Thibaut Forest
Phase 56: NEW CHAMPION logbag_rank + gbm rank-blend w=0.30 (purge 1630.3, was 1972.0)
698401a
|
Raw History Blame Contribute Delete
4.59 kB

ft — new feature engineering (temporal-purge CV)

STATUS: DONE (last updated 2026-08-22 ~20:1x)

Baseline: champion logbag/base SCORE 1972.0, swaps 9.21. All rows paired per-scenario vs /tmp/cv_all48oof.csv. Reference points from siblings: logbag_rank 1746.4 (-225.6), frontier logbag_rank+gbm-0.30 blend 1630.3 (-341.7).

Results table (full-48 runs, sorted by delta vs champion 1972.0)

label ALL48 delta SE W/L swaps
ft_fr_ddown 1675.0 -297.0 101.6 34/13 9.29
ft_fr_whit 1748.9 -223.1 84.9 32/15 9.15
ft_rkw 1758.3 -213.7 68.8 34/13 8.92
ft_fr_rwhit 1782.3 -189.7 81.9 31/16 8.96
ft_rkwd 1794.2 -177.8 68.1 32/15 8.94
ft_rwhit 1922.7 -49.3 41.0 29/19 8.96
ft_rkw2 1931.5 -40.6 37.3 26/19 8.96
ft_rdip7 1961.9 -10.1 28.4 24/17 9.27
ft_dip7 1968.0 -4.0 33.2 20/23 9.31
ft_rddown 1977.8 +5.8 25.6 22/15 9.00
ft_whit 2015.1 +43.1 39.2 18/30 9.44
ft_wd 2035.1 +63.1 42.4 20/26 9.31

CRITICAL READING: the negative deltas in the top half are NOT my features — they are the r_vlast/r_tau rank pair and the gbm blend those configs sit on. The correct paired comparisons:

  • ft_rkw vs logbag_rank alone: +11.9 SE 33.1 (30W/18L) — r_whit adds nothing significant.
  • ft_fr_ddown vs frontier: +44.8 SE 25.7, 17W/28L — DEGRADES.
  • ft_fr_whit vs frontier: +118.6 SE 48.9, 16W/32L — DEGRADES.
  • ft_fr_rwhit vs frontier: +152.1 SE 42.3, 16W/31L — DEGRADES.
  • ft_rkwd vs ft_rkw: r_dip7 costs ~36.
  • ft_rkw2 (r_whit REPLACING r_tau): -40.6 vs champion — collapses from rkw's -213.7. r_tau is load-bearing; whit is no substitute.

What each feature was

  • whit: analytic Wiener 42d first-hit probability (staleness-conditioned inverse-Gaussian CDF from wnr_m/wnr_s2/v_corr_last, per-origin fleet-median shrinkage), fed as logit. r_whit = within-origin pct-rank.
  • dip7: v_last - v_min_7 (recent dip). newlow: v_min_30 - v_min_life. ddown: drawdown from running max of temp-corrected V (raw rebuild).
  • b250_60/b242_60: days below 2.50/2.42 V in last 60d (raw rebuild). readslope60: 60d slope of daily read count. macd: EWMA(10d)-EWMA(30d) momentum. divmm: min-vs-median slope divergence. All from a causal per-device rebuild of battery_metrics.parquet (/tmp/ft/rawfeat.parquet).
  • cohort_overdue/rel_age: install-batch calibrated clock (already in table, never in FEATURES).
  • ghost: (days_below_245>42 & v_corr_last<2.55) label-process ghost indicator; ft_gz = hard GHOST_ZERO.

Triage graveyard (17 origins; cut, not confirmed)

cohort +20/+43, b250_60 +66 (5W/11L), obsr +75, stsl +52, b242_60 +7, readsl +2, fiterr +22, ghost +44, ghostv -22 (noise), GHOST_ZERO +533 (swaps 8.5->5.6, count rule gutted — phase-15 non-adoption emphatically confirmed under the purge; historical: never adopted).

Structural findings worth keeping

  1. Slot squatters vs wall misses (purged fits, 9 origins): the alive devices occupying top-12 ranks are largely label-process ghosts (median days_below_245=43, b250_60=24.5); the missed deaths at ranks 13-30 sit HIGHER in voltage (v_corr 2.479 vs 2.441, tau_projected 23 vs 9). BUT ghost-flagged devices still die at 19% (vs 29% non-ghost at same depth), so both demoting them (feature) and zeroing them (GHOST_ZERO) lose more late-swap than they free in slots. The squatter problem is real but not solvable by a ghost flag.
  2. Rank transforms are the transfer device on this instrument: raw whit +43 -> r_whit -49 (same information, rank form). Confirms the sibling logbag_rank finding from the feature side.
  3. Offline marginal-signal screens (AUC beyond v_last) DO NOT predict cost: b250_60 screened best (0.643) and hurt worst; the "extra signal" pointed at ghosts.
  4. 17-origin even-panel triage can cut losers but cannot confirm winners: ft_whit read -71 on the subset and +43 on the full 48 (the full panel block alone flipped to +57).

Conclusion / recommendation

NO FEATURE SURVIVOR. Recommend the frontier config UNCHANGED (logbag_rank + gbm-0.30, 1630.3) — every tested addition (whit raw/ranked, dip7, ddown, and by triage all others) is neutral-to-harmful on top of it. The champion's BASE_FEATURES remain fully distilled; the r_vlast/r_tau rank pair subsumes what my candidates offered.

Scripts: /tmp/ft/run.py (all variants), /tmp/ft/rawbuild.py (+ /tmp/ft/rawfeat.parquet). CSVs: /tmp/cv_oofcand_ft_{whit,dip7,wd,rwhit,rdip7,rddown,rkw,rkwd,rkw2,fr_whit,fr_rwhit,fr_ddown}.csv; triage /tmp/ft/tri_ft_*.csv.