Every number here is out-of-sample: the board's weekly player ordering, scored against what actually happened, week by week, on data the model never trained on. Most projection products never publish their misses. This is the whole point — the metric is honest because you can check it.
Our published board graded against free market projections on identical player populations (a player-week counts only if every provider projected it). The claim rule is strict: we say "best" at a position only when we beat every provider on it — and we publish the positions where we lose, too.
Each week the published board is graded against what actually happened: top-5 hit rate, slate ordering (ρ), measured floor–ceiling range coverage vs the 70% target, and boom-call precision (players we flagged at P(boom) ≥ 50%). Measured numbers, misses included — QB shows no range because we don't serve a QB band.