Skip to content

feat(ISSUE-263): W5-C1 golden set 标定 (#286) - #84

Merged
cloudbrid-agent[bot] merged 1 commit into
mainfrom
feat/issue263-w5c1-w2c4
Aug 23, 2026
Merged

feat(ISSUE-263): W5-C1 golden set 标定 (#286)#84
cloudbrid-agent[bot] merged 1 commit into
mainfrom
feat/issue263-w5c1-w2c4

Conversation

@cloudbrid-agent

Copy link
Copy Markdown
Contributor

Summary

  • W5-C1 (AC-8): golden set 建设——含已知不合格样本(构造独立于判定脚本)+ 盲化注入
    (剥 id/source/type/expected_verdict 元数据,与正常样本同批混排);每次 run 全量重放
    • 回归断言(不合格样本仍不合格;变合格 = gate 失效 = 红);criteria 变更重新标定
      (SHA 强一致,标定记录留档)。
  • W2-C4 (AC-3/AC-17): holdout 注册/校验——sealed_sha256 校验复用 ADR-0068 揭封公式;
    PR 引用 hash 与注册一致性;验证者 APP 身份校验(verifier_app.id=null 时降级留痕,
    代码先实现,IFACE-01 时序约束)。

Test plan

  • python -m pytest pipeline/adversary/tests/test_golden_set.py — 14 passed
  • python -m pytest pipeline/adversary/tests/test_holdout_registry.py — 11 passed
  • python pipeline/adversary/golden_set.py regress --mode native — 4/4 回归通过
  • python pipeline/adversary/calibrate.py run/verify — SHA 强一致校验通过
  • python pipeline/adversary/holdout_registry.py verify-hash/check-pr/register
    tamper 检测 / 未知 id 拦截 / 非验证者 403 全命中

Closes Cloudbird-Software/.github#286
Relates to Cloudbird-Software/.github#276

golden set 建设(W5-C1 .github#286 / AC-8):
- pipeline/adversary/golden_set.py:golden set 加载、盲化注入(剥来源元数据同批混排)、
  全量重放、回归断言(native 模式确定性,免 LLM;llm 模式可选)
- pipeline/adversary/calibrate.py:criteria 变更检测 + SHA 强一致校验 + 标定记录落盘
  (标定 SHA / golden 版本 / 定标时间机器可查;criteria 变更未重标定 = fail-closed 判红)
- pipeline/adversary/fixtures/golden/samples/:已知不合格样本(构造独立于判定脚本:
  全零聚合分 / 边界 0.69 vs 0.7 / token 账 void / 非 golden 随机低分反特判)
- pipeline/adversary/holdout_registry.py:holdout 注册 / sealed_sha256 hash 校验
  (复用 ADR-0068 揭封公式)/ PR 引用一致性 / 验证者 APP 身份校验
  (W2-C4 .github#276;verifier_app.id=null 时降级留痕,代码先实现)
- pipeline/adversary/tests/:golden_set 14 例 + holdout_registry 11 例,全绿

Closes Cloudbird-Software/.github#286
Relates to Cloudbird-Software/.github#276
@cloudbrid-agent
cloudbrid-agent Bot enabled auto-merge (squash) August 23, 2026 09:43
@cloudbrid-agent
cloudbrid-agent Bot merged commit 4a12443 into main Aug 23, 2026
21 checks passed
@cloudbrid-agent
cloudbrid-agent Bot deleted the feat/issue263-w5c1-w2c4 branch August 23, 2026 09:43
cloudbrid-agent Bot added a commit that referenced this pull request Aug 23, 2026
PR #84 added files at CI-Workflows/pipeline/... (nested) instead of pipeline/... (repo root) due to a path-prefix error. Canonical implementations already merged via PR #82. Removing duplicates.

Co-authored-by: cloudbrid-agent[bot] <318150333+cloudbrid-agent[bot]@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

W5-C1: golden set 标定(盲化反例 + 逐 run 重放 + 变更重标定)

0 participants