bgkf-CV improvement campaign — live agent reports (2026-08-22 night)
Second full campaign, on the settled CV: building-GroupKFold, deployment-faithful (cv.py v5, folds summed per scenario).
THE SCORE: venv/bin/python -m batteryswap_example.cv bgkf
Champion (plain logbag) baseline: 1760.2
(warmup 2616.1 / panel 1773.4 / nearend 544.0; swaps 9.5/scenario;
all-building out-of-fold totals, mean over 48 scenarios).
Baseline cells: /tmp/cv_bgkf.csv (fold x origin). Candidates pair on the 48 per-scenario all-building sums. Survivor bar: delta < 0 beyond ~2*SE with swaps ~flat per scenario. Quarantine (public-proven, phase 56-58): count-coupled and label-process wins are report-only; split-file leak forbidden. This instrument measures CROSS-FLEET transfer (unseen buildings) - the deployment path both public and private splits take.
| file | theme |
|---|---|
| ft_findings.md | feature engineering (relative/rank forms favored here) |
| md_findings.md | model families & hyperparameters |
| pl_findings.md | planner: timing, routing (count-coupled = report-only) |
| bl_findings.md | blends, ensembles, calibration |
| tr_findings.md | training composition: weights, labels, sampling |
| rk_findings.md | learning-to-rank objectives |
| wd_findings.md | RADICAL: drop p42 entirely + registry sweep + wild cards |
Runs are ~40-120s each - no triage subsets needed, full runs always.