The H8 stop recorded a COUNT of fit-side rates changed by a physical reorder (10,891/10,940). What is H8 at the grain a forecast consumer reads, across three held-out seasons; is all of it efficiency.py:182; and does it block a volume contrast such as H-153?

Owner lane: Not recorded

Question

The H8 stop recorded a COUNT of fit-side rates changed by a physical reorder (10,891/10,940). What is H8 at the grain a forecast consumer reads, across three held-out seasons; is all of it efficiency.py:182; and does it block a volume contrast such as H-153?

Prediction

Frozen 2026-09-26T2117Z in prereg.json before any arm ran: H8 at consumer grain small against its count (mean |dFP| 0.01-0.05, max ~2; mean |dtargets| < 0.02; ORD-S MAE interval spans 0 every season); S2==S and ORDP==SP on 100% of rows; H-153 top-of-room volume bias under S within +/-0.3 per game and not exceeded by targets_l3's bias (premise NOT supported); S-vs-ORD change in that bias < 0.05.

Finding and verdict

REFINED STOP (the H8 stop stands for exact replication; its scope is narrower than stated). H8 at consumer grain, 2023-25 held out: reordering identical rows moves 6,134-6,155 of ~6,840 fantasy forecasts, 5,376-5,423 target and 4,648-4,746 carry lines per season, but by mean |dFP| 0.028-0.036 pts (p99 0.19-0.29, max 1.97), mean |dtargets| 0.013-0.018, mean |dcarries| 0.008-0.012; walk-forward MAE ORD-S is -0.0003..+0.0002 pts with every 95% interval inside +/-0.0021 (targets inside +/-0.0012, carries +/-0.0010). S2==S, COPY==S and ORDP==SP on every row of every season: efficiency.py:182 is the ONLY source. H-274's candidate fix (latest position) itself moves accuracy by -0.0004/+0.0007/-0.0001 fp MAE, all intervals spanning 0. H-153 (registered, S-ranked top of room): WR/TE top over-projected by +0.63/+0.69/+0.99 targets/game vs targets_l3 +0.20/-0.10/+0.14, but MAE model minus trailing -0.03/-0.05/-0.03, all spanning 0; RB top not over-projected (+0.24/-0.10/-0.03 carries) and the model beats carries_l3 by -0.75/-0.32/-0.51 (2 of 3 exclude 0). H8 moves that bias by <=0.009. My registered prediction for WR/TE bias (within +/-0.3) was WRONG; the RB half and the 'trailing had it right' half held.

Reasoning

The stop's statistic (10,891 rates changed) is a count on the fit side; the consumer reads forecasts. At forecast grain H8 is a precision defect of size ~0.03 pts/row and ~0.0003 pts MAE: it forbids bit-identity claims and any refit contrast under ~0.002 pts MAE or ~0.001 targets, and it forbids nothing larger. The H-153 top-of-room contrast is 100-400x its H8 floor. Post-hoc: ranking the room by targets_l3 instead of the model flips the asymmetry (trailing +0.68/+0.68/+1.01 vs model +0.25/+0.13/+0.67), so most of the 'board over-projects the top' is selection on the predictor (winner's curse), not a volume defect. On the rows where both rankings agree (391/362/376) both sources over-project by +0.30..+0.96 targets; that shared residual is not explained here (candidate: in-game exits inside a played=1 population).

Next action

H8 stays red until H-274 ships; a lane may run a refit contrast while it is red provided its effect exceeds ~0.002 pts fp MAE / ~0.001 targets MAE and it claims no bit-identity. H-153 released TESTED.