falsify

Teams ship AI accuracy claims that evaporate under scrutiny not from malice, but because the claim was never structured to be falsifiable. A "94% accuracy" number becomes 71% in production and nobody can tell whether the model drifted, the metric got relaxed, or the threshold moved after seeing the data. Falsify is a CLI that forces pre registration: you cryptographically seal the metric, threshold, and dataset with SHA-256 before the experiment runs. Post hoc edits break the hash. CI exits 3. The audit trail writes itself. Deterministic exit codes are the API — 0 pass / 10 fail / 3 tampered. CI gates on the code, humans read the trail. The claim either survives contact with the data, or it didn't happen. Built entirely with Claude Opus 4.7: 5 skills, 2 subagents, 3 slash commands, 1 MCP server. 514 tests passing.