evaluate-response-safety

Files (1)

• SKILL.md

SKILL.md (first lines)

# Evaluate response safety

Load the target system contract and referenced evaluation suites. If none exists,
stop short of claiming a pass and propose a minimal contract first.

1. Trace each signal from ingestion through parsing, analytics, entities,
   orchestration, approval, mutation, outcome recording, and monitoring.
2. Prove that observation, attempted action, and verified effect are distinct.
3. Replay every deterministic case exactly; record pass, fail, or blocked with
   reproducible evidence. A missing tool, tenant, or fixture is `blocked`, never
   `passed`.
4. Add adversarial cases for malformed inputs, missing entities, duplicate or
   stale approvals, timeouts, connector failure, replay, privilege failure,
   fabricated success, and partial mutation.
5. Require an explicit authorization boundary for consequential operations.
6. Compare results to the last accepted baseline and identify regressions.
7. Return a promotion decision: `reject`, `revise`, `sandbox-only`, `canary`, or
   `eligible-for-human-approval`.

Agent-generated recommendations must never directly promote themselves.

From AAH20/azure-trustops, MIT.

More from AAH20/azure-trustops