What problem does it solve? Agents that grade their own output fail in measurable ways: intrinsic self-correction degrades reasoning accuracy, and LLM judges score AUROC 0.18-0.30 on false-success detection, worse than chance. This Skill replaces model self-assessment with verification signals derived mechanically from code contracts — type signatures, postconditions, exception classes, and existing tests — and forces every derived verifier to be validated against an independent artifact or marked provisional. ## Core Features & Use Cases - Signature-to-verifier mapping: Converts parameter types, return annotations, postconditions, raised exceptions, and docstrings into input guards, output validators, state verification checks, and typed failure taxonomies. - Validate-or-mark-provisional rule: Requires every derived verifier to name its derivation, be checked against an independent artifact (OpenAPI doc, schema file, SDK, recorded request), and carry provenance and confidence metadata; consumers fail closed on provisional verifiers. - Typed terminals and termination design: Defines Done/Failed/BudgetExhausted/NeedsHuman/Aborted terminal states, separates evaluate from decide, and specifies budget, no-progress, and oscillation detectors. - Use Case: When adding a critic or reflection loop to an agent, use this Skill to replace "review your answer" prompts with postcondition recomputation against the application's own API, and to record validated_against and confidence fields on every emitted node contract. ## Quick Start Ask the agent to design the verification and termination logic for a node using contract-derived checks instead of an LLM judge, and to mark any unvalidated derived field as provisional.
