Review of 2026-09-23T1520Z-h218-existence-within-position

Owner lane: Not recorded

Reproduction

REFUSED. protocol_sha256 is None and input_artifacts is []. The as-run snapshot (h218.duckdb, 2026-09-23 09:16 MT) is gone. The fresh snapshot ref-h218 is a different byte set (played flags revised on ~30 rows). Bound by hash and unchanged: runner sha256 be973a9b (full hash in bound), result out.json sha256 8ee14b1e..., source record sha256 59d32246....

Finding and verdict

UPHELD

Review scope

Narrow. Upheld: (1) rapm_theta and uplift_cate are 1.000 vs 1.000 played vs not-played inside each of RB/WR/TE and have zero QB rows, so the pooled 0.90 vs 0.46 is position mix; (2) regime_gate and kalman_share presence equals the SQL rule on 23935/23935 keys AND equals the output of the real builders re-run on the snapshot; (3) defshare's defender-keyed joins carry no played/not-played gap. NOT upheld as more than discovery: the frozen control (kalman within-position gap > 0.5 in RB, WR and TE) fails on RB again (0.409), so the record's own terms keep the confirmatory reading void; this review does not restate the bar. The record says TESTED, which is the right label.

Next action

(1) surfaces.py surface_existence docstring says the not-played pool is 57% QB; the measured share is 54% (1394/2592 as run, 1392/2567 now). (2) kalman_share row existence at week W depends on weeks after W; if any consumer reads existence as a served-time feature (share.feature_sql COALESCE(ks.kal_c, rg.gated_c, p.c)), the 246-row/223-played flip at a week-8 cut says a live week-8 board would fall through to regime_gate or the raw column for those keys where the backfilled one had a kalman value. Whether the served path rebuilds kalman_share per week is not checked here; it is the natural probe.