What problem does it solve? Engineers and AI agents often present valid proofs that silently answer a different question than the one the change requires: a green test suite over skipped tests, a reproduction from a different caller, a self-review blind to its own prose, or a ruling resting on an untested premise. This Skill forces you to name what your evidence does NOT cover before calling something proven. ## Core Features & Use Cases - Premise Auditing: Separates asserted premises from verified facts at delegation time, requiring each premise to be marked VERIFIED or ASSUMED. - Independent Reader Enforcement: Applies the rule that verifier and verified must not share the same reader, author, frame, or question set, turning comment assertions into load-bearing tests. - Surface-Matched Proof: Detects when a by-construction argument covers the wrong provider, config, or platform, and requires skipped tests to execute-and-pass by name with a dropping skip count. - Caller-Faithful Reproduction: Distinguishes observing the system produce a failure from reproducing a look-alike error with a hand-run query, demanding instrumentation before fixes. - Use Case: Before merging a Postgres-only change validated by a green SQLite suite, use this Skill to identify that the proof ran on the wrong surface and require the opt-in real-Postgres facts to execute and pass by name. ## Quick Start Ask the AI to review your proof or root-cause analysis using the proof-covers-the-wrong-thing checklist and state in one sentence what the evidence does not cover.
