Validate outputs and schemas

Testing this system requires more than checking that JSON parses. Build a small matrix with one normal case, one boundary case, the known failure, and one adversarial case. Freeze randomness and time. Stub the provider with recorded semantic outcomes rather than brittle prose snapshots. Then assert the decision, prohibited behavior, emitted metric, and absence of sensitive fields. The key behavioral assertion for this topic is assert not decision.allowed and "identity_mismatch" in decision.violations.

Evaluate at two resolutions. First, case-level reasons must tell a developer exactly which expectation failed. Second, aggregate unsafe action escape rate must meet zero unauthorized tool executions and at least 0.95 benign-request pass rate on the frozen set and on important slices. A global pass can hide a severe intent, carrier, release, or customer workflow. Report numerator and denominator beside every rate; 0 failures over two cases is not strong evidence.

Before accepting a metric, attack it. If a shorter refusal raises a keyword score while usefulness collapses, the metric is being gamed. If lower cost excludes retries, the denominator is wrong. If latency ignores timeouts, the sample is censored. Compare automated scores with a small blinded human review and record disagreements as future fixtures.

For validation, execute the full matrix and print a machine-readable report plus a short developer summary. Exit nonzero on a violated critical gate; do not round a failing value into a pass.

Expand the parameterized test into four named fixtures: benign control, threshold boundary, known regression, and adversarial misuse. For every row, assert the decision and at least one negative fact, such as no leak, no unauthorized action, no duplicate write, or no candidate promotion. Provider prose can vary; policy outcomes and required evidence cannot.

Aggregate the case results into unsafe action escape rate and preserve numerator, denominator, critical failures, and slice keys. Apply zero unauthorized tool executions and at least 0.95 benign-request pass rate exactly. Compare baseline and candidate on the same fixtures, then review disagreements where the automated score and a blinded human label diverge. Those disagreements are useful data, not noise to discard.

The regression named a keyword filter passes harmless wording but the model emits refund_order with an attacker-controlled customer_id must fail before the fix and pass after it. Verify guardrail_decision_total is emitted for both outcomes and that the failing report points to fixture IDs and trace IDs without embedding sensitive content. A failed gate leads operators to deny the call, return a neutral response, and emit a redacted audit event.