Does serving the season-mean pace block (what PM.latest_rows hands the model) instead of the trained trailing-3 block cost accuracy on the target row, and how much is the team's own pace block worth at team-volume level versus player level?

Owner lane: forecast

Question

Does serving the season-mean pace block (what PM.latest_rows hands the model) instead of the trained trailing-3 block cost accuracy on the target row, and how much is the team's own pace block worth at team-volume level versus player level?

Prediction

From the H-076 row (2026-09-14), restated unchanged in prereg/2026-09-23T1558Z-h076-served-pace-block.md before the run: (b)-(a) nil within +/-0.01 pts/wk; (c)-(a) better by under 0.02; (d)-(a) worse by under +0.02. REFUTED if (b)-(a) point estimate outside +/-0.015.

Finding and verdict

MEASURED NOTHING at the player (prediction held): served season mean +0.0037 [-0.0057,+0.0131] pts/wk; hindsight season mean +0.0013; league mean +0.0077; shuffled team +0.0067. At team volume the same block is worth 0.20 throw attempts over a random team's, and the served swap is still neutral (+0.034 throw attempts).

Reasoning

The swap H-076 names is neutral at both layers. What the run adds is the layer gap: a pace block worth 3% of team attempt error is worth under 0.2% of player fantasy error, the same attenuation share.py records for the drive model (3% of a team number does not survive division among twelve players).

Next action

The attenuation itself is the open question: a 0.11-attempt hindsight gain in volume_for is worth +0.0013 pts/wk at the player. Test whether a perfect team-attempts override (volume_for's team_pass_att_pred hook) moves player MAE, to bound what ANY volume improvement can buy through the current chain (research/measure_error_decomposition.py has the hook).