simin-yuan's picture
self-auditing-agent — Volume 1: a forensic audit, with the auditor's own mistakes included (Simin Yuan)
51943f5
Raw History Blame Contribute Delete
2.77 kB
name: Verify the archive
# 这份档案的主张是「每条结论都能被第三方复现」。
# 那么它自己就必须在每次推送时被复现一遍 —— 主张不能只写在 README 里。
#
# 这里刻意区分两种"红":
# ① 主张的对称性破了(门禁不会说不 / 或只会说不)→ 必须让 CI 变红
# ② 盲区数量变了(gatecheck 找到的漏过条数) → 只记录,不让 CI 变红
# 理由:盲区条数会随输入和算子波动,把它设成门禁等于制造假警报;
# 但"基线还过不过"是硬事实,破了就是主张破了,必须红。
on:
push:
branches: [main]
pull_request:
workflow_dispatch:
jobs:
repro:
name: 两条命令 · 干净机器
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: actions/setup-python@v5
with:
python-version: "3.11"
- name: ① 门禁必须能说不、也能说是
run: python repro/verify_gate.py
- name: ② 复现核心发现(只读 SQL 接口的缺口)
run: python repro/verify_sql_gap.py
- name: ③ 重建合法基线(证明基线是可复现的,不是手搓的)
run: |
python repro/fixtures/make_valid.py
python repro/verify_gate.py
gatecheck:
name: 给自己门禁做体检
runs-on: ubuntu-latest
continue-on-error: true
steps:
- uses: actions/checkout@v4
- uses: actions/setup-python@v5
with:
python-version: "3.11"
- name: 变异 175 个输入,逐个撞门禁
run: |
python gatecheck/gatecheck.py \
--gate "python repro/validate_meta_model.py {target} --source repro/fixtures/valid-source" \
--target repro/fixtures/valid-meta-model \
--workers 4 --timeout 60 \
--report gatecheck-report.json || true
- name: 断言基线仍然可用(这条会真的让 CI 变红)
run: |
python - <<'PY'
import json, sys
r = json.load(open("gatecheck-report.json", encoding="utf-8"))
print(f" 基线通过 : {r['baseline_ok']}")
print(f" 变异总数 : {r['mutants_total']}")
print(f" 拦住 : {len(r['caught'])}")
print(f" 漏过 : {len(r['missed'])}")
if not r["baseline_ok"]:
print("::error::门禁连合法基线都不通过了 —— 主张已破")
sys.exit(1)
if r["mutants_total"] < 100:
print("::error::变异数异常偏低,体检本身可能失效了")
sys.exit(1)
PY
- uses: actions/upload-artifact@v4
with:
name: gatecheck-report
path: gatecheck-report.json