H-151: does playprob serve P(DRESSED) rather than P(takes an offensive snap), 18-28% apart on the bench? Refined before running: for a dressed man the served multiplier is the table (confirmed_active only ever stamps 0.0), fitted on on_roster=1=ACT; is the bench's served error in p or in the conditional line?

Owner lane: allocation

Question

H-151: does playprob serve P(DRESSED) rather than P(takes an offensive snap), 18-28% apart on the bench? Refined before running: for a dressed man the served multiplier is the table (confirmed_active only ever stamps 0.0), fitted on on_roster=1=ACT; is the bench's served error in p or in the conditional line?

Prediction

Frozen (protocol 0017Z): H-151 REFUTED on its wording; |E1| < 0.03 in every bench cell; bench over-projection in the conditional term, largest for QB2; played-minus-offense-snap gap under 3 points in every bench cell.

Finding and verdict

H-151 REFINED. As worded it is REFUTED: for a dressed man production serves P(played | dressed, cell), not P(dressed), and on the bench it is served too LOW (-0.09 to -0.13), not too high. The claim the data carries: the table's population changes meaning at 2019 (no INA before), so it is fitted on a mixture of P(dressed x played) and P(played | dressed), and the bench under-call is that mixture. The fix is not free: refitting on 2019+ improves Brier and worsens fantasy MAE +0.011, because the bench conditional line over-projects on offensive-snap rows (+0.3 to +3.9 pts/row) and the low p was offsetting it -- ~2.5 pts per team-game each way. A vacant starting QB job (79 rows) is the largest single cell: -7.69 pts per row. Discovery/execution only; nothing enters CHANGES-PROPOSED.

Reasoning

The literal H-151 mechanism (P(dressed)) is absent on the confirmed path by design (inactives.py only stamps zero) and the table's target is conditional on ACT; the tier session's intuition was right for 2016-18 rows and for vacancy weeks only.

Next action

(a) the bench conditional line over-projection on offense-snap rows is the object that must move BEFORE the 2019 window can ship (window alone costs MAE); (b) vacancy: filed as a new row -- promote the backup when the ACT room holds no chart-1 man, then grade on the same 79+ rows; (c) the H8 invariant was red at run time (efficiency.py:182 scan order); this run uses one bundle per season on main's scan order, same as production, and no arm reorders player_week, so H8 does not separate the arms.