H-259 asked whether the structured roster feed carries a practice-squad elevation no later than text (predicted >= 80%, Q adds little). Refined: the race that decides what the board can know is between OUR TWO CAPTURES' first-receipt clocks, not between the sources; and the parser's precision against an independent label had never been measured.

Owner lane: q-language

Question

H-259 asked whether the structured roster feed carries a practice-squad elevation no later than text (predicted >= 80%, Q adds little). Refined: the race that decides what the board can know is between OUR TWO CAPTURES' first-receipt clocks, not between the sources; and the parser's precision against an independent label had never been measured.

Prediction

P1 S names < 20% of labelled Sunday man-games pre-kickoff in wk1-2 (no S capture between the Saturday deadline and Sunday kickoffs), so H-259's >= 80% is refuted as captured. P2 T names 40-70% pre-kickoff, precision >= 0.6. P3 where both name one, T leads by >= 12 h median. P4 the served board zeroes >= 70% of labelled men, so on the served object both sources are worth zero until a post-Saturday rebuild.

Finding and verdict

H-259's premise REFINED AND ITS PREDICTION REFUTED AS CAPTURED. The structured feed was NOT no-later-than-text: S first or tied on 3% (wk1) and 2% (wk2) of elevations either named, against the >= 80% predicted; the cause is capture schedule, not source latency (P1). Text named 79-81% of labelled elevations pre-kickoff, but the cue-near-name parser is 14-19% precise (8x a shuffled floor), 38-43% at best in a post-hoc family, so it is CAPTURE_ONLY_FACTOR_NOT_ADMITTED: its errors are list-neighbour names, not language the model cannot read. On the served object the whole question is moot today: every labelled offensive man on the wk1-2 board was served 0.0, the wk2 board was built Tuesday 09-15, and wk2 lost 19.2 of 19.2 elevated points that text held before kickoff (15.4 of them Cody White).

Reasoning

Weakest joint, for the referee: (1) the label is incomplete (weekly file hides repeat elevations; LABEL2 repairs it only from S), so precision is a lower bound and recall an upper one; (2) two weeks, one season, no held-out; (3) the parser family is post hoc; (4) S's wk1-2 result is an outage, and wk3's 8/19 with S running is the fairer read of the rival, still censored.

Next action

WAITING not before 2026-09-30: re-run both runners unchanged on a fresh snapshot after the 09-27 capture and nflverse's final wk3 weekly file. Wk3 is the confirmation set for (a) S pre-kickoff recall with the daily capture running (the fair test of the H-259 rival) and (b) V4 precision/recall. A precision-worthy parser needs clause-level attribution (subject of 'elevated'), not a character window; that is a Q build item, not a finding.