Does the one-banded-residual-distribution-per-quantity calibration (props.conformal_cal: bands by projected level, pooled over positions) over-price RB receiving OVERs by 10-13 points and QB rushing-attempt OVERs by 14 points, and does a per-position (Mondrian) calibration remove it?

Owner lane: s-observation

Question

Does the one-banded-residual-distribution-per-quantity calibration (props.conformal_cal: bands by projected level, pooled over positions) over-price RB receiving OVERs by 10-13 points and QB rushing-attempt OVERs by 14 points, and does a per-position (Mondrian) calibration remove it?

Prediction

Frozen in prereg/2026-09-23T1317Z-h181-prop-position-pooling.md before any run: REFUTED on magnitude for RB receiving -- bias_A in [0.00,+0.08] for both RB receiving cells; QB carries bias_A in [+0.03,+0.12]; per-position bands reduce |bias| in all three cells and are Brier-better or equal. Refuted if any RB receiving cell has bias_A >= +0.10 with 99% CI excluding +0.08, or QB carries bias_A < +0.03.

Finding and verdict

TESTED -- H-181 MAGNITUDES REFUTED; POSITION HETEROGENEITY REAL BUT TWO-SIDED. RB receiving OVERs over-priced +0.045 (receptions) / +0.043 (rec_yards) at a line at our own number, not +0.10..+0.13; QB rushing attempts -0.006 [99% -0.052,+0.037], not +0.14. A per-position (Mondrian) band moves RB receptions by +0.030 (perm p 0.015 against a 200-shuffle calibration null); RB rec_yards' +0.015 is inside calibration noise (p 0.15). The pooled band's larger error is on TE, where it prices OVERs LOW by 0.036 (rec_yards) -- selected post hoc. Per-position is Brier-better on the three primary cells but Brier-worse or equal on WR/TE receiving, so it is not a straight replacement. QB carries cannot be banded per position from one calibration season.

Reasoning

Kept the drawn row. Mapped the claim onto props.conformal_cal (bands by level, pooled over positions) and reused H-180's production-path harness, adding a Mondrian arm.

Next action

(a) The frozen A-P interval method is wrong for calibration contrasts: any future calibration comparison must resample calibration rows (or use a shuffle null), not only scored weeks -- H-180's A-B CI [+0.023,+0.024] has the same defect. (b) A per-position calibration needs >1 season of calibration rows to band QB carries/rush at all; test a 2-season calibration window with position strata. (c) TE under-pricing (-0.036) and WR rush_yards over-pricing (+0.092) deserve their own frozen run, since they were found by looking. (d) Real book lines, not floor(cond)+0.5.