feedback_role_anchor_flag_gap

Rebecca Allen (NYL) 2026-08-01: snapshot said mispriced_under on points 6.5 (z=+0.58, L10 avg 3.7). The user pushed back — "she's been in the starting lineup" — and was right. Splitting by role:

The 0,0,0,0 games dragging the average were 3-7 minute bench appearances.

The guard did not fire. recent_vs_baseline_flag = stable and minutes_source_confidence = stable while minutes went 6.8 -> 26.5. The methodology-lockdown role-anchor check (project_methodology_lockdown) catches minutes DIVERGENCE between sources, not a role CHANGE over time. A promotion is invisible to it.

It compounds. adjustment_notes applied matchup: H2H L5 avg 1.6 -> -1.1 shift — that H2H sample is also bench-era, so the same artifact is charged twice: once in the baseline, once as a matchup penalty.

How to apply — before trusting ANY line_edges quality:

1. Pull recent_games and check role (or the new SDS boxscore participants[].starter, see project_sds_participation_receipts) for the L10 window. If role flips inside it, the pooled average is meaningless.
2. Re-average within the CURRENT role only, and cross-check against min_conditional at the bucket matching tonight's projected minutes — that block is role-agnostic and was right here (>=20 min: avg 10.33).
3. Ask WHY the role changed. Allen's 4 starts all had Sabally AND Fiebich out — both out again that night, so the sample was the matched comparable, not something to regress away from. A promotion caused by injuries that have since healed is the opposite case.
4. Don't quote the raw EV off a 4-game split. The arithmetic said +8.5%; the honest read is "the under case is disproven, roughly break-even to modestly positive." See feedback_model_edge_calibration.

Reinforces feedback_verify_before_extend: I led with the system's verdict instead of checking the game log first. The user's domain knowledge beat the model, and the model was wrong for a knowable, checkable reason.

FIXED 2026-08-02 (data-hydrator a8e5088 + 845499d, live-verified): detectRoleTrend compares L3 minutes avg vs prior-window avg (ratio ≥1.6 AND ≥7 min absolute AND all 3 recent games direction-consistent vs the prior avg — the consistency rule exists because the mean-only version misfired on Aliyah Boston within an hour of deploy: a 5-min anomaly read as "collapse" the night after she played 30). On a break the window re-anchors to ±8 min of the L3 (n≥3) and stamps flags.role_trend + role_trend_adjusted; un-anchorable breaks ship the flag alone. Allen's exact case is a regression test: points 6.5 flips mispriced_under → no_edge. Steps 1-2 of "how to apply" are now automatic; steps 3-4 (WHY did the role change; don't quote raw EV off n=4) remain judgment.

THIRD ITERATION (2026-08-05, live-verified): the lineup role-SPLIT — categorical, from per-game started flags now on BOTH SDS gamelog routes (153682f + 0ea22fa: the combined /stats/player/{id} route builds recent_games in its OWN query; patching only /games left DH's WNBA path blind — caught live). DH detectRoleSplit (17728e4 + ac083fa) compares the newest game's role against the ADJACENT preceding era (consecutive runs), NOT a window majority — majority-vote v1 missed Sabally live because her June starter era outvoted her July benching. Day-one changes flag without re-anchoring (run<3); established changes re-anchor to the current run. Categorical outranks the minutes-trend detector. Verified on the rebuilt 8/04 snapshot: Sabally=lineup_promotion, Harrison=lineup_demotion, Morrow=lineup_demotion (fresh activation — technically a "return", reads as demotion-era; acceptable). Rolewatch surfaces lineup_* trends sorted with the up-trends (grading 1b3db82). Remaining gap: the scout pkl still UNDER-biased in prod — its fix sits unpushed on feat/xgboost-plus-v5-missingness in sports-scout-service.

BACKTESTED (2026-08-05, 2,345 settled priced props / 20 nights):

• Flag = validated. Fresh promotions (run ≤ 2) → overs 19-12, +5.8u (61%). The market lag is SHORT: established promotions (3+ starts) → overs 42% — books reprice within ~3 games. Bet the promotion in its first 1-2 games or not at all. Demotion fades ≈ nothing (51/49).
• Re-anchor as direction-flipper = refuted. Role-window vs pooled DISAGREEMENTS: new direction −17.5u vs old +3.8u (n=33, ~z2.6). Kills near-neutral (18-11). Correction shipped (DH 6dc027f): windows_disagree quality when the two windows point opposite ways — no mispriced_* matcher fires on it, so disagreement = kill, not conviction.
• Caveats: line universe = bet_clv rows (1000/day cap hit every day → truncated), single-game outcome grading, in-sample. Forward test = rolewatch lineup_promotion candidates from 8/05 on.