Scoring
The published formula, caps, and UNVERIFIED semantics.
The overall score is one decimal from 0 to 100: spec × 0.60 + security × 0.25 + maintainability × 0.15. Each bench is a severity-weighted ratio: critical 8, high 4, medium 2, low 1. PASS earns the full weight; FAIL and DRIFT earn zero; UNVERIFIED earns zero but is reported separately.
Caps
- Any critical spec FAIL or DRIFT caps the overall score at 59.9.
- Any critical security FAIL or DRIFT caps the overall score at 49.9.
- UNVERIFIED never triggers a cap; missing evidence is not treated as a confirmed failure.
- Applied caps are stored in the result and shown in a banner above the score.
Maintainability
A static import graph yields four metrics: dependency cycles, maximum fan-out, orphan ratio (entrypoint-aware), and module count. Each degrades linearly between published good and warn thresholds. Tests and mocks are excluded; only resolvable relative JS/TS imports create edges.
No model in the score path
The score module imports only deterministic task, rule, and result types. A test asserts that the import graph contains no provider or adapter module. Deterministic FAIL overrides any model PASS; model output cannot create rules, paths, verdicts, or scores.