Plan: Q4 closes, the scaled candidate already handles category-driven spread (I074)

[bbstats]

The quantile model already handles categories that set the spread, so the planned fix (Q4) isn't needed. Plan files only.

Q4 was queued on 2026-09-23 because the model lost badly to CatBoost on catscale, a synthetic dataset where each category has its own noise level. Since then the default gained the "scaled" candidate (#177). Its spread model encodes each category by its average error, which is exactly the signal Q4 was going to add.

Rechecked on the same synthetic data (excess CRPS x1000 over the true distribution, lower is better):

catscale	default now	model on 2026-09-23	CatBoost MultiQuantile	NGBoost
1,000 rows	79.9	91.0	108.2	83.7
10,000 rows	16.0	42.4	30.9	32.8

The "scaled" candidate won all 6 fits. The fixed-width baseline reproduced its stored score exactly, so the data match the old run.

Records: CAMPAIGN_PLAN.md I074; QUANTILE_PLAN.md Q4 closed, Q1 re-read.

🤖 Generated with Claude Code