At a fixed projection level, is intervals.shipped_band's ceiling exceeded more often against defences that allow more fantasy points to the position (soft) than against those that allow fewer (hard)?

Owner lane: calibration

Question

At a fixed projection level, is intervals.shipped_band's ceiling exceeded more often against defences that allow more fantasy points to the position (soft) than against those that allow fewer (hard)?

Prediction

From the H-203 row (2026-09-15), unchanged in the prereg: a small real gradient on WR and TE, soft minus hard ceiling exceedance +1 to +3pp inside a cell, Bonferroni-4 interval excluding zero on WR only; QB and RB show no gradient; a tercile-keyed curve gains under 0.5% of pinball-0.90 pooled. REFUTED if every position's gap is inside +/-1pp with intervals spanning zero.

Finding and verdict

MIXED, PREDICTION WRONG ON LOCATION. The level-only ceiling does not under-cover receivers against soft defences (WR +0.1pp, TE +0.3pp, both inside +/-2pp). It does on running backs: +2.7pp [+0.1, +5.2] more ceiling exceedances against the top tercile of RB-points-allowed than the bottom, at the same projection level. Keying the curve on the defence tercile does not recover it -- pinball-0.90 gets worse pooled (-1.9%) because each tercile curve fits on a third of the rows. Exploratory and post-hoc: half the RB gradient is the implied team total, which shipped_band ignores and which moves RB ceiling exceedance from 6.1% to 11.5%.

Reasoning

A level-only key asserts the spread depends on nothing but the level. For receivers that assertion held against the defence scalar; for backs it fails, and the post-hoc split points at game exposure rather than defensive quality. Filed as a new frozen question rather than claimed from this run.

Next action

New hypothesis filed as H-255: does an itt-keyed fit_shipped (exposure terciles, EXPOSURE_CUTS) beat the level-only curve on pinball-0.90 for RB, frozen before a run. Premise correction: forecast.py:542 serves sim_world p10/p90 for non-QB rows when simulate_week returns them and shipped_band only for QB and as fallback, so this measures the fallback curve for RB/WR/TE.