Claude Code plugin that turns a theory, paper, model, or confident quantitative claim (including Claude's own) into a falsifiable, runnable test — instead of a debate in prose.
Use when someone shares a theoretical framework, whitepaper, derivation, or assertion and wants to know if it holds up, or when a claim can be settled by execution rather than argument.
The protocol:
- Operationalize — name every undefined symbol, flag inputs dressed up as outputs, and check the math actually matches the prose.
- Register predictions — numbered, numeric, stated before any code runs or any result is seen.
- Execute — minimal, deterministic, headless-safe code; prefers the user's own hands so the result doesn't route through Claude's judgment.
- Adjudicate symmetrically — score every prediction pass/fail, failures first, and call out results that were installed rather than derived.
No credentials or external services required.
MIT