What row fact makes H8 red, does it bind the 2026 board's own fit window, is it the WHOLE of the order dependence, and which order-free position rule agrees with the row the consumer is served?

Owner lane: Not recorded

Question

What row fact makes H8 red, does it bind the 2026 board's own fit window, is it the WHOLE of the order dependence, and which order-free position rule agrees with the row the consumer is served?

Prediction

PREDICTION (written before the script ran): A: the multi-position ids are real nflverse per-week roster moves (one clean season-boundary switch dominates: RB<->WR, QB<->TE, FB<->RB), 0 source duplicates with two positions (spine already dedups), 0 ids carrying two names. Count grows from 24 (<=2022, H-274) to ~35-50 at <=2025. B: H8 on the shipped window (<=2025) is also red, thousands of rates, both seeds. C: canonicalising ONLY the multi-position players' position makes both reorders change 0 rates on both windows (the multi-position ids are the whole defect). D: for multi-position players with a 2026 row, the 2026 position equals latest-through-2025 on >=90% and modal on fewer. Falsifier: C non-zero on any window/seed => another order-dependence exists and H-274's fix is incomplete. [written 2026-09-26T2317Z before the run; local clock only, not externally receipted]

Finding and verdict

REFINED, CONFIRMED AND EXTENDED. (1) H8 binds the 2026 board's own window: fit season<=2025 moves 12,225 and 13,721 of 13,721 rates under two reorders (max 2.72), not only the <=2022 invariant window. (2) It is exactly the multi-position ids: relabel those 32 players (1.1% of played-row targets and carries) and both reorders move 0 rates on both windows; nothing else in efficiency.fit depends on order. (3) The key: player_week.position is nflverse's per-WEEK roster label, not a player attribute. 19 of the 32 flip INSIDE a season and 12 flip two or more times (Taysom Hill QB/TE 4 flips, Andrew Beck TE/RB 3 in-season), so a 'position switch' is not a season-boundary event; 0 source duplicates; 2 ids carry two spellings of one name (Velus Jones/Jr., Joe/Joseph Fortson), not collisions. (4) The amplifier: 1.1% of volume moves 100% of rates because the player's career totals move between posr pools and every player at both positions is shrunk toward that pool. (5) The rule H-274 left as a football choice has a construction answer: of 11 multi-position players with a 2026 row, the 2026 position equals latest-through-2025 on 10 and modal on 4; modal disagrees with latest on 12 of 32. The canonicalisation is itself a model change (canon-latest vs shipped: 13,721/13,721 rates, max 2.72), so whoever ships H-274 must grade it as a model change, not a no-op.

Reasoning

Sol's stop recorded that H8 fails on the invariant's window; it did not name the row fact, check the window the served board is fitted on, show the multi-position ids are the whole defect, or give the order-free rule a basis.

Next action

Ship H-274 with arg_max(position, season*100+week) (latest), and grade on the 32 ids, not the league. Invariant H8 could also run season<=2025 (red there too, 100% on seed h2).