Models drift and prompts regress silently. Freeze a golden dataset of known-good cases; every time the Agent Constitution changes, run the new logic against it and gate the ship on the pass rate — exactly the conformance rigor EcoCloud runs on its own kernels. A missing result is a failure, and a single regression blocks the build even if you're above threshold.
Honest scope. A deterministic scorer + gate — it compares the new logic's outputs to the golden expectations (deep-equal, order-independent; or by a single key) and returns ship/block against your threshold; it does not generate the cases or run the agent (you supply both). It's the build gate the rollout kernel ramps after. FAIL-CLOSED — an empty golden set, a threshold outside [0,1], a missing result, or any regression never ships. Source: flowdesk-eval-kernel.js.