falsify

Scans provided text for empirical claims and checks them against locked verdicts. Trigger phrases "audit this claim", "check this text", "is this accurate", "does this match our verdicts", "review this PR description", "lint this commit message", "fact check against falsify".

Review a git diff (staged, unstaged, or PR branch vs base) for falsify honesty violations. Flags new specs that aren't locked, threshold changes without relock, metric_fn references to nonexistent modules, and stale lock files. Call this skill before merging any PR that touches claims/*.yaml or metric files.

Diagnose a failing falsify CI run. Reads `make release-check` output, unittest failure tracebacks, and smoke/dogfood stderr. Identifies which of the 12 gates failed and why, cross-references likely causes (spec hash mismatch, stale verdict, missing dependency, regressed test), and suggests exact fix commands. Call this skill when release-check returns exit 1 or when CI is red and the cause is not obvious.

Triggers when user makes empirical claims or asks to verify results. Trigger phrases include "testing whether", "let's verify", "does this work", "is it true that", "empirical claim", "I think X performs better", "run the experiment", "check the threshold", "verdict", "falsify this".

Use when the user makes an empirical claim (e.g., "X improves Y", "our filter works", "results show…") and hasn't yet written a falsifiable spec. Converts natural-language claims into a locked hypothesis.yaml. Triggers on phrases like "I believe", "our data shows", "this works", "we proved", "testing whether", "hypothesis".