Which producer's interval reaches a K/DEF row, and does each producer's band cover at its nominal 10/90?

Owner lane: Not recorded

Question

Which producer's interval reaches a K/DEF row, and does each producer's band cover at its nominal 10/90?

Prediction

A (0.45/1.75) misses: K floor exceedance 15-25%, DEF floor >=25%, DEF ceiling >=15%; A's floor off nominal by more than 2.5pp. B (KDEF_RESIDUAL_Q) reproduces ~10/10 on 2025 and is inside 5-20% on 2026. C within 2pp of nominal on 2021-25. D-C spans zero for K, small (<0.05 pt pinball) for DEF. C-A pinball excludes zero in C's favour on the DEF lower tail.

Finding and verdict

PREMISE CONFIRMED AND MEASURED. The 0.45/1.75 pair is live on every K/DEF row of the warehouse forecast table and reached the served K row in week 2. Held out 2021-2025, it breaks its floor 16.6% (K) and 15.8% (DEF) of team-weeks against a nominal 10%, and its ceiling 12.1% / 10.7%. A walk-forward residual pair (C) beats it on pinball: K q10 -0.0405 [-0.0631,-0.0186], K q90 -0.0151 [-0.0240,-0.0063], DEF q10 -0.0347 [-0.0608,-0.0088] (Bonferroni-8, all exclude zero); DEF q90 -0.0040 [-0.0394,+0.0288] spans zero. The shipped KDEF_RESIDUAL_Q pair (B) is indistinguishable from C on 2021-25 (|d|<=0.001 pb10) and reproduces 10.1/10.1 on 2025. Level-dependent bands (D) MEASURED NOTHING over the pooled pair (4/4 nil). Prediction: A-K floor, A-off-by->2.5pp, B reproduction, C-A DEF lower, D-C all HIT; DEF floor >=25% and DEF ceiling >=15% MISSED (15.8%, 10.7%) -- the constants are bad but less bad for DEF than predicted. C within 2pp: K ceiling 11.7% HIT at the edge.

Reasoning

The defect is construction, not modelling: one fitted pair (B) exists and is correct to within the measurement's resolution, but two producers write floor_10/ceiling_90 for K/DEF and build-serving keeps whichever arrived first. The fix is a single source for the K/DEF band (kdefmodel.interval) in forecast.py:734, not a new band.

Next action

(1) mechanical: forecast.py:734 should call kdefmodel.interval (the ONE definition build-serving's docstring already claims) so the forecast table, the archive and the served row agree -- routed via ledger, not CHANGES-PROPOSED (discovery). (2) measure the band around the SERVED projection (game_forecasts anchor) once 2026 has >=8 week-clusters. (3) K ceiling drift: a recency-weighted residual window vs all-prior, family of 2.