Eval Conductor

Stop your AI evals from quietly regressing. Fail-closed release gate for scores you already produce (claude plugin eval, pytest, promptfoo, any JSON) — BLOCKs CI on regression, NaN/missing metrics, stale baselines and edited thresholds, with registry-driven thresholds and measured noise bands. Every PROMOTE/BLOCK emits a tamper-evident, signed, hash-chained eval receipt, and pre-registration lets you seal the bar before a run so a decision can prove it was judged against a bar fixed in advance.