prove-me-wronglisted

# Prove Me Wrong > Your agent has a theory. Make it try to kill it. Use this at the commitment point: a leading claim is about to drive a substantial fix or be declared correct. It is not a prediction log for debugging steps. Skip mechanical edits. ## Write the attack card Keep the card to four lines: ```text Claim: <specific causal explanation or behavior the fix guarantees> Smallest falsifier: <one observable result that contradicts the claim> Strongest plausible competitor: <an alternative that fits known facts> Cheapest safe attack: <a controlled check that can expose the falsifier> ``` Make the claim narrow and observable. Choose a credible competitor with a different prediction. The attack must exercise the conditions under which the claim predicts an outcome. Missing the relevant path or state is inconclusive, not falsification. ## Attack the claim Run one realistic controlled counterexample before accepting the claim: 1. Preserve the baseline and change only the variable that distinguishes the claim from its competitor. 2. Prefer a focused reproduction, test, query, log, toggle, or temporary instrumentation over a broad rewrite. 3. Record observation separately from interpretation, then remove temporary instrumentation. For a completed fix, construct the counterexample within the requested workflow. Derive it from a boundary, state transition, concurrency path, integration seam, or surviving competitor—not an arbitrary edge case. When practical, show a regr