Commit 5963de5
feat(detector): findings 輸出格式紀律(evidence 裸 token + severity enum)
真實盲測照出:agent 常把 evidence 寫成帶敘述的字串("trace_path: 乘積…")或用
非 enum 的 severity("MAJOR"),導致計分時被靜默判為證據不完整/定級未達。
在 detector.md 與 security-fraud-detector.md 的證據標註步驟加入「輸出格式紀律」:
evidence 只填裸 token(trace_path/logic_invariant/state_mutation/guard_absence,
不得加冒號或敘述,理由寫進 rationale);severity 只能是 info|low|medium|high|critical。
附正確/錯誤範例。配合 validate_findings.py 的 fail-closed 驗證形成根因+執行雙層防線。
驗證:修正後重跑盲測,agent 產出乾淨 token,candidate evidence_complete 0→1.0、
severity_hit 0→1.0。
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>1 parent 96b0a9f commit 5963de5
2 files changed
Lines changed: 12 additions & 0 deletions
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
96 | 96 | | |
97 | 97 | | |
98 | 98 | | |
| 99 | + | |
| 100 | + | |
| 101 | + | |
| 102 | + | |
| 103 | + | |
| 104 | + | |
| 105 | + | |
99 | 106 | | |
100 | 107 | | |
101 | 108 | | |
| |||
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
67 | 67 | | |
68 | 68 | | |
69 | 69 | | |
| 70 | + | |
| 71 | + | |
| 72 | + | |
| 73 | + | |
| 74 | + | |
70 | 75 | | |
71 | 76 | | |
72 | 77 | | |
| |||
0 commit comments