Brain cycle 82: fix calibration mutation bias — venn_abers boosted across all 3 GA impls
Claude
d1af5f1
PROBLEM: venn_abers is the confirmed best calibration (S13 pareto_best 0.21773,
delta -0.00543), but mutation weights were PENALIZING it:
- hf-space/app.py: init=22 → mutation=14 (dropped by 36%)
- genetic_loop_v3.py: venn_abers MISSING ENTIRELY (only isotonic/sigmoid/none)
FIXES:
1. hf-space/app.py mutation: venn_abers 14→25, none 22→16
- Now mutation preserves venn_abers instead of reverting to uncalibrated
2. genetic_loop_v3.py: add venn_abers to Kaggle GPU loop
- Init + mutation both now use weighted choices matching HF island config
- GPU Karpathy sessions will now explore venn_abers for first time
Expected: GPU Kaggle loop discovers venn_abers advantage within first session.
HF islands maintain calibration gains instead of mutating them away.
https://claude.ai/code/session_01KtBfaHJZ3YzfFqzerNp23f
d1: DART dropout trees on LightGBM in GA loop — expected Brier delta -0.0025
P012: Add boosting_type=dart, drop_rate=0.1, skip_drop=0.5, uniform_drop=True
to LGBMClassifier in _build_model (genetic_loop_v3.py). DART prevents first-tree
dominance, improves probability spread across [0.3,0.7], reduces reliability
component of Brier. Consistent with Rashmi & Gilad-Bachrach 2015 (AISTATS).
Already verified working in app.py stacking path — now applied to main GA loop.
1. Brier-based elitism: protect top-2 by raw Brier in addition to composite
(prevents fossil loss like S15's 0.22159 config disappearing)
2. CatBoost CPU cap: n_estimators=60 + early_stopping=15 on CPU
(3-5x speedup on S10/S13 where catboost dominates)
3. Feature penalty: -0.05 per 200 features above 80 threshold
(breaks the n_feat=200 trap that causes convergence)
4. Default n_splits: 5→3 for 1.6x faster walk-forward CV
Expected: fleet 219→350 gen/hr, prevents config loss, better Brier pressure
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
