Does a trailing-3 window that counts a non-ACT week (INA/RES/DEV/CUT) as a consumed row, so a man back from absence is served a one-game l3, cost the production projection, and would a window over his last three ACT rows recover it?

Owner lane: allocation

Question

Does a trailing-3 window that counts a non-ACT week (INA/RES/DEV/CUT) as a consumed row, so a man back from absence is served a one-game l3, cost the production projection, and would a window over his last three ACT rows recover it?

Prediction

Frozen in research/scientist/prereg/2026-09-23T0439Z-h224-roster-window.md (sha256 9c3b1b97...9e60), written while the shipped arm was fitting and before any arm output existed: pooled nil (|delta| < 0.01); age 4+ roster3 beats shipped by -0.10 to -0.30, interval clear of zero; age 2-3 by -0.03 to -0.12 spanning zero. REFUTED if age 4+ is within +/-0.05 or roster3 is WORSE there. Addendum (premise count, no outcome): 18% of scored rows are treated, not ~7%; pooled band kept, not widened.

Finding and verdict

TESTED. Pooled nil HELD (+0.009 [-0.005, +0.022]); the age-4+ prediction REFUTED (roster3 +0.093 worse, interval [-0.024, +0.199]). Moving the six l3 columns to an ACT-row unit does not improve the production projection on any stratum, although it improves the raw fp_l3 by 0.54 pts at age 4+. Room totals: projected targets per team-game move by 0.34 mean abs (mean -0.04), carries 0.017.

Reasoning

The allocation chain absorbs a window-unit change that a naive statistic rewards, the same shape as D-198 and H-021. The thin window is not the error source it looked like from the served-row penalty (H-095).

Next action

(1) the 4+ worse-than-shipped point (p=0.065 against shuffled labels) is the interesting half: ask whether a return-from-absence ramp exists by fitting the target-share residual on weeks-since-return, with the pre-absence level as a separate term; (2) played3-strict is the untested family member; (3) the defensive windows (H-215) use a third unit, played-only defender_week rows, and nobody has compared units there.