The prop ledger path (armpaths.prop_rows -> props.projections_from_model divides the served line by playprob.p_play -> props.model_vs_market prices it with conformal bands from props.conformal_cal, fitted on (UNDIVIDED served line, actual|played)). How much does that mismatch over-price the OVER, and is it concentrated on men below p 0.95?

Owner lane: decision

Question

The prop ledger path (armpaths.prop_rows -> props.projections_from_model divides the served line by playprob.p_play -> props.model_vs_market prices it with conformal bands from props.conformal_cal, fitted on (UNDIVIDED served line, actual|played)). How much does that mismatch over-price the OVER, and is it concentrated on men below p 0.95?

Prediction

Frozen in prereg/2026-09-23T1201Z-h180-prop-band-conditioning.md before any run: H-180's 9-30 points REFUTED on magnitude -- bias_A (p<0.95, L0, pooled) positive but below +0.09; A-B between +0.01 and +0.05 and of similar size in the p>=0.95 bucket (band-wide contamination, not per-row). My prediction refuted if bias_A(p<0.95) >= +0.09 with CI excluding +0.05.

Finding and verdict

TESTED -- MECHANISM REAL, MAGNITUDE AND SHAPE REFUTED. The shipped prop path over-prices the OVER at p<0.95 by +0.072 [+0.053,+0.089] at a line at our own number -- below H-180's 9-30 floor pooled, but carries +0.120 [99% +0.072,+0.167] and rush_yards +0.079 reach it. Only +0.024 of that is the band/division mismatch H-180 names, and the same +0.020 hits men at p>=0.95 too. The other +0.048 is the conditional level itself (H-041). A consistent calibration is Brier-better at p<0.95 (-0.0029 [-0.0039,-0.0018]) and Brier-equal at p>=0.95. Decision consequence at L0: A clears the -110 break-even (0.5238) on 0.8% of rows; at p<0.95 those hit 0.421 (n~38), so the few overs the mismatch creates at low p are losers.

Reasoning

Kept the drawn row: it is the only open question on the props path whose premise could be checked in the code, and the ledger's prop_probability path is the one that feeds CLV work.

Next action

(a) Any change to conformal_cal belongs with the fix for effect 2, not before it: recalibrating on served/p alone (B) swaps +0.020 over-pricing for -0.013 under-pricing on healthy men. The fix that addresses both is a calibration on the pairs the prop path actually prices, WITH a level correction for low p -- the H-041 per-position-by-rank rescale, re-measured on P(over). (b) Real book lines (W-003, H-179) are needed before any of this can be scored as edge; these are synthetic lines at our own number. (c) pass_yards/pass_tds are outside this measurement.