-20261002.md), with [compact data](jiuwenswarm-repair63-comparison-cn-20261002.json).
 - [Final audit summary](jiuwenswarm-repair63-final-20261002.json), [source bindings](jiuwenswarm-repair63-source-depth-20261002.json), and [actual native approval bindings](jiuwenswarm-repair63-native-approvals-20261002.json).
 - [237-case retrieval acceptance](jiuwenswarm-repair63-retrieval-after-20261002.json), [independent review](jiuwenswarm-repair63-independent-review-20261002.json), and [local validation](jiuwenswarm-repair63-local-validation-20261002.json).

diff --git a/eval/knowledge-depth/jiuwenswarm-ab-retest-cn-20261003.json b/eval/knowledge-depth/jiuwenswarm-ab-retest-cn-20261003.json