Measured. Not believed

"Our model wins 68%." "The strategy returns 8% a day." "We're #1 on the leaderboard." Every one is a number computed from a model of reality and handed to you as if it *were* reality — and from the outside, an inflated number looks exactly like an honest one. You have to go and check.

This is a field guide to checking. It recomputes the AI benchmarks that rank today's models — a celebrated judge that rewards length over quality, a #1 ranking no better than a coin flip — and post-mortems a trading strategy carried to real money and lost, for reasons you could measure first: the fee bigger than the edge, the queue no backtest can simulate.

Both meet in one discipline: four questions that catch a lying number, in a leaderboard or a bot. Then the author turns them on himself — a discovery that survived cross-validation, held-out testing, and every robustness check, and was still an artifact — and on twenty-three recent papers making the same quiet mistake.

No hype. No secret edge. Just the reflex that saves a launch, a fundraise, or three months and $53 — and the nerve to run it on your own work.