share.feature_sql serves depth_chart_rank as LEAST(COALESCE(rank,0),3) (share.py:1868). Who reads the 0, and does each reader interpret it as it was fitted?

Owner lane: serving

Question

share.feature_sql serves depth_chart_rank as LEAST(COALESCE(rank,0),3) (share.py:1868). Who reads the 0, and does each reader interpret it as it was fitted?

Prediction

Frozen in the runner's docstring at 2026-09-27T0217Z before its first execution (local clock, no external receipt): E1 negative between -0.05 and -0.40 with the 98.33% interval excluding zero; E2 |delta|<0.02, may span zero; E3 spans zero or slightly positive. Invariant: rows whose rank != 0 outside a QB room holding a rank-0 QB move by exactly 0.0.

Finding and verdict

TESTED. The COALESCE end is a real fit/serve encoding mismatch -- playprob/qbshare are fitted with NULL=3 and served 0, which they read as 'position average, snaps ignored' (P(play) 0.885 vs 0.716 on 6% of 2023-25 roster-weeks). But correcting it in either direction measured WORSE on mean-sensitive accuracy: the frozen remap wins MAE on the rank-0 man (-0.046, excluding zero) only by median shrinkage (MSE +0.81 on played rows, Brier nil) and gives it back to his teammates through _reconcile (pooled MAE nil); fitting a depth-0 cell is worse still (pooled MSE +0.146, excluding zero), because unranked meant 'not playing' in 2016-18 (0.27-0.55 play) and 'omitted depth man' from 2019 (0.78-0.91). The accidental encoding is the best of the three measured, by accident. Do not ship the remap. The live board carries the 0 on 1 of 450 rows.

Reasoning

Refines H-118 from 'the COALESCE end is unnamed' to: it is read differently by four consumers; the one that matters (P(play)) is miscalibrated for every encoding tried because the NULL population's meaning is source-era-bound; the redistribution in _reconcile makes any per-man P(play) change a team-week change, so a per-man endpoint alone overstates any gain.

Next action

Pre-register (before running) an era-bound unranked cell: playprob/qbshare depth-0 cell fitted on season >= 2019 only, scored held-out 2024-25 with fit through S-1, endpoints pooled unconditional MSE and team-week MAE (not the rank-0 man alone), P(play) Brier on rank-0 rows. Predict: rank-0 P(play) within 0.03 of realised, pooled MSE nil -- the live reach is ~1% of served rows, so the ceiling on value is small; this is worth one run, not a line of work.