Audit on 6-node training data (151,329 frames) found 21 multicollinear
pairs (|r|>0.85), one dead feature (amp_min constant 0), and only node[0]
used in 8 of 15 features. Top per-feature F-stat = 15,497 but accuracy
stuck at 44.4% — classifier couldn't extract the signal that physical
sensors were already capturing.
Refactor:
- Drop 8 dead/redundant features (amp_min, amp_range, breath_bp,
spec_pow, motion_bp, amp_mean, amp_max, amp_iqr, amp_kurt).
- Keep 4 globals: variance, mean_rssi, dom_hz, change_pts.
- Add per-node features × all 6 nodes: amp_std, amp_skew, amp_entropy.
- New N_FEATURES = 22 (was 15). Z-score normalisation kept.
API change: features_from_runtime now takes &[(u8, &[f64])] — caller
must supply per-node amplitudes. New helper current_per_node_amps()
reads AMP_HIST.nbvi_history.back() for all live nodes.
Old data/adaptive_model.json removed (incompatible 15-feature schema).
Retrain result on same 151k frames:
44.4% → 49.58% accuracy (+5.2 pts)
Total improvement vs 2-node baseline (40.4%): +9.2 pts.
Live confidence distribution now meaningful (0.30-0.85) vs pre-fix
near-uniform 0.04-0.10. Sensor placement matters: n6 (near door, far
from AP) sep_ratio=0.60 best; n1/n5 (near AP) ~0.01-0.06 nearly dead.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
e86f650681