How many graded men (played=1) were served an all-zero conditional line on the LOCKED boards, what did they score, and is the zero produced by a zero trailing share (the headline's mechanism) or by a later stage?

Owner lane: s-observation

Question

How many graded men (played=1) were served an all-zero conditional line on the LOCKED boards, what did they score, and is the zero produced by a zero trailing share (the headline's mechanism) or by a later stage?

Prediction

Prereg 2026-09-23T1916Z (research/scientist/prereg/2026-09-23T1916Z-h114-all-zero-lines.md), written after seeing only the wk2 count of 21: E1 wk2 21 men, <=40 half-PPR pts; wk1 15-35 men. E2 majority class A (no trailing usage), class C (non-zero trailing input) <=30%. E3 nonzero-line comparables outscore the zero-line men. Falsifier: C majority => headline misattributes the mechanism.

Finding and verdict

PARTIAL. Headline count REPRODUCED (21 on wk2; 25 on wk1). Cost measured: 1.25% / 1.67% of graded half-PPR points, 40% of it (Wentz + White) in two men. The mechanism clause ('a zero share') is NOT established: the registered falsifier fired on inputs, but the class map could not tell a zero model share from a later-stage zero.

Reasoning

The prereg classified by trailing inputs because that is what the stored row supports; the mechanism lives inside lines_for.

Next action

Replay playermodel.lines_for at model_sha of the served wk2 board on the 21 (+25 wk1) men, logging the share model output and each zeroing stage (predict_many clip, MIN_RAW_FOR_SCALE, confirmed_active, room reconciliation). Blocker is WORK. Wentz alone (QB new to team, depth 2, started) is a class of its own: a QB with no trailing rows gets 0 conditional on playing.