Thibaut Forest Claude Opus 5 (1M context) commited on
Commit
80003d7
·
1 Parent(s): 4be2283

Phase 57: building-GroupKFold champion baseline SCORE 352.0 (per-test-fleet denomination, 22s harness)

Browse files
Files changed (1) hide show
  1. EXPERIMENTS.md +8 -0
EXPERIMENTS.md CHANGED
@@ -3129,3 +3129,11 @@ label it trains on beyond what deployment itself would know.
3129
  CLI: `cv bgkf` / `cv cand <label>` / `cv report`. Purge harness recoverable
3130
  at f5f3fda; walk-forward family at 2000281. Candidate rule strengthened by
3131
  phase 56: wins must hold at ~flat swaps PER CELL, not just at the mean.
 
 
 
 
 
 
 
 
 
3129
  CLI: `cv bgkf` / `cv cand <label>` / `cv report`. Purge harness recoverable
3130
  at f5f3fda; walk-forward family at 2000281. Candidate rule strengthened by
3131
  phase 56: wins must hold at ~flat swaps PER CELL, not just at the mean.
3132
+
3133
+ Phase 57 baseline (champion = deployed plain logbag): **SCORE 352.0**
3134
+ (folds 489.0/322.6/365.9/289.1/293.7; blocks wu 523.2 / panel 354.7 /
3135
+ nearend 108.8; swaps 1.9/scenario on ~90-device test fleets - the count rule
3136
+ scales with fleet size as designed). NEW DENOMINATION: per-test-fleet cost,
3137
+ never cross-compare with the full-fleet instruments (1972.0 purge / 1661.7
3138
+ self-fit). Harness runtime 22s - cheap enough to run per candidate without
3139
+ triage subsets.