On close same-position start/sit pairs, does sim_world's centre (the centre of the served RB/WR/TE band, never published) order the pair better than the published lines_for projection, on seasons the discovery never scored?

Owner lane: forecast

Question

On close same-position start/sit pairs, does sim_world's centre (the centre of the served RB/WR/TE band, never published) order the pair better than the published lines_for projection, on seasons the discovery never scored?

Prediction

C1 about +2.5 pp, 99% interval excluding 0; C2 similar; 4 of 4 seasons positive. Frozen in experiments/2026-09-26T0834Z-h292-sim-centre-heldout.prereg.json (commit 4b802378).

Finding and verdict

H-292 CONFIRMED HELD OUT. On 2019-22 (seasons H-207 never scored), across 41,558 close same-position RB/WR/TE pairs (lines within 1.0, both >= 6, played=1 SELECTED), the man with the higher sim_world centre outscores the other 54.0% of the time, against 51.5% for the man with the higher published line: sim minus line +2.53 pp [99% +0.99, +4.06], outside the shuffle null; +3.40 pp [+1.64, +5.21] with both men at p_play >= 0.95; positive in all four seasons (2019 +2.98, 2020 +2.64, 2021 +3.63, 2022 +0.85). The board already computes a centre that orders close start/sit calls better than the number it publishes, and serves it only as the middle of the band. For scale: H-017 measured the free consensus beating shipped on close calls by 3.09 pp.

Reasoning

A second centre that orders close calls better, but is the worse level forecast, points to a blend or a pair-level combiner rather than a replacement. That is a promotion-grade question and needs review first.

Next action

(1) a walk-forward blend line + b*(sim_mean - line) fitted on seasons < S, graded on BOTH MAE and close-pair order (the sim is the worse level); (2) the market rival: re-run sim_world with team_week total/spread set to the league mean and see if the order edge survives; (3) the start/sit gap to the consensus (H-017) on the same pairs.