provelisted

# prove `grill` stress-tests one item. `prove` checks many claims by recomputing each from the source of truth and running adversarial refutation against the survivors, then tags every claim with a confidence verdict. Use it when being wrong in public is expensive. ## When to use it - Numbers are going to stakeholders. - A finding is contested or a strategy rests on a few key measurements. - A claim must survive independent scrutiny before you assert it flatly. - Skip it for a small exploratory finding where the cost of being wrong is low. ## Required inputs - The claims to check, one per line or bullet. - The source of truth for each claim: the file, field, or dataset the number should come from. - The cheapest falsifier for each claim: the test that would break it first. ## Steps 1. Frame falsifiable hypotheses. Rewrite each claim as something that could be shown false, and name the cheapest test that would break it. 2. Recompute each hypothesis from the source of truth itself, not from a prior summary or doc. Record the value, the method, the file and field, and whether the result matches the claim. A claim with no named source of truth cannot be recomputed: route it to UNKNOWN and name the source that would let it be checked, rather than confirming it. 3. Refute the survivors, cheapest falsifier first: cherry-picking, circular definition, selection bias, wrong baseline, arithmetic slip, hidden dependency. A refutation attempt ends one of three way