VibeGuard precision report

June 13, 2026 · View on GitHub

Precision, recall and F1 on the labeled AI-PR corpus under tests/fixtures/corpus/, measured at the actionable tier (HIGH/CRITICAL) — the decision vibeguard gate --fail-on high makes. A case is blocked when it produces a finding at or above HIGH in the relevant rule family.

  • Precision — of the cases the gate blocked, the fraction that were genuine risks.
  • Recall — of the genuine risks, the fraction the gate caught.
  • Detection recall — the fraction of true-positive cases that produced any finding, regardless of severity (advisory-tier rules like risky_diff/ai_footprints detect but don't block).
Rule familyPrecisionRecallF1TPFPFNTNDetection recall
ai_footprints1.000.500.6710111.00
auth1.001.001.0030011.00
dependencies1.001.001.0020011.00
packagingn/a0.000.0000111.00
risky_diffn/a0.000.0000321.00
secrets1.001.001.0040031.00
sourcemaps1.001.001.0010011.00
Overall1.000.690.811105101.00

The live regression guard is tests/test_corpus_precision.py, which fails CI if any tp_ case stops firing or any fp_ case starts producing an actionable (HIGH/CRITICAL) finding. Regenerate this report with make bench-precision after adding corpus cases.