What problem does it solve? Skill claims are often accepted without rigorous testing, leading to unreliable automation. This Skill applies Popperian falsification to test skill declarations against real task inputs, producing auditable evidence chains and quantitative trust grades instead of subjective assessments. ## Core Features & Use Cases - Falsification Testing: Executes tests with real task inputs (never mock data), including counter-evidence cases designed to fail, and archives full evidence chains to test-results/. - Bayesian Trust Scoring: Updates per-skill trust scores via scripts/update_trust.py, mapping results to four grades (High Trust 0.9-1.0 down to Unreliable <0.5) based on accumulated evidence. - Golden Test Set: Ships with golden cases and expected outputs covering basic validation and counter-evidence scenarios with weighted pass criteria. - Use Case: Verify a claim like "citation-verification detects ≥95% of fabricated references" by running it against a real paper corpus, archiving inputs, logs, outputs, and metrics, then updating the skill's trust grade. ## Quick Start Run falsification tests on the citation-verification skill's claim using the real pima-crispdm paper directory and update its trust score with the results.
