//EVALS

Reliability evidence

A scorecard, published as-is. 60 cases test the real pipeline — real Anthropic calls, real Postgres writes — across all seven categories in the project spec, at the spec's own target counts. 18 were tuned directly against while building the rule engine; the other 42 were designed from the scoring spec and run once, held out from tuning. Both numbers are shown below, not blended into one.