Eval Gate in CI Diagram

Give every control teeth

• A failing eval blocks the build instead of filing a finding
• The threshold traces to a named failure mode or obligation
• The gate is built at the earliest point it can block

Evidence accrues automatically

• Structured result: suite id, model version, score, threshold, verdict, timestamp
• Filed against the registry entry for the model
• Regressions are caught before production

Trade-offs

• Eval maintenance and run-time cost in CI
• Thresholds tuned to avoid flaky gates