EXPERIMENTS.md

Experiment history

Chronological record of everything tried, with the CV protocol and scores. All numbers are mean total_cost per scenario from the official evaluate_plan (lower is better).

CV protocol

• Unit: (device, origin) pairs; an origin is a planning time-zero. Features use only data ≤ origin; the label is death within the 42-day window after it.
• Walk-forward: at every scored origin the model is refit on all pairs whose window closed before that origin (fully observed labels, built-in embargo).
• Current protocol (since the 2026-08-18 protocol change, after phase 10): champion selection on the merged 34-origin panel s_8–s_41 with repeated-split stability - accept a challenger only if selected in ≥60% of 2000 random half-splits AND modal under blocked splits; the public leaderboard is the only remaining holdout and is spent deliberately.
• Three-regime scorecard (since phase 20): the panel alone samples one regime (native fleet, mature calibration, expensive false swaps). Standing additional blocks: warm-up s_0–s_7 (cold-start calibration; where the phase-19 floor was proven) and near-end s_42–s_47 (censored shipped truth - the same scoring the remote evaluator applies to a held-out split's near-end scenarios; where the phase-11/14 mechanisms are measured). Adoption of a deployment-regime mechanism requires: no panel regression (tie-aware selection stability) AND a win in its target regime block AND the regime being reachable by a local harness - "architecturally sound" is not sufficient (see the foreign-fleet floor, built and refuted within one day in phase 20). selection_stability counts decided splits only; summarize() reports late/early/ops components by default.
• Other blocks: warm-up s_0–s_7 (3–16 unique training deaths) is excluded from champion selection but scored since phase 17 for cold-start behavior; s_42–s_47 stay unscored (windows extend past data end 2026-07-31).
• Historical blocks (phases 1–10): dev s_8–s_23 (all tuning) / test s_24–s_41 (counted looks only, ledger at the end); merged once 8 looks had spent the test block. Per-phase "dev"/"test" columns below refer to these.
• Reference points (test block): shipped example 6779 · defer-everything 2512 · lowest-voltage-15 heuristic 2034 · oracle (true EOL) 186.
• Protocols live in batteryswap_example/cv.py (phase 40), not in /tmp: all48 (full-fleet walk-forward, three-regime scorecard), static (the deployment gate - one foreign-fleet fit, no refits, ORDERING questions only), screen (offline model screen grouped by BATTERY, never by scenario or row). All three report ranking precision at the operating N beside the cost: cost alone cannot distinguish a change that reordered the list from one that only cut it deeper, and AUC 0.947 has already coexisted with recall 0.000.

Phase 0 - understanding the problem

finding	consequence
Plans past day 42 are dropped free (evaluate.py)	the decision is swap in-window vs park; ~9 of ~450 batteries/scenario matter
Penalty asymmetry 0.5/day early vs 10/day late	cost-optimal swap day ≈ 5th percentile of EOL distribution
Example model predicts constant 120d for every battery	its 43 in-window swaps are arbitrary; 8673 all-48
Defer-everything scores 3325 all-48 (−62%)	the real bar for any model
Last in-band voltage: AUC 0.95, 100% coverage	age (install+2y) is useless: AUC 0.63, P@5=0.00 (survivorship: median alive age > median lifetime)

Phase 1 - first model lineage (v1, submitted 2026-08-17)

step	dev	test	notes
voltage-k heuristic, k=15	—	2034	beats defer 14/18; k tuned on cost, not precision
72-config sweep: {rank-v, logistic, gbm, aft} × {onepd, sched, batch} × p* × day-q	best 2580 (aft/batch/p0.35)	—	batch > sched > onepd for every model
logistic/batch, p*=0.5	2473	1735 (look 1)	threshold minimum confirmed (0.35: 2678, 0.65: 2601, 0.8: 2989)
diagnosis: p42 scale drifts as training grows; at s_38–41 the 0.5 threshold includes 0 batteries while 5–7 die			fixed thresholds are structurally brittle
topsum: include top-N by p42, N = 0.8·Σp42	2502	1368 (look 2)	beats fixed on 15/18; v1, public LB 1964.29 (rank 3)

Rejected in phase 1: AFT expected-cost planner (3182 dev - uncalibrated probabilities), Platt calibration (+40–50 dev), GBM (behind logistic at every data size).

Phase 2 - features, expected cost, routing

step	dev	test	notes
recall diagnostic			deaths ranked fine (median rank 13) but excluded; worst misses had NaN slopes (sparse traces)
+5 features (slope_lastk, drop_recent, below_frac, accel, n_30)	2390	—	window-free slope closes sparse-trace gap
ecday: include/park + swap day by expected cost under predicted day-distribution	2327	—	low-p batteries automatically scheduled later
ecroute: + local search over building-group days vs exact mobility simulator	2197	1478 (look 3)	simulator reproduces evaluator to 0.000 after replicating its return-leg double-count quirk
tsroute hybrid (topsum inclusion + ecroute scheduling)	2229	1502 (look 4)	tied on dev
physics: master discharge curve			validated: EOL = 2.40 V crossing (±0.034 across all 82 deaths), temp coeff +4.9 mV/°C; tau_hat AUC 0.927 alone but redundant with voltage features → rejected (2274 dev)

Phase 3 - training-data augmentation

step	dev	test	notes
sudden-death check			only 2/79 deaths lack voltage droop → late_swap is recoverable
synthetic origins every 4 days back to Feb-2025	tsroute 2126, ecroute 2161	1476 (look 5)	positives at s_24: 253 → 688; recovers 3 deaths no scenario window covered
day-distribution blend toward uniform	2174–2214	—	rejected by dev

Phase 4 - error screen and targeted fixes

Error buckets on test (per origin): miss→emergency 718 · false-positive early 340 · timing of caught deaths 41 (solved).

step	dev	test	notes
tau_projected = C⁻¹(v_corr_last) − staleness	2147 (neutral)	1439 (look 6)	targets the 8 "dark" devices (staleness ~62d) = 25 of 60 misses; AUC 0.676 where model was blind
break-even top-up (include p42 ≥ 0.2)	2487–2742	—	rejected by dev
placement floor for uncertain inclusions	2263	—	rejected by dev

Recurring lesson: every absolute-probability knob motivated by test-block failures (blend, top-up, floor) was refuted on dev - the p42 scale shifts with death prevalence. Only scale-free mechanisms survive: count-based inclusion, ranks, physics projections.

Phase 5 - spending the 30-minute runtime budget (current, fc9c286)

step	dev	test	notes
within-day visit ordering fix (building/room sort) + merge-move search with restarts	2120	—	interleaved rows paid repeated building changes the search never saw
logbag: 15-fold bagged logistic	2074	1398 (look 7)	variance reduction near the inclusion cutoff
aggressive challenger α=1.15	—	1516 (look 7)	leader-like profile (late 722/early 550); kept as BATTERYSWAP_ALPHA knob, not default
adaptive runtime refit (submit.py)			refits at every scenario origin on the target split's own self-masked history; honest all-48 train metric 2202; 280 s/split native; deadline guards → static-pickle fallback

Public leaderboard (2026-08-18, adaptive submission): 1767.87, rank 2 - v1 was 1964.29. Mobility+limits 315.9 now beats the leader astar's 334.6; the whole remaining gap to astar (1652.6) is timing balance (our late/early 1296/156 vs their 660/658). An α=1.15 change motivated by that profile was committed, then reverted before submission: the local α sweep rejected it (see phase 6) - selection stays on CV, not public feedback.

Phase 6 - voltage-silent deaths (exploration, all rejected by CV)

Silent deaths are fast plungers, not a separate mechanism: all 16 deaths that looked healthy 30d out (v_corr > 2.55) were at ~2.47 by 7d out and ~2.41 the day before - the same 2.40 V crossing, compressed into ~3 weeks.

candidate	evidence in isolation	dev CV	verdict
room/building failure clustering	AUC 0.52-0.54 (chance)	not attempted	dead on arrival
lifetime mean temperature	AUC 0.64; corr(lifetime, temp) 0.21	not attempted	too weak alone
install-cohort clock (age - batch median lifetime)	AUC 0.906 on silent pairs; batch explains ~90% of lifetime variance	2074 -> 2279 (raw), 2433 (overdue-neutral encoding)	rejected: 14%/54% coverage - the no-information majority degrades the fit in every encoding
alpha re-sweep 0.7-1.3 (public-LB-motivated 1.15)	public late/early ratio 8 vs leaders' ~1	minimum stays at 0.8 (2074; 1.15: 2219, losing 12/16 origins)	1.15 reverted; late saturates ~1100 for alpha >= 1.15 - remaining misses are not an inclusion problem

Phase 7 - full coordinate sweep (nothing adopted; champion is a local optimum)

One-knob-at-a-time sweep around the champion, ~25 knob values across both axes. Acceptance rule fixed in advance: adopt only if dev improves by >= 60 AND wins

= 10/16 origins head-to-head. Full cartesian grids were ruled out: dev's resolution is ~150 points (16 correlated origins), so a large grid guarantees spurious winners.

axis	best variant	Δ vs champion 2073.9	verdict
bag count	30 bags	−9.4	closest of all; 6x short of the bar
search budget	restarts 4 / iters 120	−0.4	saturated
day-model GBR	dtrees=50	−0.6	insensitive
regularization	C=0.3	+64.8	C=1.0 confirmed (0.1: +148.6, 3.0: +76.7)
training set	dthaz (discrete-time hazard, open windows)	+15.4	statistically tied while using a different likelihood and more data - the natural committee partner if one is ever wanted
recency weighting	half-life 240d	+30.7	no drift-adaptation gain (120d: +99.7)
blends	logbag+gbm	+190.6	GBM drags; blend_dh +1367.9
routing-aware inclusion	tspig membership repair	+46.5	right shape, does not pay at these travel costs
piggyback inclusion	tspig2 α₂=1.3 / 1.6	+293.1 / +497.7	over-includes
determined labels	logbag_open	+3080 (std 11k)	blows up at early origins - tiny determined sets

Disclosures: emergency_day is structurally inert for tsroute (it prices parking, which the count rule never consults); the q6 day-quantile grid is inert because cost_curves reads only the four standard quantile columns (CDF-resolution axis untested); logbag_aux* variants await their feature columns in the table.

Conclusion: every axis is flat or worse around the champion - remaining cost is structural (fast-plunge deaths), not hyperparameters.

Classifier-family addendum (RandomForest and friends)

model	dev	Δ	note
RandomForest (500, leaf 5)	2273.1	+199	ranks well (AUC 0.94 at s_23 vs logbag 0.85) but probabilities too compressed for topsum's count
RF deep (leaf 1, 1000)	2325.9	+252
RF class-balanced	2890.2	+816	worst config ever tested - inflated Σp42 over-includes massively
ExtraTrees	2245.3	+171
MLP (32,16, early stop)	2768.7	+695
rank_rf hybrid (RF ranking, logbag scale - Σp42 preserved)	2166.2	+92	RF's ranking edge does not transfer across origins even with the scale problem removed

Complete p42-classifier ledger: logistic-family (champion), HGB/GBM x4, Weibull AFT, discrete hazard (tied), RF x3, ExtraTrees, MLP, and five blends - linear + bagging wins throughout. With ~1000 positives, one dominant monotone signal (voltage level), and a planner that needs a stable probability scale, that is the expected outcome.

Phase 8 - the silent-death campaign (closed: unactionable within this data)

Target: the ~27 test-block misses with p42 < 0.10 - fast plungers whose voltage collapse starts AFTER the planning origin.

Precursor screen (9 candidates, silent subpopulation = 11,599 healthy-voltage pairs, 24 deaths): only volatility carries signal - rng_30 AUC 0.833, vstd_30 0.810; diurnal-amplitude drift 0.60, temp-sensitivity drift (beta) 0.57, reading-rate change 0.58, low-side outliers and range-ratio ~ chance.

attempt	dev	test	verdict
rng_30 + vstd_30 as global features	+17	1438.7 vs 1397.5 (worse)	rejected - signal diluted across the healthy majority
SilentFix override (blind-zone logistic re-ranks silent batteries only; count untouched, scale-free)	w=0.15 inert, w=0.3 +6, w=0.6 +38	w=0.3: byte-identical plans (1397.5)	rejected - inert at safe weight, harmful at active weight
cohort clock (phases 6)	+205..+359	-	rejected twice
inclusion expansion (alpha, p_be, floors)	all rejected	late saturates ~1100 regardless	not an inclusion problem

Why this is a wall, quantified: blind-zone death prevalence is 24/11,599 = 0.2%. Even at AUC 0.83 the precursor's precision at any usable operating point is a few percent, so acting on it loses in expectation: adding a silent battery costs 50-90 (early) against ~0.02-0.05 x 215 (expected emergency saved); swapping one in displaces a ~30%-p42 candidate. Every mechanism - global feature, ordering override, count expansion - hits the same arithmetic. The remaining silent misses (290/origin on test) are the price of predicting a 42-day window from a signal that only turns on ~3 weeks before death. This ceiling moves only with new information: more observed deaths maturing the cohort clock (batch explains ~90% of lifetime variance - the physics is there, the coverage is not), or upstream data (duty cycles, battery lot IDs).

Trajectory

champion	dev	test	combined (34)	all-48 metric	public LB
shipped example	—	6779	—	8673	—
defer-everything	—	2512	—	3325	—
voltage-15	—	2034	—	—	—
logistic + fixed p0.5	2473	1735	2082	—	—
v1 topsum	2502	1368	1901	2150	1964.29 (rank 3)
+ features + ecroute	2197	1478	1816	—	—
+ synthetic origins (tsroute)	2126	1476	1782	—	—
+ tau_projected	2147	1439	1772	2081	—
current: + bagging + search + adaptive runtime	2074	1398	1716	2202*	1767.87 (rank 2)

* honest walk-forward (adaptive path), not comparable to earlier in-sample-ish pickle numbers.

Current champion in full: logbag (15-bag logistic, C=1.0, 13 features) + tsroute (topsum inclusion α=0.8, expected-cost day placement, merge-move day search against the exact mobility simulator) + adaptive per-origin runtime refits, confirmed a local optimum by the phase-7 sweep.

Oracle bound: 186 (test). Remaining gap is dominated by deaths with p42 < 0.10 - batteries whose voltage has not moved by the origin (27 of 60 test misses).

Phase 8 - restart-from-baseline sweep of unexplored directions (2026-08-18)

Systematic pass over mechanisms not tried in phases 0-5. Champion reproduced first (logbag/tsroute α=0.8: dev 2073.9). All tuning on dev only.

idea	dev	notes
EV-based piggyback: post-routing add/drop by expected value vs exact mobility delta (tspig)	2120	EV decisions inherit the p42-scale brittleness - again
rank-gated piggyback: extend inclusion to next-ranked batteries in already-visited buildings (tspig2, α2=1.0/1.3/1.6)	2162/2367/2572	monotonically worse: α=0.8 is the true inclusion optimum; even zero-travel extras lose
recency-weighted training (half-life 120/240 d)	2174/2105	rejected
discrete-time hazard on censored open windows (dthaz)	2089	close but no win; open windows add little given synthetic-origin augmentation
determined-label training (open-window deaths as positives, logbag_open)	5154	prevalence bias wrecks sum(p42) → runaway inclusion
logbag+gbm blend / logbag+dthaz blend	2265 / 2078	rejected
logistic C sweep 0.1/0.3/3.0	2223/2139/2151	C=1.0 was already optimal
in-origin fleet ranks of v_last, tau_projected (leak-free)	2055	helps bare logbag, redundant with aux features
v_proj = v_last + slope_lastk·staleness	2122	redundant with tau_projected
aux usage-dynamics features from raw hourly data (aux_features.py: daily discharge depth rng_90 AUC 0.78, vstd_30 0.77, rng_ratio, gap_max_60)	2035 (α 0.85: 2030)	only dev-positive find; paired stats soft (7/16 wins, t≈-0.7)
UCB inclusion (rank by bagged mean + k·std)	2038/2053	neutral
log1p-compressed heavy-tail features	2098	rejected
more bagging (30/50) / more search restarts (8)	2046-2058 / 2031	flat
cohort_* columns (batch-age clock)	—	skipped: AUC ≤ 0.55 standalone

Test look 8: logbag_aux/tsroute/α0.8 → 1510 (vs incumbent 1398); combined 1757 vs 1716. Dev gain does not transfer - the incumbent stays champion. Conclusion: the (device,origin) table is mined out around the current operating point; inclusion count, probability model class, training-set construction, and routing search are all at local optima that reinforce each other. The remaining gap to oracle is the no-signal-yet deaths, and neither usage stress, hazard pooling, nor uncertainty boosting reaches them on test.

Phase 9 - literature-driven approaches (2026-08-18, all rejected by CV)

Two approaches from the prognostics/decision-learning literature, implemented and walk-forward tested. In-run reference: logbag/tsroute α0.8 = 2091 (FEATURES had gained vstd_30/rng_30 from the phase-7 sweep by this point).

Wiener first-hitting-time RUL (fht_features.py + WienerFHT): temp- corrected voltage as Brownian motion with per-device drift (shrunk toward the population), EOL = first passage of the validated 2.40 V threshold, closed-form barrier-crossing CDF, dark devices conditioned on surviving their silence. Ranks well (AUC 0.91-0.96 spot-checked) but probability mass is overdispersed (sum p42 ≈ 35-54 vs 8-15 true deaths), and the planner runs on probability mass:

config	dev
wiener standalone, α 0.8 / 0.3	2868 / 2884
0.5·logbag + 0.5·wiener, α 0.8 / 0.6	2309 / 2329
wiener overrides logbag where staleness > 14 d	2503
FHT projections as logistic features (fht_days, wnr_m, wnr_s2 / fht_days)	2405 / 2222

Redundant with tau_projected where data is dense, noisy where it is sparse.

Conformal risk-controlled inclusion (ConformalCut + tscut): per-refit cutoff = risk_q quantile of the model's scores on dying calibration pairs from the last 120 d of closed windows - scale-free by construction, the literature's answer to exactly our probability-drift failure mode.

risk_q (allowed miss rate)	dev
0.15 in-sample / 0.15 split-calibration	3266 / 5907
0.25 / 0.40 / 0.50 in-sample	2480 / 2286 / 2296

Even at the cost-matched operating point (~0.4-0.5) it loses ~200 to topsum: the cutoff is a quantile of ~10-30 recent death scores and inherits their sampling noise, while topsum's count integrates over the whole fleet's p42 mass. Verdict: count-based inclusion beats threshold-based inclusion even when the threshold is conformally recalibrated at every refit.

Not attempted from the same review, with reasons: decision-focused learning / SPO (16 dev origins cannot support cost-gradient tuning without overfitting the block), deep sequence RUL (82 deaths), RL scheduling (sample complexity), MNAR state-space voltage models (gap features were already dev-neutral; highest-effort/lowest-evidence of the batch).

Phase 10 - late_swap gap: cohort-gated rescue (rejected; bucket is empty)

Motivation: public-LB late/early = 1296/156 vs leader 660/658; misses→emergency is ~718/origin of test cost. Hypothesis: the phase-6 cohort clock (AUC 0.906 on silent pairs) was unfairly rejected because dev-era coverage was 35%, vs 72% test / 84% deployment-era - so deploy it gated to covered devices only and validate on the late-dev block s_18-23 where coverage is representative.

Quantifying first where the cohort signal lives (covered pairs, s_18-41):

subgroup	n	deaths	cohort_overdue_m AUC
all covered	6134	131	0.584
dark >14d / >30d (the actual miss bucket)	1282/1144	26/23	0.426 / 0.410 - worse than chance
looks healthy (v_corr > 2.55)	4734	4	0.902

The 0.906 was real but rests on 4 positives: fast plungers that were still healthy at origin. Even perfect recall there is 4 catches per 24 origins, and reaching them means including high-overdue healthy devices at hopeless precision (60:1) x ~100 early-cost each.

CV confirmation (CohortGate: logbag + per-device lift to a cohort-hazard logistic on covered devices; 'silent' mode lifts only p42 < 0.2), late-dev s_18-23: logbag 1943 · cg_all 2421 · cg_silent 2421 (identical plans - all lifts land in the p<0.2 bucket) · cg_silent α0.9 2588. Rejected.

Conclusion on the late_swap gap: the remaining misses' plunge onset is after the planning origin. No origin-time feature in this dataset sees them: voltage is healthy, usage stress is flat (phase 8), FHT drift is flat (phase 9), and the batch clock is uninformative exactly where the misses live. The late/early imbalance vs the leader is not recoverable by rebalancing (phase 6 α-sweep) or by any tested signal; parity would require either mid-window replanning (not allowed by the submission format) or a signal source we do not have.

Protocol change (2026-08-18): dev and test merged into one 34-origin block

After 8 counted looks the test block no longer provides an unbiased holdout, and dev s_8-23 is structurally unrepresentative of deployment (smaller refits, 35% cohort coverage vs 84% at deployment-era origins). Before merging, every near-miss candidate without a test number was scored on s_24-41 (look 9):

model	dev16	test18	combined34
logbag (incumbent)	2073.9	1398.0	1716.1
logbag_rank	2055.3	1446.0	1732.7
logbag_aux	2035.5	1510.2	1757.4
dthaz	2089.3	1502.6	1778.7
aux_proj	2036.5	1596.7	1803.7

The incumbent wins the combined block against every dev-winner - merging the blocks retroactively changes no decision ever made, which is the strongest available evidence the dev-only selection was sound. Going forward: tune on all 34 origins (s_8-41) with a pre-registered acceptance rule (mean improvement AND per-origin win count AND no era-block regression); the public leaderboard is the only remaining holdout and is spent deliberately.

Repeated-split selection (adopted with the merge): training stays walk-forward (time-locked), but champion selection is resampled - choose the best config on a random half of the 34 origins, measure it on the other half, repeat (2000 iid splits + all 20 blocked 3-of-6-chunk splits, which respect window-overlap correlation). On the 5-finalist panel: logbag is the modal champion (63% iid / 50% blocked selection frequency, and just as often the true best on the held-out half); logbag_rank is the only live challenger (27%/35%); aux/dthaz variants essentially never win. Mean selection regret ~38-64 cost units. The historical dev->test split turns out to have been an unlucky draw: on the current feature set, dev-16 alone would select logbag_rank while the test half's best is logbag. Acceptance rule for future candidates: selected in >=60% of random splits AND modal under blocked splits.

Phase 11 - price-adaptive inclusion (ADOPTED: break-even top-up, fresh-only, m0.8/pm25)

Trigger: new public leader (sackriti 1477.48, late/early 804/343 + 179 in worker-limit penalties). Their +187 early for -492 late vs us is a 2.6:1 winning trade; our own dev frontier re-measured that same trade LOSING at 0.5:1 (α0.8→1.2: -213 late, +406 early). Not a modelling gap - a price gap.

The price structure: a false swap costs 0.5/day against the imputed EOL (device end_time + 30d), so it depends on distance to the split's data end: dev origins 81-133 per false swap (day-42 placement), test 18-77 falling monotonically, s_41 17.7 - and ~0.5 for scenarios adjacent to a split's data end, which the runtime splits contain. α=0.8 is the CV-optimal response to dev's prices, not a universal constant. Crucially the runtime planner and the evaluator read the same /tmp/data metadata, so a price-based rule cannot disagree with the scorer about prices.

Counterfactual harness (/tmp/counterfactual.py): same dev origins, same real deaths, same evaluator - end_time/EOL observability truncated at origin+50d to reproduce cheap pricing (~19/false swap) from local data only. There, break-even inclusion (ecroute m1.0) wins 1678 vs champion 1754; α widening still fails (+10 at best) - a fixed count cannot express "include everything break-even at today's prices".

Iterations under the merged-block protocol (panel = all 34 origins + repeated-split selection):

candidate	panel mean	random-split sel%	verdict
champion (fixed α0.8)	1715.8	-	baseline
ungated break-even top-up m1.0	+24 vs champ on real dev, worst +461	-	rejected: EV errors are punished where wrong costs ~100
+ price gate pm25 (v1 e_far clip)	1760.5	5.5%	rejected: clip made long-stale devices (imputed EOL in the past) look FREE; evaluator charges those swaps LATE 10/day (s_34 +411, s_36 +850, s_37 +668)
+ two-sided unclipped pricing	1706.6	42.5%	closer; price-0 "stale, imputed EOL in-window" fiction still costs mobility (s_23/s_41 flips)
+ fresh-only (e_far >= 42), m1.0	1703.8	25.6%	s_39 -78 / s_40 -406 / s_41 +76
+ fresh-only, m0.8	1702.0	63.0%	ADOPTED: s_40 -438.7, s_41 -102.2, s_39 +72.5; 31/34 exact ties

Blocked splits: 50/50 with champion, but every champion "win" is an alphabetical tie-break on byte-identical costs (splits excluding the last chunk are exact ties); the hybrid wins every decided split. Mean selection regret drops 36 -> 8. The cost_curves fix (two-sided, unclipped survivor pricing) is byte-inert for the old champion on all 34 origins and is adopted as a pure correction.

New champion: logbag / tsroute α=0.8 + break-even top-up (ec_margin=0.8, ec_pmax=25, fresh devices only) + fixed cost curves. The top-up is structurally empty at dev-like prices (all 16 dev origins byte-identical) and widens exactly where a wrong inclusion is provably cheap - which is where the runtime splits live. Expected effect is largest on public/private scenarios near their data end; the panel contains only ~3 such origins, so the local delta (-14) understates the deployable one.

Follow-up sweeps, both rejected by the panel: pm40 (1705.1, s_38's -721 does not survive the fresh-only restriction) and pm40/m1.0 (1715.1). pm25/m0.8 is the local optimum of the family. Docker all-48 verification of the adopted champion: 2175.4 (was 2202); late 1460 / early 451.

Phase 12 - spatial peer features (rejected: panel +106, 0% selection)

User idea: cross-room/building features, age-scaled ("what would my peers have measured at my age?"), to catch fast plungers relatively. Attractive on paper because it fixes phase-6's fatal flaw: batch death-clocks needed peers' DEATHS (14-54% coverage), while peer VOLTAGE at matched age has ~100% coverage. Recon: rooms too thin (median 4 peers); buildings workable (median 9, max 117); install-week batches age-match by construction (median 8). Building death rates genuinely cluster (4% to 100% across 24 buildings).

Screen on the blind zone (11,599 healthy-voltage pairs, 24 silent deaths):

feature	blind AUC	all-pairs AUC
deficit vs building peers (age-regressed)	0.761	0.869
deficit vs install-batch peers	0.663	0.808
building death-rate prior (smoothed, leak-free)	0.646	0.721
rng_30 (phase-8 incumbent)	0.832	0.764
v_corr_last (already in FEATURES)	0.867	0.945

Additive test: blind logistic rng_30 alone 0.843; +peer features 0.825-0.837 (worse). Panel adjudication anyway (logbag + deficit_bldg + bldg_rate on the phase-11 planner, 34 origins): 1807.6 vs 1702.0, champion selected in 100% of random and blocked splits. Per-origin AUC is unchanged (s_41: 0.967 vs 0.968) - the features do not add ranking signal, only calibration noise, which the count rule converts into worse inclusion counts.

Why it cannot work here: the master curve already performs fleet-level age-matching (tau_projected IS "voltage vs expected-at-this-point"); local peer medians re-estimate the same quantity from 5-20 samples instead of the fleet. The building death-rate signal is real but weak exactly where the misses live, and largely confounded with install batch (~90% of lifetime variance, phase 6).

Phase 13 - plunger mechanism hypotheses: gaps, heat, room cascade (all rejected)

Three user hypotheses for what precedes a fast plunge, each built leak-free per (device, origin) over all 168 origins and screened on the blind zone (11,599 healthy-voltage pairs / 24 silent deaths), then panel-adjudicated:

hypothesis	features	blind AUC	panel (champ 1702.0)	sel%
telemetry gaps precede death	gap_max_90 / gap_frac_90 / gap_trend	0.42 / 0.34 / 0.59 - INVERTED: plungers report MORE cleanly than survivors	1875.0	0.0%
hot-running or temp spikes	temp_mean_life / temp_rise_30 / tmax30_rel / temp_hot_frac_90	0.68 / 0.42 / 0.39 / 0.49 - only lifetime heat mildly ages (phase-6 confirmed); recent temp is INVERTED	1988.0	0.1%
room cascade (peer fails -> room heats -> I fail)	room_death_90 / bldg_death_90 / room_peer_temp_rise_30	0.54 / 0.52 / 0.37 - peers' temp FALLS before a death, recent peer deaths ~chance	1871.8	0.1%

Champion selected in 99.8% of random and 95% of blocked splits. Even the gap family's small additive screen nudge (rng_30 0.843 -> 0.867) becomes +173 in CV - three noisy features degrade p42 calibration, and the count rule turns that into worse inclusion counts.

Mechanism picture after phases 6-13: the fast plunger runs slightly COOLER than its own baseline, spikes less, reports cleanly, and shows no environmental or neighbourhood stress - but its voltage wobbles (rng_30 0.833, the one real precursor, itself unactionable at 0.2% prevalence). This is the signature of an internal cell failure, not an external stressor: nothing in temperature, telemetry health, or spatial context flags it before the plunge starts.

Phase 14 - planner structure audit (ADOPTED: building day-splitting + censoring-aware pricing)

With WHO-prediction mined out (phases 6-13) and ops already beating the leader's, the audit targeted WHEN precision and pricing fidelity. Four candidates, all CV-adjudicated:

candidate	verdict	evidence
split_days=7: sub-group buildings by best-day bucket so the day search may split a building across days (merge moves can rejoin at zero travel cost - same-day groups share the mobility simulation)	ADOPTED	panel 1679.2 vs 1702.0, selected in 100% of random AND 100% of blocked half-splits; sd14 also wins (1694.4) but sd7 dominates
censoring-aware cost curves: death mass beyond the split's data end cannot be charged late/emergency (never in shipped eol_times) - price it via the end_time+30 imputation branch and drop it from park	ADOPTED	byte-inert on all covered windows (panel identical to the decimal); counterfactual TRUNC=30 (last 12 window days unobservable): 1302.2 vs 1334.4 (-32), 11/16 wins - early drops 572->480 as the planner stops paying bookkeeping on unscoreable deaths
hazard-based WHEN (logbag WHO + DiscreteHazard conditional bucket-CDF day quantiles)	rejected	panel 1728.4 (+26); s_8's -26 did not generalize - the 4-knot GBR quantiles are the better day estimator
park_mobility (emergency also pays travel/overtime)	rejected	+0.8 to +2.9 on the counterfactual - noise; emergency-cost underestimation was not the binding error

One-visit-day-per-building had been implicit in plan_tsroute since phase 5 (offs = bldg.map(day_of)): a building holding both a dies-day-5 and a dies-day-40 inclusion was forced onto one compromise day, paying both sides.

New champion: logbag / tsroute alpha=0.8 + break-even top-up (m0.8/pm25, fresh-only) + split_days=7 + censoring-aware two-sided cost curves. Panel 1679.2 (from 1715.8 pre-phase-11).

Phase 15 - undocumented swaps & unlabeled deaths (all three fixes rejected; the "noise" is the label process)

User finding (voltage_all_devices notebook): timeseries extend beyond battery lifetimes - unrecorded preventive swaps. Parallel forensics (3 agents) established the full scale: 31 devices with undocumented swaps (32 events, clustering on building maintenance days - 2025-07-14 x8 and 2025-07-23 x7 in b_afdd8b1a066d, 2026-05-08 x6), detectable by a rule with a wide dead zone (consecutive-day corrected-voltage jump from <= 2.85 up to >= 3.00: identical 33-event set for any jump threshold 0.10-0.25 V; real swaps jump >= 0.28, nearest noise landing 2.989). Beyond the 2 suspected unlabeled deaths: 26 confirmed ghosts (crossed 2.45, never recovered, stopped reporting, no EOL; dropout days cluster on batch-removal dates: 2025-12-07 x13, 2026-02-19 x10) plus 32 deaths-in-progress still limping below 2.45 at data end.

Three CV-adjudicated treatments (panel, sd7 champion 1679.2):

treatment	panel	why it failed
A: truncate history at detected swap (age_days restarts, windows never span the swap)	1684.5 (+5.3; ties except s_9 +101, s_14 +122, s_32 -43)	the age error only touches age_days (weak, linear); cohort features were already rejected; truncation shortens post-swap history and loses more than it fixes
B: exclude ghosts from plans (p42=0 if below 2.45 > 42d, still < 2.55, no EOL)	2226.0 (+547)	WRONG PREMISE: EOL is the ~2.40 crossing and batteries spend >42d in the 2.45->2.40 twilight - the rule zeroed genuinely dying devices
C: quarantine "looks dead, labeled survives" training rows (3,027 rows)	2519.1 (+840), per-origin AUC collapses to ~0.51	those negatives are LOAD-BEARING: they teach the model that deep-low voltage does not reliably produce a recorded death - which is exactly what the metric scores

Conclusion: what looks like label noise IS the label-generating process, and the incumbent model has already learned it (via staleness and the low-voltage negatives). The evaluator scores recorded deaths, not physical deaths; ghosts are unscoreable, and their "survives" labels are the correct training signal for that objective. detect_swaps() and the GHOST_ZERO/QUARANTINE knobs remain in the code, inert by default, with the production table and runtime paths unchanged. Champion unchanged: panel 1679.2, all-48 Docker 2146.9.

Phase 16 - dataset semantics (two community observations verified; no model change needed)

Categorical metadata leak: the train parquets' device_id dtype carries 2,451 categories while train uses 461 - the held-out fleets total 1,990 devices (4.3x train). Implications: (a) per-split adaptive refits are the right architecture (each split has its own large fleet); (b) worker limits (fixed 100/day) bind much harder at 4x fleet size - the public leader's daily_limit penalties (108/scenario) are rational there; our optimize_days prices limits exactly, but local CV never exercises limit-saturated regimes.

end_time semantics: locations.end_time is the ADMIN snapshot end, not per-device last data - 445/461 devices share 2026-08-01. All 82 dead slots keep reporting past EOL (mean 107d, max 305d): replacement batteries under the same device_id, invisible to scoring. Verified harmless everywhere: master_curve filters tau >= 0 (replacement data never enters), dead devices are excluded from planning, and the 15 admin-removed unscoreable devices (early end_time, no EOL; e.g. 11 removed 2025-12-26) are scored p42 ~ 0.01 by the model - the "stale + frozen voltage -> survives" pattern learned from exactly these rows (phase 15's load-bearing negatives) keeps them out of every plan. The e_far imputation pricing in cost_curves makes them prohibitively expensive to the break-even top-up as well (two-sided fix, phase 11). Example decoded: d_57e99ff00802 - EOL 2026-02-02 (scored death), replacement battery reports until 2026-05-13, admin end_time 2026-08-01.

Phase 17 - label-process modeling & count correction (all five directions closed)

A structured attack on the 1679 -> 1400 gap along the label-process and count-error channels (user-provided plan). Every direction adjudicated:

candidate	panel	verdict
label-arrival hazard (two-regime: logbag for not-crossed, KM P(EOL recorded in window | crossed a days ago) for the death zone)	2837.7 (+1158, 1 win/33)	refuted: the classifier's p42 for crossed devices is ALREADY the correctly calibrated mixture - dies42 labels contain the arrival process; overriding ~0.15 with KM arrival ~0.4-0.9 inflates sum(p42), explodes N
+ domain-split training (classifier never trained/queried in death zone)	2938.1 (+1259)	refuted: most positives ARE crossed near death - splitting starves the classifier of deaths
cost-weighted training (miss ~200 vs era FP price)	2693.7 (+1015)	refuted; worst exactly at early origins where FP weights are largest
oracle-N diagnostic (per-origin N sweep, real evaluator)	mean regret 222.6/origin, median 185	the pool is real but the oracle direction is era-inconsistent (winner's curse around a shallow optimum)
backtest-alpha corrector (pick alpha from last-K closed origins' replayed curves - simulated offline from the saved sweeps)	K=4: +113, K=6: +89, K=8: +36 vs fixed 0.8	refuted: converges to picking 0.8 (18/26 at K=8); per-origin optimal alpha is unpredictable noise

The label-process channel has now been attacked five ways across phases 15 and 17 (binary zeroing, hard quarantine, domain split, KM arrival hazard, cost weighting) - each loses by hundreds. Conclusion: the bagged logistic's implicit calibration of the label-recording process is better than every explicit model of it we can impose, and the fixed CV-chosen count (alpha=0.8) is better than every adaptive corrector tested.

Per the plan's own decision rule ("if the first two deliver under ~120 combined, 1400 on the panel likely isn't reachable"): they delivered negative amounts; 1400 is not reachable on this panel without new information channels. The honest target remains deployment-era origins (ghost-dense, cheap-FP), which the shipped mechanisms - adaptive per-origin refits, break-even top-up, censoring-aware pricing, building day-splitting - already serve, and whose panel deltas systematically understate their deployment value. Champion unchanged: panel 1679.2, all-48 Docker 2146.9.

Phase 18 - limit-aware constructive initializer (rejected; the limit tail is a correct trade)

Last item of the phase-17 plan: greedy urgency-ordered construction (each group placed on the day minimizing exact partial timing+mobility cost) as an additional best-of-starts candidate in optimize_days, targeting the overtime/limit bucket (~172/origin, spikes 500-850 at s_21/23/33-38).

Panel: 1692.3 vs 1679.2 (+13.1), champion selected 88.5% random / 75% blocked. Three findings close the direction: (1) 28/34 byte-ties - the argmin-start climb already finds these optima; (2) the limit spikes are UNTOUCHED (s_21 781, s_23 832 identical) - with exact limit pricing in the objective, the optimizer genuinely prefers paying penalties because deaths cluster in time and spreading costs more in timing than it saves in limits; (3) where the constructive start did win the search objective (s_34/s_35), realized cost got WORSE (+215/+226) - at the margin, deeper search harvests model-curve error rather than real savings. The knob (optimize_days.constructive) stays available, default off.

This completes the phase-17 plan: all five directions adjudicated and rejected. Champion unchanged: panel 1679.2, all-48 Docker 2146.9.

Phase 17 - late-swap forensics: the misses are mostly NOT silent plungers (cold-start floor ADOPTED)

Question: which batteries get replaced too late? Attributed the all-48 submission's late cost through the evaluator's own transitions (its emergency loop books missed deaths as late_swap at 10/day, day ~43+):

• 98% of late cost (73,430 of 74,850) is 263 parked deaths swapped as post-window emergencies; only 1,420 is caught-but-late (deaths on days 1-7 swapped a few days after).
• State at origin of the 263: 157 already below 2.45 (48,460 cost) · 71 in visible droop 2.45-2.55 (17,710) · only 35 healthy+fresh (7,260). The phase-8-13 "silent plunger wall" is ~10% of the late cost, not the story.
• (The repo submission.csv was the phase-5-era artifact, all-48 2202; all numbers here use same-code regenerations.)

Mechanism 1 - cold-start count starvation (s_0-s_8 = 46% of all-48 late): at warm-up origins the ranking is fine (AUC 0.97 at s_0; missed deaths at ranks 2-8) but sum(p42) undercounts prevalence ~4x (3-16 unique training deaths, lower-prevalence era), so n = 0.8*sum(p42) includes 1-8 batteries while 9-19 die - catch rate 0.00/0.00/0.06 at s_0/1/2. Invisible to every prior selection round: CV discards s_0-s_7 as warm-up. Cold-start late ~700/scenario ~= the public-LB late gap to the leader (1296 vs 660) - phase 10's "not recoverable" verdict was derived on the test block only.

Mechanism 2 - label-lag suppression (bulk of the remainder): long-crossed devices (59-275d below 2.45) score p42 0.001-0.3 - the phase-15 load-bearing negatives working as trained; the minority whose EOL does arrive in-window pays 300-500 each. Still open; LogbagLAH (KM crossing->EOL arrival curve) is the drafted candidate.

Fix for mechanism 1 (ADOPTED): gated inclusion floor in plan_tsroute - psum_eff = max(sum(p42), k * prev_est), with prev_est = max(deaths recorded in the last 42d, fresh devices first crossing 2.45 within the last 42d), active only while unique observed training deaths < 20 (true exactly at s_0-s_7; the 34-origin panel is byte-identical to the champion by construction).

variant	warm-up s_0-s_7 mean
champion (no floor)	4256.1
ungated k=1.0/1.5/2.0	3994/3741/3354 - leaks into the panel (+21/+95/+299; recent-deaths overshoots post-peak), rejected by selection stability (champion 89.5%/80%)
gated, recorded-deaths only, k=1.5	3740.5 - recording lag blinds the label signal exactly at s_0-s_2 (recent42 = 1 while 9-17 die)
gated + zone signal, k=1.0/1.5/2.0	3313/3014/2532
gated + zone, k=2.5	2379.7 - wins 8/8 origins; k=3.0: 2619, k=4.0: 3191 (interior optimum)

Registered caveat: k and the gate threshold were chosen on the same 8 origins (no holdout exists there). Any k in [1.5, 3] recovers 29-44%, so the qualitative result is robust; the exact k=2.5 is not.

End-to-end verification (submit.py adaptive path, floor vs same-code control): s_8-s_47 byte-identical costs (gate closed), warm-up 4764.2 -> 2976.6 (-37.5%, 8/8 scenarios), all-48 2146.6 -> 1848.7 (late 1465 -> 1125, early 434 -> 515; the control reproduces the recorded Docker 2146.9). Champion and panel (1679.2) unchanged. Knobs: BATTERYSWAP_FLOOR (k, 0 disables the cold-start branch), BATTERYSWAP_FLOOR_GATE. Held-out-split caveat: pooled training passes the gate earlier in calendar time there (bigger fleet, faster death accrual) and sum(p42) carries the training fleet's scale - the foreign-fleet floor branch in submit.py (own_frac gate) targets that regime; the local train split cannot exercise it.

Phase 19 - cold-start count floor (ADOPTED, gated) & ungated variants (rejected)

Public resubmission of the phase-14 champion read 1717.8 (rank 3): late 1296→1234, early +14 - the deployment mechanisms delivered ~-50, so public mostly does NOT sit in the cheap-FP regime, and the remaining -240 to the leader is late-swap conversion. Cold-start diagnosis (warm-up block): at s_0-s_7 the model RANKS deaths at the very top (AUC 0.97) but sum(p42) undercounts prevalence ~4x (3.8 when 17 die) - the count starves, not the ranking; 46% of the all-48 late cost. Fix: floor the count with observable signals - max(sum_p42, k x max(deaths recorded in last 42d, fresh 2.45-crossings)) - gated to fire only while training has seen < 20 unique deaths. Warm-up CV: 4256 -> 2380 at k=2.5, winning 8/8 origins; the 34-origin panel is byte-identical (gate closed everywhere). All-48 Docker: 2146.9 -> 1851.0 (late 1465→1125).

Ungated variants, panel-adjudicated and rejected: k=1.0 recent-only +20.9, k=1.0 +zone +75.1, k=1.25 +zone +113.9 (champion 96%/80% selection). Where sum(p42) is natively calibrated it beats the empirical death rate (deaths are bursty; last-42d != next-42d), and the death-zone signal counts never-dying limpers. Sixth refutation of explicit count/label correctors in calibrated regimes.

Phase 20 - CV-harness upgrade: the harness catches a blind ship (foreign floor refuted)

Five CV improvements, all implemented:

1. Fleet-holdout harness (train the pool on half the devices, plan/score the full fleet): built to validate a "foreign-fleet floor" (fire the count floor when training positives are mostly another fleet's - the public early-origin hypothesis) that had been wired into submit.py on architectural grounds. The harness REFUTED it the same day: the probability scale transfers across device subsets (half-fleet training: sum(p42) 15.8 vs 9.1 true deaths - it OVER-counts if anything); the floor was inert on 15/17 origins and lost where it fired (+15.1, 0 wins). Branch removed before it could cost a submission. Public late-heaviness is therefore NOT foreign-scale starvation.
2. Near-end block s_42-s_47 (censored shipped truth, previously discarded): native retro-adjudication of the deployment mechanisms - phase-11 top-up worth -273 (877.6 -> 604.8), phase-14 censoring-aware pricing -64 (669.2 -> 604.8). Now a standing block.
3. Three-regime scorecard codified in the protocol section.
4. selection_stability counts decided splits only (all-tie splits no longer break alphabetically to the incumbent).
5. summarize() reports late/early/ops decomposition by default.

Champion: logbag / tsroute a0.8 + top-up (m0.8/pm25) + sd7 + censor-aware pricing + gated cold-start floor (k=2.5, gate 20). Panel 1679.2 (unchanged), warm-up 2380, near-end 604.8, all-48 Docker 1851.0, public 1717.8.

Phase 21 - forced-emergency timing: exact pricing tested, rejected

Evaluator fact (established by reading evaluate_plan's emergency loop): every recorded in-window EOL the plan never swaps gets a FORCED emergency visit - late penalty + full unoptimized ops - and the queue does not start at day 42. It starts at the end of the calendar week containing end_time (end_day = end_time + (6 - weekday)), then processes one battery per day in alphabetical order. All 48 train scenarios start on Mondays, so the first emergency is always day 48, and the k-th miss lands at day 48+(k-1). Our cost model prices parking against EMERGENCY_DAY = 45 - a systematic ~30 underprice per expected death, plus an unmodeled queue-position term.

This supersedes the phase-10 disclosure that emergency_day is "structurally inert": the phase-11 top-up and optimize_days' add/drop passes both consult park, so the knob is live now.

candidate	verdict	evidence
emergency_day=48 (exact)	rejected	panel 1682.8 vs 1679.2 (+3.6); 31/34 plans byte-identical - the repricing only flips decisions at near-data-end origins (s_39 -74, s_40 +67, s_41 +130); near-end block s_42-47 631.9 vs 604.8 (+27), 0 wins / 4 losses / 2 ties
emergency_day=50 / 52 (exact + queue pad)	rejected	+9.4 / +13.1, monotonically worse

Why the exact price loses: the deaths we park are mostly voltage-silent (p42 ~ 0.05), so +30 x p42 adds ~1.5 to their park price - never enough to include them. What the pricier park DOES flip are marginal p42 ~ 0.3-0.5 devices that mostly survive: late improves -3.5 while early pays +6.6 (panel decomposition). The champion's underpriced park is a de-facto regularizer against over-inclusion where FP prices collapse. Same conclusion as the phase-14 park_mobility rejection from the other side: emergency-cost underestimation is not the binding error - p42 on silent deaths is.

Phase 22 - optimizer-constraint challenges: sd<7 and NN visit order (both rejected)

The evaluator's "constraints" are soft prices (weekly limit flat 100/blown week, daily flat 100, overtime 2x/h past 8h; ONE worker hard-wired in the state machine - no worker field exists in the plan schema) and mobility_cost replicates its cost machine exactly, so the ~152/origin the champion pays in penalties is a purchased trade (phase 18 already adversarially confirmed). The two remaining SELF-imposed constraints were challenged:

candidate	verdict	evidence
split_days=3 / 1 (finer group granularity)	rejected	panel 1685.0 / 1691.1 vs 1679.2; ops rises 216.6 -> 223.7 -> 228.2 - finer groups fragment the local search's argmin start and it converges to worse optima with more travel; more freedom != better local search
nn=True (visit a day's buildings in NN route order from base instead of name order, priced in the sim AND emitted identically)	rejected	panel 1679.4 vs 1679.2 - tie; 8 wins / 13 losses / 13 ties; per-origin moves up to +-40 are search-trajectory perturbations, not route savings (with a symmetric travel matrix 2-building days are order-invariant; 3+ building days are rare and route deltas are <1 each)

Machinery kept inert: optimize_days.nn_order (default False, cfg key 'nn'), assemble(sort_key=). Ops pot accounting: 217/origin total = 152 penalties (correctly purchased) + 54 travel/changes + 11 swap/room time; the recoverable slice was bounded by ~5-10/origin and the tests found ~0. The ops channel is CLOSED - remaining cost concentrates where it always was: late_swap (p42 on silent deaths).

Provenance note: the first nn panel run was invalidated by a parallel-session file race (a peer harness's frozen-file guard checked models_cv.py out to HEAD mid-experiment; nn configs silently ran as no-ops - detected because knob configs came back byte-identical to baseline). Long background runs now print the imported module's md5 and assert the knob under test exists.

Phase 23 - worker-limit saturation campaign (all rejected; non-adaptation is optimal)

The one regime local CV never samples: held-out fleets are ~4.3x train (phase 16), where fixed worker limits bind and the public leader pays ~108/scenario in daily-limit penalties. Harness (research/run_sat.py, frozen): divide each scenario's worker limits (weekly/daily/overtime_start) by 4 on real panel origins - planner and evaluator share the transformed settings (the phase-11 property). Champion baseline on the block: 2115.1 (vs 1679.2 native) = late 1022 / early 446 / OT 112 / daily 262 / weekly 212 / travel 41 - so ~585/scenario of limit-driven penalties, with late/early unchanged from native: the champion does not adapt its plan to congestion.

Autoresearch protocol (mutable surface = opt-knob JSON only): s0 = 6-origin saturated screen; s1 = all 34 saturated origins + native no-regression leg. 10 s0 experiments + 1 s1, all rejected:

knob family	best variant	s0 delta	verdict
constructive start (phase 18)	on	+16.2	worse basin under saturation too
search budget	r6/i120	-13.4	one -80 escape in six origins; 5 ties
split_days	12 / 4	-4.1 / +19.4	both directions wash; limits untouched
day_q	0.2	0.0	identical plans - dead knob under saturation
top-up gate	pm12	+116.3	starving the top-up raises late AND limits
alpha	0.75 / 0.7	-59.8 (4/6) / -10.5	promoted at 0.75+sd12 combo -> s1

s1 (alpha 0.75 + sd12): sat +20.0 (18/34 wins but 4 blow-ups sum +1835: s_17 +628, s_8 +459, s_31 +391, s_23 +357 vs 18 wins summing -1317), native +42.0. Decomposition: late +43.5 / early -27.0 / daily +5.9 / weekly -2.9 - inclusion cuts buy ZERO congestion relief; they are a mortality bet with fat-tailed downside, made worse by saturation (emergencies consume the same scarce worker-hours). Screen lesson: 6-origin s0 is high-variance for inclusion knobs (-59.8 flipped to +20.0 at s1); the s1 mean gate caught it.

Conclusion: the ~585/scenario limit spend is the shadow price of scheduling MASS, not of bad choices - the day-assignment search is converged (extra search, day quantiles, both sd directions, constructive start: all flat or worse), and paying the flat fines beats the timing cost of leaving peak mortality weeks. The champion's non-adaptation to congestion is optimal within its plan space; the leader's limit penalties on the LB are likewise a purchased trade, not an exploitable gap. The saturated-regime channel is CLOSED; ec_margin was left untested as the same condemned inclusion family.

Phase 23 - unmasked foreign pooling in deployment (ADOPTED; user change, holdout-validated)

User change to submit.py: the packaged train table is pooled UNMASKED on held-out splits (all 66,890 closed-window pairs, 82 dead devices, explicit window_closed filter), while the target split's own pairs - and any train-table rows for devices in the target fleet - stay under the T-42d mask. The old uniform mask was CV hygiene misapplied to deployment: it starved a held-out split's early origins to 3-20 unique training deaths (the cold-start mechanism). Byte-inert when the target split IS train (own-fleet guard), so panel / warm-up / all-48 are unchanged by construction.

Validation (pooling-holdout harness /tmp/pool_holdout.py: fleet A = packaged table, fleet B = target, B-only scenarios scored):

regime	warm-up s_0-7	spots s_16/24/32/40
old masked pool, no floor	1478.9 (late 1274)	1141.9
old + gated k=2.5 floor (shipped before)	1091.0	1141.9 (gate closed)
new unmasked pool (ships now)	1108.6	1088.4

• Count calibration transfers: held-out s_0 sum(p42) 0.5 -> 3.4 vs 4 true deaths. The starvation the phase-19 floor patched is gone at its source.
• Warm-up vs the floor: statistical tie (5W/3L; s_4 -827 for the floor - a local prevalence burst that own-fleet recent-death signals catch and foreign calibration lags). Profiles differ: floor buys insurance early (480 early), new stays lean (177 early / 786 late).
• The undersold gain is MID/LATE held-out origins: -53.5 (3W/1L, s_16 -438) - the unmasked pool improves every held-out origin's training set, and a real split is mostly mid-history origins.
• Deployment consequence: on held-out splits the floor gate NEVER opens now (uniq >= 82 from origin 0); the floor remains live only for genuinely history-poor local regimes (train warm-up block, all-48 metric).

Phase 24 - pretrained time-series foundation models (rejected at tier-0 screen)

Question: can a pretrained model improve the metric? Rules check (nora.ai docs 04): pretrained models ARE allowed if open-source, documented, reproducible, and fully self-contained at evaluation (no network; 30-min CPU-only, 32 GB). The runtime image ships torch 2.13 but NOT transformers/chronos/timesfm - deployment would mean vendoring inference code

• committing weights. Applicable class: zero-shot TS foundation models (Chronos-Bolt, TimesFM, TTM, MOMENT - all Apache-2.0/MIT). No public battery-specific pretrained RUL model fits this data (BatteryML et al. target Li-ion charge/discharge cycling curves, not discharge-to-death telemetry).

Test: amazon/chronos-bolt-small zero-shot on per-device daily temp-corrected voltage (365-d context, 64-d quantile forecast, leak-free at each origin; 0.2-0.5 s per ~400-device origin on MPS; scripts /tmp/fm_screen.py, /tmp/fm_build_all.py, /tmp/fm_tier0.py; features for all 71,658 (device, origin) pairs built in 26 s). Derived: P(cross 2.40 V in window) per day (quantile-interpolated), q10/median margins at window end, days-to-crossing.

Signal screen (6 tier-0 origins): fm_margin_end AUC 0.937 all-pairs vs v_corr_last 0.948; crossed bucket 0.621 vs slope_lastk 0.672; blind/dark "wins" (0.815/0.986) rest on 3-4 deaths. Raw crossing-prob mass overdispersed 4-9x (sum 52-62 vs 6-15 true deaths) - the exact WienerFHT failure (phase 9).

Tier-0 (exact replica of run_experiment.tier0, lab.py untouched; baseline reproduced bit-exact): +fm_margin_end auc_crossed 0.757, +fm_pmax42 0.801, +trio 0.752, +fm_pmax42 dark-gated 0.781 vs baseline 0.802 (gate >=0.822 or dying_p >=0.544): all four REJECT - neutral at best, and the margin/trio variants actively degrade the crossed bucket.

Why, mechanically: a zero-shot forecaster of the voltage series is a smoothed level/trend extrapolator - exactly the signal v_corr_last / tau_projected / slope_lastk already saturate (phase-9 redundancy, third confirmation). The misses that matter have flat voltage at origin (phases 8-13): no forecaster OF THAT SERIES can see them, pretrained or not. Other FM families (TimesFM, TTM, MOMENT embeddings) read the same series and inherit the same wall; embeddings-as-features on 82 deaths is the already-rejected deep-RUL sample-complexity problem. Direction CLOSED absent a new signal source.

Phase 24b - static fit for held-out splits (ADOPTED; deployment simplification)

Follow-up question to phase 23: with the full foreign pool available at every origin, do the per-origin refits (foreign pool + target's own accumulating masked history) still earn their runtime? Holdout harness (/tmp/static_holdout.py, fleet A=packaged / B=target, 16 origins):

regime	warm-up s_0-7	mid/late s_12-40	all 16
rolling refit (phase-23 shipped)	1108.6	1065.4	1087.0
static: ONE fit on A_full	1072.8	1042.6	1057.7

6W/7L/3T - a statistical tie with a static lean (mean gap driven by s_36 -372 vs s_40 +142). The sum(p42) columns are near-identical everywhere: with 82 foreign deaths in the pool, the target's own few masked deaths perturb the fit without informing it. Phase 20's "the probability scale transfers" holds all the way down: even fleet-specific own-history adds no measurable calibration.

Adopted in submit.py: held-out splits (own-fleet fraction < 0.5 vs the packaged table) fit ONE model per split on the foreign pool and skip the split-table build entirely - equal-or-better score, minutes instead of ~17 per split, near-zero time-budget fallback risk. Train-fleet splits keep the rolling self-masked path (their labels are the answer key; also how the all-48 metric is produced). Verified: synthetic held-out split (B-only data, A-only table) takes the static branch and plans sanely (4/5/7 swaps vs 4/6/7 true deaths at s_0-2); all-48 Docker must reproduce 1851.0186805555559 byte-exact (train path untouched).

Deliberately NOT adopted: training on the target split's own post-T labels (the runtime ships each split's full eol_times.csv). That is answer-key training - near-oracle scores, unmistakable on a leaderboard, and the organizers run and read this code. The label-integrity assertion stays.

Phase 25 - HVAC environment class as a feature (rejected at tier 0; gradient is real but absorbed by voltage state)

EDA first (research/hvac_room_profiles.py, hvac_signatures.png): device temperature signatures cleanly classify the equipment environment - 8 classes from level, weekday-vs-weekend diurnal schedule, and seasonal swing (scheduled AHU in unheated fan rooms sits at 1.5 C nights/weekends and steps to 8 C on weekday operating hours; hot plants sit >30 C year-round; the -18 C minima are Nordic AHU intakes in winter, not freezers). Raw death rates are U-shaped in the class: cold classes 25-40%, hot plants 20-30%, mild indoor ~10%; median dead lifetime 750 d (cold) vs 1000 d (warm). Motivation vs the three prior temperature rejections (phases 6/8/13): those were monotone or recent-window features; two separate exposure arms can express the U, and a static class is low-noise where the phase-12 building death-rate prior was sampling-starved.

Leak-free encoding (research/hvac_origin_features.py, expanding life-so-far per (device, origin_idx)): cold_frac (days with min < 5 C), hot_frac (days with max > 32 C), sched_amp (weekday-weekend daily-range gap). Candidate hvac_env_logbag = exact logbag replica + the 3 columns in the WHO classifier.

Tier 0: REJECT - auc_crossed 0.776 vs baseline 0.802 (gate +0.02), dying_crossed_mean_p 0.450 vs 0.453; only count_err improved (5.23 vs 5.70).

Post-mortem screen (66,890 closed pairs): cold_frac standalone AUC 0.623 all-pairs (same league as phase-6 lifetime temp) but 0.400 in the crossed bucket - inverted, sched_amp 0.419; additive in-sample gains +0.001/+0.009. Within v_corr_last deciles, cold_frac AUC is 0.23-0.56 (noise-to-inverted): the lifetime mortality gradient is fully absorbed by voltage state at origin - a cold-stressed battery's depressed state of charge IS the voltage the model already reads, and among crossed devices the environment prior points the wrong way (consistent with phase 13: plungers run cooler than baseline). Class one-hots would inherit the same wall (functions of the same signatures). Fourth and final temperature rejection; the environment channel is CLOSED for WHO-ranking. The EDA artifacts stay in research/ (the class map itself is sound and may serve ops/reporting, e.g. cold-room devices as a maintenance batch - just not the model).

Phase 26 - deployment alpha under the static foreign fit (ADOPTED: ALPHA_STATIC=1.5)

Public 2026-08-19 (phase-23 submission 3c7b883): 1661.45, rank 3 (from 1717.8; the holdout harness predicted the 30-80/origin band correctly). Remaining gap to sackriti (1477.48) = 184, entirely the late/early exchange: ours 1212/137 vs theirs 804/343 - they buy -408 late for +206 early. Ops now slightly better than theirs.

Mechanism found: on a held-out fleet the static foreign fit's sum(p42) UNDERCOUNTS (5.5 vs 8 true at s_16, 6.2 vs 9 at s_6), so N = alpha*sum(p42) needs a larger alpha than the panel optimum. The panel's 0.8 is the optimum for a CALIBRATED count on the native fleet - a regime public never sees. Holdout sweep, 4 runs (3 seeds + reversed fleet direction), 16 origins each:

alpha	0.8	1.0	1.15	1.3	1.5	1.7	2.0
mean total	1156.3	1045.1	1006.8	956.5/967.2*	916.4	908.3	898.8

(*separate run batches; 1.3 anchors both.) alpha=1.3 beat 0.8 in ALL FOUR runs; the curve flattens from 1.5 (1.5->2.0 gains only -18, within noise). ALPHA_STATIC = 1.5 adopted (env BATTERYSWAP_ALPHA_STATIC): captures ~93% of the gain from 0.8 while staying off the plateau's unbounded edge - over-inclusion exposure grows there if the real public fleet is better calibrated than the half-fleet proxies. At 2.0 the late/early profile (446/331) matches the leader's shape almost exactly.

Applies ONLY to the static held-out path; the panel champion (rolling, native fleet) keeps alpha=0.8, where the same sweep was decisively rejected (phases 1/6). The train path never reads ALPHA_STATIC - all-48 unaffected.

Phase 27 - held-out-parity audit & dress rehearsal (2 fixes shipped)

Question: does our setup correctly reflect public/private mechanics? Audit of what can vary per scenario + a Docker dress rehearsal of the EXACT remote invocation (main() over fake public+private splits with renamed device IDs so the own-fleet guard routes to the static path).

Findings:

1. Per-scenario settings DO vary in train - 27 of 48 scenarios differ, but only in base_location/base_room (16 distinct technician bases). The mobility sim reads the live settings object, so this was already handled. All penalties/limits/window are constant 0.5/10.0/42 across train.
2. Hardening shipped: penalty wiring. cost_curves priced with hardcoded EARLY/LATE while the evaluator charges per-scenario settings values - now wired through per scenario (byte-inert at default values; checksum confirms). planning_window_days != 42 now logs loudly (cannot be rewired: the 42d grid is baked into features/quantiles).
3. Hardening shipped: failure isolation. The rehearsal exposed that a data-shaped exception (e.g. a check_dataset assertion) killed the adaptive path AND the fallback identically - the exception escaped main(), and NO submission.csv was written: one bad split destroyed all splits' plans. Now three layers: load_best_effort (consistency check demoted to a logged warning), individually guarded pickle fallback, and a raw-file last resort (defer-everything plans; finite score, cannot poison other splits), itself guarded so a split can at worst lose only itself.
4. Rehearsal after fixes: public + private both take the static path (82 deaths, own-fleet 0.00), the injected inconsistency logs as a warning, complete 2-split submission.csv in 35s. All-48 checksum byte-exact (1851.0186805555559).
5. Not verifiable locally, accepted as residual risk: held-out fleet size (~4x per the metadata leak; count rule is a ratio, runtime proven by two real submissions), scenario dates/weekdays, and the held-out label-lag process (the ALPHA_STATIC=1.5 bet rides on the undercount direction).

Phase 28 - public-repo port: Attention-Heads solution mined (features/floors/queue rejected; cliff-gated static appetite CONFIRMED by the barrier-event resampler - adoption candidate)

Source: team Attention-Heads (public rank 5, 1798) left their full solution public at huggingface.co/Attention-Heads/BatterySwapAI2026-attention-heads, including a SOLUTION.md with measured ablations and a component breakdown against us. (Other public repos are untouched templates; sackriti's is private again.) Everything transferable was ported and adjudicated. Runner: exp_x28.py (new file; frozen modules imported, never edited, md5-asserted per run); outputs /tmp/x28/. Parity: champion byte-equal on s_8/s_9 vs research/baselines/tier1_champion.csv. Their independently-derived core matches ours (recorded-EOL-only force visits, exact evaluator replica as the objective, censored-anchor survivor pricing, return-leg double count) - the deltas tested here were their hourly "cliff" machinery, per-battery appetite scaling, miss-queue park pricing, and peer-offset features.

Recon (facts stage): the data is exactly hourly (tele_n7 median 168). Their headline mechanism - hourly mins/hours-below-threshold see a knee days before the OFFICIAL smoothed series (temp-gated daily median + 7d rolling median), ~30% of their false negatives - does NOT transfer as detection: our features never used the 7d roll, so we never had that lag. Cliff mask (>=35% of the last 7d's readings below 2.45 V, >=20 readings, corrected v < 2.46) catches 54% of dying panel rows but only 2/61 of the champion-blind pre-crossing deaths; its catches are already-crossed knee devices - i.e. an inclusion signal, not a detection signal (consistent with phase 30's label finding).

candidate	tier-0	panel (champ 1697.4)	static holdout, α=1.5 (base 922.9, 5 fleet splits x 16 origins, paired)
hrly / hrly2 / peer features	REJECT (auc_crossed 0.802/0.795/0.790 vs 0.802)	not run	-
cliff_floor25 / 50 (floor p42 on cliff rows)	floor25 promoted (auc_crossed 0.835)	+68 / +149 regression - floors distort the calibrated count channel the native path lives on	−16 / −35.5, worst +818
cliff_floor50p (park-mode: floor rank/count only, timing raw)	-	not run	−11.3 - their scale_mode='park' lesson does NOT transfer
cliff_rank22 (rank-only reorder)	-	+13 tie	-
cliff_count22 (rank_score & p42_sum = clip(2.2·p42,1) on cliff rows; raw p42 keeps timing honest)	n/a (p42 untouched)	+27 (late 904 / early 591 vs 1022/442)	−54.5 (t −1.9, 4/5 splits, worst +319; late 452 / early 308 vs 580/226)
cliff_count_lg (+ lingerer 0.7 downweight)	-	1698.9 tie (modal 40% blocked, 31% random < 60% bar)	−25.7 - lingerer downweight subtracts value
logbag+tsq (miss-queue park penalty in the ec top-up)	-	+1.7 inert	+0.8, 0/48 wins - the queue term never clears the gated break-even channel; REJECT

Controls: plain alpha raise gets the same mean (α=2.0: −55.3) but noisier (sd 298 vs 260), worse tail (worst +421), and the curve TURNS at α=2.5 (+32.5) - phase 26's over-inclusion fear is real; the cliff gate is self-limiting (only knee devices scale, appetite cannot run away on survivors). α-pairing sweep: 1.5 stays optimal under the boost (1.2: −31, 1.35: −23, 1.7: −44 noisier).

Verdict (initial): cliff_count22 @ ALPHA_STATIC=1.5 looked like a CANDIDATE for the static held-out path only (same gating pattern as ALPHA_STATIC; the native rolling path decisively prefers the raw count - panel +27), pending a label-noise robustness check because the paired t ≈ −1.9 sat inside the range where phase-29's label-arrival luck can decide the sign.

Label-noise check (final): REFUTED. Fixed holdout plans for logbag@1.5, logbag@2.0 and cliff_count22@1.5 re-scored under 120 common-random-number label draws (phase-29 lognormal+cure lag model refit on the full table: mu 4.08 / sigma 1.33 / cure 0.205; crossing dates real; per-cell synth eol covers ALL fleet-B batteries - verified, so any crossed device can die in any draw; /tmp/x28/holdout_noise.py, results holdout_noise.csv). Paired delta vs logbag@1.5: cliff_count22 +60.5 ± 51.2, better in 8% of draws; plain α=2.0 +79.6 ± 50.9, better in 6%. The real-label −54.5 edge flips sign: it was one favourable realisation of which knee-pool members got labelled in-window, not mechanism. Caveat the check inherits: the fitted unconditional lag median ~59 d is longer than the observed-event median 33 d (cure/censored mass drags the fit), so the simulator over-weights long-lag worlds.

Sensitivity bracket (short-lag world): same fixed plans and draws, lag refit on observed events only (median 37 d, sigma 1.06, pi fixed 0.205; /tmp/x28/holdout_noise_short.py): cliff_count22 flips to −34.8 ± 41.1, better in 81% of draws (plain α=2.0: −18.2, 68%). The candidate's sign is therefore decided by the label-process model, so a third, mechanism-true test was built.

Barrier-event resampler (decisive): the edge is REAL. Phase 30 established the recorded EOL IS the 2.40 V corrected crossing (+ recording lag, measured here independently: median 5 d, p90 21 d, n=60) - so the right randomness for a parked knee device is a barrier event (descent continues through 2.40 vs stalls above it), not a lag draw; both lognormal simulators conflate descent time with recording delay. Resampler (/tmp/x28/barrier_resampler.py): reality kept everywhere EXCEPT the contested pool (in the 2.45 band at origin, not yet through 2.40, staleness<=7; pool sizes 2-25 per cell), whose outcomes are redrawn from empirical level-bucketed continuation rates (P(cross 2.40 within 42d): 0.47 / 0.69 / 0.79 for [2.44,2.45)/[2.42,2.44)/[2.40,2.42), pooled 7d state grid) plus empirical delay and recording-lag quantiles, common random numbers across plans. Result over 120 draws: cliff_count22@1.5 −66.8 ± 24.1 vs logbag@1.5, better in 120/120 draws (p95 −28.3); plain α=2.0 −51.2 ± 15.6 (also 100%, but ~16 worse than the cliff gate in every regime that rewards appetite).

Verdict (final): cliff_count22 @ ALPHA_STATIC=1.5 is a CANDIDATE FOR ADOPTION on the static held-out path - real-label holdout −54.5 (4/5 fleet splits), barrier-event resampler −66.8 (sign held in every draw), and the one adverse result (censoring-corrected lag world, +60.5) is attributable to that model's known mechanism mismatch. The native rolling path keeps the raw count (panel +27; same gating pattern as ALPHA_STATIC itself). Wiring cost: submit.py computes 3 hourly columns (frac245_7, tele_n7, corrected-voltage gate) from the split's own raw timeseries and wraps the static model's predict; all-48 must stay byte-exact (train path untouched by construction). Adoption is a human decision - not shipped in this phase. Side finding: under the censoring-corrected draws, appetite raises beyond ALPHA_STATIC=1.5 price negatively across the board, retroactively supporting phase 26 stopping at 1.5. All other phase-28 candidates REJECTED.

Independence scrutiny (closed). Continuation events are NOT independent across devices: per-device first-band-entry states (n=127, 54% continue) overdisperse strongly across calendar quarters (dispersion 28.5 vs binomial 6.0, bootstrap p<0.001 - hot/cold continuation regimes exist; buildings only mildly, p=0.10; /tmp/x28/dispersion_check.py). A correlated variant of the barrier resampler (shared per-quarter logit shock, sigma=0.88 calibrated to the observed ~4.7x variance inflation, maximally correlated within draws; /tmp/x28/barrier_corr.py) leaves the verdict unchanged: cliff_count22 −66.1 ± 26.1, better in 120/120 draws (p95 −20.2) - cold-quarter tails are bounded because a wrongly-included knee device sits near its imputed EOL (cheap false swap), while hot quarters pay the 20:1 late rate. The quarterly continuation regime itself is new label-process structure future phases should model - see phase 31 ideas 3-5 for two candidate mechanisms recorded there: seasonal temperature (warmer devices live longer, r=+0.20, so winter = hot-continuation regime) and install-batch waves (85% of lifetime variance is batch; batches descend together, confounding quarter with batch). The correlated-shock verdict above is agnostic to the mechanism: the shock was calibrated to the observed dispersion, and batch waves cluster in calendar time, so per-quarter sharing proxies either cause. Not ported (already ours or already refuted): unconditional recorded-EOL target, censored-anchor pricing, exact-cost-model optimization, building group moves (phase-7/22 search saturation), isotonic rejection, CDF lag (they over-swap early; we under-swap).

Public outcome (2026-08-20, submission 6ea03da): REFUTED - and it refutes phase 26 with it. 2188.78 vs 1661.45 (3c7b883, still the live phase-23 rolling config at alpha=0.8). This was the first public exposure of the ENTIRE static stack at once (24b static fit + 26 ALPHA_STATIC=1.5 + 27 wiring + 28 cliff boost). Component signature: battery_swap 3.02 -> 5.08 (+68% swaps), early 137 -> 582 (+444), ops +70, late 1212 -> 1225 (+13 - ZERO recall gain). That is pure over-inclusion: on the real held-out fleet the probability scale transfers (phase 20's finding: half-fleet sum(p42) 15.8 vs 9.1 true, it OVER-counts), and the half-fleet holdout construction that phase 26 and phase 28 both relied on manufactures the undercount their gains priced in. Late cost on public is saturated by voltage-silent deaths (the phases-1/6 result, again). Defaults reverted (ALPHA_STATIC=0.8, CLIFF_SCALE=1.0; static-fit machinery and phase-27 hardening retained). Standing lesson for every future phase: the half-fleet holdout is refuted as an appetite adjudicator - it disagreed with phase 20 and public sided with phase 20; appetite changes are public-look-only decisions now. WHY it mispredicts (parallel session, deployment-faithful sweep same day): the half-fleet scenarios carry ~6 deaths against public's much larger fleet, so added inclusions look cheaper than they are; the measured marginal exchange rate is 0.46 late-saved per early-spent at alpha 1.5->2.0 locally (0.21 at 2.0->2.5) and ~0.03 on public - all far below the 1.0 break-even. The half-fleet holdout remains valid for structural/mechanical questions (static-vs-rolling, failure isolation), not for anything count-related. The cliff boost's solo effect at alpha=0.8 is unproven (it was only ever measured on top of the refuted 1.5 baseline) and stays off by default.

Phase 29 - fresh-eyes slate: generative label model, arrival rank/features, planner add/drop, policy control (all rejected; label-noise floor quantified)

Trigger: user-requested outside review ("maybe we went in the wrong direction from the start"), ideated deliberately BLIND to this file (code + raw data only), then every proposal tested. Runner: exp_x27.py (new file; frozen modules imported, never edited; md5s logged per run - models_cv.py 7ead54af... matched the pinned hash in all four runs). Parity: champion reproduced byte-equal on s_8/s_9 vs research/baselines/tier1_champion.csv. All runs at the tier-1 budget (r0/i20); champion reference 1697.4 on the 34-origin panel. Outputs in /tmp/x27/. (Numbered 29: phase 27 = parity audit, 28 = public-repo port, both from parallel sessions.)

Raw-data facts established during the blind review:

• All 82 recorded deaths crossed 2.45 V before EOL, and 42d before their crossing ALL sat in the 2.47-2.77 descent band (median 2.575); the

  2.75 -> 2.45 descent takes median 229d (p25 108, min 4). Detection is saturated; the residual is timing (independently re-derives phase 15/17).

• Crossing -> recorded-EOL lag: median 33d, p90 ~140d - wider than the window; the label process is the dominant randomness.
• All 82 deaths transmitted up to their EOL day (median pre-EOL data silence 0d, p90 also 0): labels REQUIRE telemetry, devices dark at their would-be death ghost out. 33 undocumented swaps detectable (median pre-swap v_corr 2.667: proactive replacements; 30/33 unlabeled).

candidate	mechanism	panel	sel% rand/blk	verdict
lagrank_b03	lognormal+cure arrival prob as RANK blend (w=0.3), count untouched	1716.6 (+19; 10W/7L/17T)	14.8/5	rejected
lagrank_ovr	same, rank override for crossed devices	2698.2 (+1001; 1W/33L)	0/0	rejected - closes the phase-17 loophole: LAH fails through ORDERING too, not only count explosion
twostage	generative EOL date: Wiener 2.45-crossing CDF (x) lognormal+cure lag	2924.6	0/0	rejected
ts_anchor	logbag p42 multiset (exact count), twostage order	2659.6	0/0	rejected - the ordering itself is destructive, not just the count
logbag_db245	days_below_245 as plain logistic feature	1747.2 (12W/20L)	9.3/15	rejected
logbag_arr	stacked arrival probability as feature (soft LAH; count cannot explode)	1755.5 (15W/15L)	7.2/20	rejected
logbag_dark	staleness>14 cliff flag	1733.0 (13W/9L; one +1245 blow-up s_36)	18.9/5	rejected
adrop planner	add/drop moves in the day search, EXACT marginal-mobility pricing	1769.4 (+72)	-	rejected: buys late -46 for early +130 - phase-21's losing trade reached from the planner side
policy rule	3-param direct rule (tau_projected<=theta, dark cliff, day frac), odd-tuned / even-holdout	2094.8 full (holdout 1997.6)	-	rejected: the ML stack is worth ~400/scenario; the grid independently selects the 14d dark cliff

Champion selected 49.9% random / 55% blocked against the whole slate - stands. Static-holdout regime not run: nothing passed the panel gate (necessary condition per the three-regime protocol).

Mechanism findings:

1. Why generative label models fail: cure-tail unidentifiability. The walk-forward lognormal+cure MLE collapses pi -> 0.00 at every panel-era refit (63-69 events): sigma inflates to absorb the censored ghosts, every crossed device "eventually gets a label", and sum(p42) overcounts 2-3x (s_24: twostage 26.9 vs 9 true deaths; logbag 15.2). On the FULL table the cure mass IS identified (pi=0.205, mu=4.08, sigma=1.33) - the ghost-vs-tail decomposition becomes identifiable only with data no walk-forward fit can have. This is the structural reason every explicit label-process model (phases 15, 17, 29) loses to the discriminative classifier: dies42 already integrates the arrival process over exactly the horizon that matters, using each of the 82 events once at full efficiency.
2. Label-luck noise floor (new standing capability: exp_x27.py noise). FIXED champion plans re-scored under 150 resampled label processes (real crossing dates; lognormal+cure lags, per-scenario truncation so devices unlabeled at an origin stay unlabeled): panel-mean sd ~710 across label worlds; the PAIRED champion-vs-challenger delta (common draws) has sd 49 (p5 -60 / p95 +103) for a pair whose real-panel gap is 19. Caveat: the fit makes synthetic worlds harsher than reality (mean 3895 vs 1697), so this overstates - level-scaled it is still ~+-22. Reading: panel deltas under ~50 are substantially label-arrival luck; this is the mechanism behind the 61-76 selection regret the stability runs measure. Corollary: lagrank_b03 BEATS logbag by 25+-49 in the synthetic worlds (where ghost labels do arrive) - its mechanism is real; the real world's label generosity decides the sign, and it decided against.

Verdict: champion unchanged (panel 1697.4 tier-1 / 1679.2 full budget). The wrong-direction-from-the-start hypothesis is adjudicated NO from four independent angles (generative date model, arrival-informed ranking and features, EV-driven planner inclusion, model-free policy rule): at 82 events, bagged logistic + count-based topsum remains the sweet spot. The durable addition is the label-noise simulator - a future candidate whose panel delta is under ~50 should be treated as undecidable on this panel regardless of selection percentage.

Phase 30 - label event identified: recorded EOL IS the 2.40 V crossing (knowledge adopted; both exploitation channels closed by forensics, no panel runs spent)

Follow-up to phase 29 ("try harder"), forensics-first: identify the label mechanism, then test exploitability directly on the champion's actual panel misses (the saved phase-29 plans), BEFORE building any model.

1. Mechanism. 81/82 recorded deaths reached v_corr < 2.40 before their EOL; lag from first-2.40-crossing to recorded EOL: **median 5d, p25 1d** (vs 33d from the 2.45 crossing - phases 15/17 measured the wrong clock: most of the "administrative lag" is descent time 2.45 -> 2.40). v_corr at EOL median 2.396. Ghosts: of 34 devices crossed-2.45 >= 42d with no EOL, 20 NEVER went below 2.40 (min v_corr median 2.405) and the 14 that dipped sit at 2.400 exactly. To first order the label process is a barrier event at ~2.40 with days-scale recording, gated on the device transmitting.
2. Channel A - knee-projection inclusion: closed. Champion panel misses: 132 (3.9/origin; v_corr median 2.468: 60 crossed / 48 droop / 24 healthy). Under per-device sub-threshold extrapolation (below-2.55 slope over the last 90d, projected days to 2.398 minus staleness) the misses rank median 35 among non-swapped transmitting devices (best 6, ZERO in top-5, 7 in top-10); their own projections say median 53d - the final descent accelerates past any linear fit. Catching the median miss means ~35 false swaps x ~100-150 early cost against ~300-450 saved. The knee pool (30-90 devices/origin) is near-exchangeable at the 42d horizon: WHICH member punches through is not predictable from trajectory state.
3. Channel B - hover-ghost demotion: closed (and instructive). The rule "v_corr in [2.395, 2.47), stale <= 14d, knee slope < 0.5 mV/d" looks like the ghost signature at device level, but at pair level that cell dies at 34% in-window - and the champion's inclusions from it die at 47% vs 17% for its exclusions. logbag's p42 already discriminates within the knee pool BETTER than mechanistic extrapolation (drop_recent/below_frac carry it). Phase-15's GHOST_ZERO loss now has a mechanism: "hover" is a device-level survivorship artifact, not a pair-level state.

Conclusion: the panel's residual late (~1022/origin) is the exchangeable knee-pool hazard, not a selection error. Combined with the phase-29 noise floor (paired sd ~49), the per-device channel on the native panel is closed with a mechanism, not just a scoreboard. The 2.40-barrier fact stands as the correct mental model of the metric's label side - and of ALPHA_STATIC's bet: held-out knee pools are larger and FP is cheap near data end, so buying the pool wholesale is right exactly there. Champion unchanged. Artifacts: /tmp/x27/miss_forensics.csv; zero panel runs consumed by this phase.

Phase 31 - user ideas screened: donor-curve k-NN and best-case-deficit (both rejected at tier 0; tau_projected residual re-ranking discovered as a deployment candidate)

Two user-proposed similarity channels, adjudicated by the phase-30 forensic protocol (rank-of-misses on the champion's saved panel plans + cell AUCs) - zero panel runs spent. Screen code: exp_knn_screen.py, /tmp/x27/knn_screen.log.

Idea 1 - donor-curve k-NN EOL (align recent v_corr history against each INDIVIDUAL dead device's trajectory, k=5 weighted by fit error, donors restricted to deaths <= origin): the only channel all day that beats the champion's own residual ordering - misses rank median 16.5 (knee channel: 35), 27/102 in the residual top-5 vs logbag-next's 21. But it is DOMINATED by plain tau_projected re-ranking (35 in top-5), loses global AUC to it (0.883 vs 0.926, 1/34 origins) and times deaths worse (median |err| 14.7d vs 11.5d). Verdict: rejected - the median master curve + per-device alignment already saturates the donor-similarity signal; individual donors add k-NN variance, not information (descent shapes are stereotyped; speed varies but speed is not identifiable early - phase 30).

Idea 2 - deficit vs the longest-lived donor (age-aligned v_corr gap to a best-case reference curve from the top-3 longest-lived donors dead by origin; target: EARLY degradation sign on the healthy cohort): AUC on healthy pairs (v_corr > 2.55) 0.565 vs v_drop 0.795 and tau_projected 0.725 - worse than the existing own-baseline features exactly where it was supposed to help; all-pairs 0.812 (dominated); residual top-5 23 vs tau 36. Verdict: rejected. Mechanism: cross-device age alignment inherits per-device voltage offsets (own-baseline v_drop cancels them) and the ~190d-sd lifetime spread means same-age devices sit at wildly different life fractions on a flat plateau - the deficit is offset noise until the knee, where level features already fire.

Discovery (the useful residue): tau_projected as residual re-ranker. Ordering the champion's NON-swapped descent-band devices by tau_projected catches 24/36/44 of its 132 panel misses in top-3/5/10 - better than knn (16/27/41) and than logbag's own next-by-p42 (15/21/35). The linear logistic underweights its best timer exactly in the descent band where its misses live. At panel FP prices (~120/false swap) even 36% top-5 precision is EV-negative (+5 adds ~600 vs ~360 saved), so NOT a panel candidate. It IS the natural ordering for the static path's ALPHA_STATIC marginal inclusions (held-out FP ~20-40 -> strongly +EV if precision transfers): adjudicate in the fleet-holdout harness with the phase-28/29 label-noise check before any adoption. Logged as the open deployment-regime candidate.

Idea 3 - V-T coupling (rolling correlation/slope of voltage vs temperature, seasonality removed): rejected. Physics hypothesis: rising internal resistance couples voltage to temperature before the level moves. Four leak-free variants screened (exp_vt_screen.py): within-day slope bvt_60 (seasonality-proof by construction), its own-baseline change bvt_delta, within-day correlation cvt_60, and the literal detrended 90d rolling correlation rcorr_90. The physics is REAL but lagging: bvt_60 carries AUC 0.779 on all pairs - and is dominated by v_drop (0.888) and tau_projected (0.909) in every cell, including the healthy early-sign target (best variant bvt_delta 0.675 vs v_drop 0.795, beats it on 3/12 origins). Residual-miss test: 2-5 in top-10 vs tau_projected's 44. Twist: the correlation form is ANTI-predictive on healthy pairs (cvt_60 AUC 0.288, inverted 0.712) - on the healthy cohort the informative direction is V-T DE-coupling (erratic voltage), which is the already-adjudicated vstd_30/rng_30 plunge-precursor channel (phase 6/SilentFix, rejected). Same verdict shape as phase 25: the gradient exists, the voltage state has already priced it in.

Idea 4 - temperature std as plunge precursor: rejected (no signal). User recollection check: the recorded precursor was VOLTAGE std (vstd_30/rng_30, phase 6/SilentFix, tested and rejected); temperature std itself had never been tested. Three leak-free variants screened (/tmp/x27/tstd_screen.log): within-day temp std (30d median), its own-baseline delta, and detrended 30d rolling std of daily medians. All at or near chance (best cell: detrended rolling std on healthy pairs, AUC 0.628 vs v_drop 0.795); residual-miss test 3-8 in top-10 vs tau_projected's 44. Curiosity only: within-day temp std is INVERTED on healthy pairs (AUC 0.217, i.e., 0.783 flipped - flat temperature readings precede death), still below v_drop and consistent with a sensor-quality artifact of the sagging supply voltage rather than an independent channel.

Idea 5 - infer per-device load, improve the theoretical (coulomb) model: rejected at the physics level. Load IS partially inferable (reads/day 17-24 p10-p90, thermal exposure, within-day sag all vary), but lifetime variance decomposition on the 82 deaths (full-life averages; mechanistic study, not an origin feature) shows there is nothing for it to explain: install-date batch alone R^2 = 0.853 (30 batches; confirms the recorded ~90% claim), all four load proxies together R^2 = 0.102 alone and +0.04 incremental over batch-month; within-batch residual sd 90d with best proxy correlation r = +0.23 (t_mean) - perfect load inference would shrink the 90d within-batch spread by ~2d. Death timing in this fleet is set at manufacture (lot capacity), not by usage; the load-adjusted clock collapses to the batch clock, already rejected (phases 6/10). Sign note: warmer devices live LONGER (r = +0.20), opposite the self-discharge story - consistent with cold raising internal resistance and pulling the 2.40 crossing earlier.

Phase 32 - "is p42 dangerous?" - window-portability bug FIXED, miss-cost ranking rejected

Two user hypotheses about p42, both tested; one found a real catastrophic-tail bug, the other closed a channel.

(a) Window portability. Worry: public/private could use a window other than 42d. The evaluator honours planning_window_days per scenario, so this is directly measurable: score unchanged plans against 28/56/84d evaluators.

evaluator window	before fix	after fix
28d	728.5	728.5 (inert)
42d	1086.6	1086.6 (byte-identical)
56d	23581.9	1688.0
84d	(same failure mode)	3283.1

The model side degrades gracefully; the danger was in the OUTPUT convention. plan_tsroute parks non-included batteries at day 43 and relies on the evaluator cutting rows past the window - with W > 42 the evaluator EXECUTES all ~200 parked rows as real swaps and charges early penalties on the whole fleet. Fixed (repark_for_window, applied in all three plan paths): parked rows move to max(W, 42)+1 read from the live settings. Inert at W <= 42 by construction (all-48 checksum byte-exact). Residual rise at 56/84d is legitimate, not a bug: a longer window simply contains deaths a 42d-ahead planner never scheduled for.

(b) Miss-cost inclusion ranking (rejected). Hypothesis: late_swap is high because inclusion ranks by raw p42, which treats a day-5 death (430 if missed) like a day-40 death (80). Tested rank_score = p42 x E[late penalty if parked] on the static held-out path (same alpha=1.5 count, same planner, only the boundary ORDER changes), 3 fleet splits x 16 origins:

run	p42	miss-cost	delta
s0	925.9	894.9	-31.0
s0 reversed	778.6	815.1	+36.4
s1	914.0	898.7	-15.3

13W/12L/23T - no unanimity, rejected under the phase-26 standard. The mechanism does what it claims where it wins (late 462->439, 524->504) but the gains are small and reverse on one split: p42 and death-day are strongly correlated (a confident p42 usually means already-deep decline, i.e. dying soon), so the boundary is already nearly cost-ordered and the differences trade on day-quantile noise. Conclusion for the record: the residual late_swap is NOT mis-ordering among visible deaths - it is the voltage-silent deaths (channel closed, phases 6-13) plus label-lag noise.

Phase 33 - late-swap root cause: the count is optimal, ranking is the wall

Public 6ea03da = 2188.78 (vs 3c7b883 1661.45) triggered a full attack on late_swap. Six mechanisms tested, all rejected; the diagnosis is now closed and quantitative.

Forensics (panel s_8-41, alpha=0.8, full evaluator)

• late is 97.9% MISSES (31070) vs 2.1% caught-but-swapped-late (660 over 34 origins). We already swap a median 12.2 days BEFORE EOL, so the "shift swap days earlier" 20:1 lever DOES NOT EXIST.
• recall 176/308 = 57%. Caught deaths: max rank 20, median rank 5. Missed deaths: median rank 35 of ~450 - not marginal, not invisible.
• miss signature = LINGERERS: median days_below_245 63 vs 29.5 for caught; tau_projected 17.7 vs 6.4. They entered the 2.45 V zone long ago and decline slowly.
• marginal economics: a marginal swap costs ~54 early and catches 8% x 251 = EV 21. Break-even density is 54/251 = 21.5%. Ranks 11-12 run 32% (keep), 13-15 run 11.7%, 13-30 run 8% (reject).

The alpha question, settled on the SHIPPING path

Definitive curve (rolling native fit, panel, full cost):

alpha	0.6	0.8	1.0	1.2	1.5	2.0
swaps	13.6	15.2	17.4	19.4	22.8	28.5
late	1270	1019	944	866	817	691
early	330	444	604	758	1017	1409
total	1844	1679	1771	1862	2084	2369

Clean U with the minimum exactly at 0.8 (1.0 is better on 7 origins, worse on 24). More swaps DO cut misses - late falls 1019->691 - but early rises 3x faster. At 28.5 swaps our late (691) still never reaches the leader's 499 while our early (1409) is 2.3x theirs: there is no alpha that reaches the leader profile. Their 19 swaps are a CONSEQUENCE of better ranking (marginal hit rate ~40% vs our 8%), not a cause. Matches the public result (+8 swaps -> late -13, early +444) and the 0.46 -> 0.21 -> 0.03 marginal rates from phase 28's addendum.

Ranking attacks - all rejected, all with the same signature

Metric note (IMPORTANT): score WHO candidates on top-12 recall, never AUC. Every candidate below RAISED AUC while LOWERING the recall that converts to cost.

candidate	AUC	top-12 recall (base 175/308)	verdict
zone features (days_below_245, in-zone, zone rate)	+0.0061	167	rejected
v_corr_last + thresh_dist (the label IS the 2.40 crossing)	+0.0088	173	rejected
both	+0.0111	169	rejected
RF ranking (rank_score) + logbag count	+0.0140	161	rejected
rank-avg 50/50, 75/25 logbag	+0.014	172 / 176	noise (4W/3W)
cohort model (train only on at-risk)	+0.0163	170	rejected
cohort + zone	+0.0185	169	rejected
sub-daily/intra-day (daily min, dip depth, dip accel, min-slope, dispersion)	+0.0009	band hits 73->72	rejected
miss-cost ranking (p42 x E[late if parked])	-	13W/12L/23T	rejected
prevalence recalibration (floor count by recently RECORDED deaths)	-	corr -0.03..-0.18 vs sum(p42) +0.39..+0.52	rejected

Cohort diagnostic (why the AUC/top-k divergence): the at-risk cohort is 28.9% of devices and holds 99% of deaths (7.6% death rate inside vs 0.03% outside). A global model spends most capacity on the healthy-vs-sick boundary that AUC rewards and we never act on; but restricting to the cohort did NOT help either (170 vs 175), so the dilution hypothesis is refuted too.

Sanity check - ML ranking vs raw physics (top-12 of 308): logbag 175, own-trajectory time-to-2.40 119, thresh_dist 108, v_corr_last 108, master-curve tau 51. Our ranking is far better than physics; it is simply not good enough at ranks 13-40 to make extra swaps pay.

Conclusion: late_swap is bounded by WHO-ranking quality among lingerers, not by appetite, timing, features-on-hand, or model family. The count channel is closed on the panel AND on the public leaderboard.

Test-look ledger

8 looks total (now historical - see protocol change above); look 9 = the near-miss backfill table. Original ledger: 8 looks (threshold stack, topsum, ecroute+ablations, tsroute, aug-tsroute, +tau_projected, logbag+α-challenger, aux-features). Dev/test disagreed persistently by ~0.7σ (dev prefers richer stacks, test the simpler v1); selection followed dev + the combined mean, with the public leaderboard as final arbiter.

Phase 34 - cross-fleet CV produces its SECOND confident false positive

Four candidates confirmed on the real leak-free full-fleet all-48 walk-forward (/tmp/all48_rank.py, /tmp/all48.py shape). Base reproduced 1661.7 exactly, so the harness is sound.

candidate	all-48	delta	W/L/T	warm-up	panel	near-end
base (champion)	1661.7	-	-	2379.7	1679.2	604.8
rankavg (.75 logbag + .25 rf rank)	1717.0	+55.3	12/19/17	2610.2	1703.1	604.8
cohort_hybrid (at-risk cohort refit, gated)	1713.1	+51.5	13/20/15	2573.3	1706.4	604.8
both	1679.2	+17.5	16/23/9	2519.2	1670.7	607.4
shape features (sh_*)	1791.4	+129.7	-	2500.6	1830.0	627.1

rankavg had won cross-fleet CV at -18.4, 27/32 folds, fold-level p=0.0002, robust to outlier removal and surviving Bonferroni across 5 candidates. It is +55.3 worse on real data. This is not a warm-up-weighting artifact: it loses on the panel alone (+23.9), where models are well trained.

Second validity boundary of cross-fleet CV (the first was half-fleet appetite bias, which refuted ec_pmax=inf): every fold trains on the FULL history, so the harness structurally cannot reproduce cold-start scarcity. RF ranks well when well-trained and is noise at 3-20 deaths; blending it helps in a world where every model is well-trained, and hurts in ours (warm-up +230.5 of the +55.3).

Verdict: cross-fleet CV is retired as a selection instrument. Both effects it called significant failed to transfer. It over-rejects nothing and over-accepts anything that (a) shifts count/appetite or (b) benefits from a large train set - which is nearly every candidate worth testing. Screen on full-fleet all-48 only.

Also closed this phase

• Warm-up physics prior (/tmp/warmup_prior.py): at the cold origins s_0-7 logbag is already the best ranking source - top-16 recall 75 vs blend25phys 72, blend50 70, phys_ttc 66, master-curve tau 50. At s_8-11 logbag dominates outright (39 vs 31/30/29). No physics prior to graft on. Line closed.
• Optimizer budget (/tmp/a48opt.py): 2 restarts x 60 iters is at the optimum. 6x120 = +2.1 (13W/12L/23T), 12x200 = +6.7 (19W/20L/9T). More search buys nothing; the day-placement objective is not search-limited.
• Feature LOO at bags=120: best removal is -age_days at top12 177 vs 175 (+2/308, well inside noise); -v_last collapses it to 160, confirming v_last as the load-bearing feature. No removal adopted.

Note on regime weighting (not acted on)

Warm-up is 23.9% of the all-48 checksum, but submit.py fits held-out splits with a single static foreign model (line ~218) - those splits never see cold-start. The checksum therefore over-weights a regime the scored public and private splits do not have. Left as-is: the leak-free alternatives all reintroduce the half-fleet appetite bias, and every phase-34 candidate loses on the panel too, so nothing hinges on the weighting.

Phase 35 - full re-screen on the surviving instrument (72 configs)

Everything ever judged on a retired instrument (cross-fleet CV, the half-fleet static holdout, panel-only, top-12 recall) re-run on the leak-free full-fleet all-48 walk-forward. Base reproduced 1661.7 in every group. /tmp/screen48.py.

Nothing beat the champion. Summary of deltas (negative = better):

group	result
cliff (8 flavours)	count 1.5/2.2/3.0 = +19.0/+21.0/+3.0; rank-only +22.9; linger +9.5; p42-floor 0.25/0.50 = +40.3/+104.0; park-floor +102.5
appetite alpha	0.9 +67.8, 1.0 +113.8, 1.2 +228.2, 1.5 +479.1 (W3 L40)
ec_margin	0 +44.2, 0.4 +12.5, 0.5 -10.9, 0.6 -3.2, 0.7 -5.9, 0.75 -4.2, 0.9 +7.1, 1.0 +19.2, 1.2 +47.8
ec_pmax	10 +41.9, 60 +16.0, inf +28.8
route	sd0 +17.9, sd3 +16.3, sd14 +16.6, nn +0.6, emerg48 +6.0, emerg52 +21.3, ecroute 0.8/1.0 = +441.1/+436.3
floor	scale 1.5/2.0/3.0/3.5/4.5 = +105.8/+25.3/+39.8/+83.3/+172.5; gate 0/10/30/40 = +530.4/+158.2/+127.3/+150.8; nozone +113.3
daycurve	day_q and p_conf inert (T48, delta exactly 0); pbe0.3 +125.8, pbe0.5 +2.4, fday3 +7.7, fday7 +12.7
model	rf +151.3, C0.3 +48.5, C3 +60.4, bags200 +11.9
features	zone +104.9, vcorr +32.8, ts +116.3, shape +129.8

Cross-fleet CV: four false positives and one sign inversion

mechanism	cross-fleet said	all-48 says
ec_pmax=inf	-35.3, p=0.031	+28.8
rankavg	-18.4, p=0.0002, 27/32 folds	+55.3
ecroute1.0	-27.7 (862.1 vs 889.8)	+441.1
rf	-26.4 (863.4 vs 889.8) at 8.2 swaps vs 7.3	+151.3
alpha	monotone DOWN to 1.2 (805.4)	monotone UP (+479 at 1.5)

Every one is an appetite artifact of scoring half a fleet. Retired for good.

gbm is not merely bad, it is undeployable

HistGradientBoostingClassifier raises ValueError: window shape cannot be larger than input array shape binning the walk-forward training sets - at s_0-7 AND at s_8-47. It only ever ran under cross-fleet CV because every fold there trains on the full history. Shipping it would have been a hard crash.

The break-even top-up is a switch, not a dial

ec_margin/ec_pmax move only s_39-46; s_0-38 and s_47 are byte-identical. Cause: the gate is fp_price = EARLY*(e_far-42) <= ec_pmax with e_far = (device end_time - T) + 30, and end_time is one global constant (the dataset's last timestamp, 2026-08-01). Eligibility is therefore a step function of calendar distance to the data end:

scenarios	eligible
s_0-38	0-2.6%
s_39-46	96.1%
s_47	0% (e_far < 42, gate fails)

s_39 sits on the edge at fp_price 24.7 vs the 25.0 cap. Consequences:

1. The apparent ec_margin=0.5/0.7 wins (-10.9/-5.9) rest on 8 scenarios, one of which (s_40) swings +389/+463. The curve is non-monotone across 0.5->0.6->0.7. Noise, not an optimum. Not adopted.
2. park_mobility 1.0/2.0/3.0 are byte-identical and move exactly ONE scenario (s_39, -247). Single-scenario luck. Not adopted.
3. Deployment has the SAME geometry - our harness is faithful here. iterate_scenarios (batteryswap_public/utils.py:163) truncates the timeseries per scenario (flat.loc[:scenario_start]) but passes the locations table through untruncated, so end_time at runtime is the split's global data end, exactly as locally. And load_context (models_cv.py:1814) calls iterate_scenarios itself, so the all-48 harness reproduces the runtime geometry exactly. The switch therefore fires on the same ~8-of-48 tail fraction on a held-out split, assuming a similar scenario spread; it is NOT secretly active everywhere at runtime. (An earlier draft of this section claimed the opposite, reading the plan_tsroute comment's "widens automatically ... at runtime" as meaning every held-out scenario. It does not - only those within ~12-61 days of the split's data end.) Residual unknown: the held-out splits' scenario spread relative to their data end is not observable from here.

Latent hole: unobserved_eol_days is hardcoded

The evaluator imputes EOL = location end_time + settings.unobserved_eol_days (evaluate.py:291). models_cv hardcodes + 30.0 twice: cost_curves:1169 (the survivor branch of EVERY day-cost curve) and plan_tsroute:1451 (the top-up gate). submit.py wires EARLY/LATE per scenario and warns on planning_window_days, but not this.

Exposure measured (/tmp/unobs.py, planner-assumes-30 vs planner-wired, the evaluator using U in both arms; U=30 control reproduced 1661.7 with the arms byte-identical):

U	unwired	wired	gap	worst scenario gap
10	1623.9	1619.4	+4.5	+790
20	1642.7	1621.4	+21.3	+747
30	1661.7	1661.7	0.0	0
45	1699.6	1698.8	+0.9	+170
60	1737.6	1756.8	-19.2	+289

Wiring is a provable no-op at the default and helps in 3 of 4 off-default cases, but is 19.2 WORSE at U=60 - because ec_pmax=25 was itself tuned assuming 30, so correcting the price without retuning the cap just moves the switch. Not a free correctness win.

Risk assessment: all 48 local scenarios carry an explicit settings block, and the ONLY fields the organizers vary are base_location (16 values) and base_room (27). Every economic setting - window, penalties, unobserved_eol_days, limits, overtime - is constant. So the hole is real but unlikely to fire. Left unwired pending a decision; models_cv.py is hash-pinned (exp_x28.py PINS) and the fix is not a measured gain.

Phase 36 - the fit-protocol axis: a real warm-up mechanism that deployment never runs

Features, model families, hyperparameters and planner knobs are all swept out. The untouched axis was how the training set is built and weighted. Motivation: at s_0 the training set is ~10k rows holding 3-20 positives, each device contributes ~168 autocorrelated grid rows (so long-lived devices dominate by negative count), and tau_projected is entirely NaN there - the imputer silently drops it. /tmp/screenfit.py.

Ungated arms, full-fleet all-48 (base 1661.7):

arm	warm-up	panel	all-48
base	2379.7	1679.2	1661.7
quar (QUARANTINE on)	2202.4 (-177.3)	2172.7	+395.0
devw (device-equal weight)	2211.0 (-168.7)	2053.0	+254.7
bal (class_weight balanced)	2467.2	4709.4	+2442.9
recw180 / recw365	2465.6 / 2505.1	1748.4 / 1747.0	+51.5 / +64.8
thin2 / thin4	2401.4 / 2533.6	1697.4 / 1906.5	+20.3 / +191.1
logcal	2756.1	1772.6	+139.6
ddepth3	2499.7	1680.9	+16.7

Clean regime split: quar and devw help warm-up and wreck the panel. Mechanism is coherent - dropping the 26 "looks dead, labeled survives" ghosts and de-weighting long-lived devices removes label noise/imbalance that only bites when positives are scarce; with hundreds of positives those same rows are informative hard negatives.

Gate them the way the floor is gated (train_deaths < 20). Verified the gate splits exactly at s_0-7: gated arms leave panel (1679.2) and near-end (604.8) byte-identical.

gated arm	warm-up	all-48	delta	W/L/T
devw_g12 / g20 / g30	2318.1 / 2211.0 / 2211.0	1651.4 / 1633.5 / 1653.2	-10.3 / -28.1 / -8.4	6/2/40 at g20
quar_g12 / g20 / g30	2249.8 / 2202.4 / 2202.4	1640.0 / 1632.1 / 1663.9	-21.6 / -29.6 / +2.3	7/1/40 at g20
dq_g20 (both stacked)	2124.7	1619.1	-42.5	7/1/40
dq_g30	2124.7	1659.4	-2.3	7/5/36

Gate 30 leaks into the panel (1736.0) and destroys the gain; 20 is the optimum, matching FLOOR_GATE. Seed stability for the devw resample (/tmp/seedcheck.py, 6 seeds): warm-up 2081.2 +/- 36.0, all-48 delta -49.8 +/- 6.0, range [-58.8, -42.5]. Every seed improves, and the reported seed 0 is the worst of the six. The mechanism is real and not a resampling artifact.

...and it is inert in deployment

submit.py scores public,private only; both are held-out, so both take the static branch where train_uniq = static_uniq = unique dead devices in the full foreign train pool. That number is 82. Every gate tested (12/20/30/40) is far below it, so the branch never executes on a scored split.

This generalises beyond the candidate: the whole warm-up regime - 23.9% of the all-48 checksum - is structurally absent from the scored splits, and the existing cold-start floor has the same property. It retro-explains the record: the floor moved all-48 2146.9 -> 1851.0 (a huge local gain) while public did not move until phase 23's unmasked pooling landed independently.

Consequence for measurement: s_8-47 is the deployment-faithful window - the gate is shut there locally exactly as it is in deployment. Base on that metric is 1518.0. dq_g20 is byte-identical on it.

Not adopted: inert where it is scored, and it would reset the checksum baseline and break comparability with every prior phase.

Free re-analysis: all 86 screened configs re-ranked on s_8-47

If warm-up had been dragging a good config down, re-scoring on the deployment window would surface it. It does not. The only negatives are the same tail artifacts already rejected in phase 35 - ec0.5 -13.0 (still hostage to s_40 +389), ec0.7 -7.1, parkmob -6.2 (one scenario), ddepth3 -4.0 (21W/18L, noise). No hidden winner.

Phase 37 - can the phase-36 mechanism be made always-on? No.

dq_g20 is worth -49.8 on all-48 but never fires in deployment (static_uniq=82 vs gate 20). Two hypotheses for why the UNGATED version cost +374/+493 on the panel, both tested. /tmp/always.py.

H1: devw's harm was a double-bootstrap artifact - REFUTED

The phase-36 devw resampled with replacement and then handed the result to BaggingClassifier, which bootstraps again. Re-implemented with true sample_weight inside a manual 15-bag loop, with a uniform-weight control (mbag) to absorb the implementation difference (that control is itself +90.5, so comparisons are made against IT, not against base):

arm	vs base	vs mbag control
mbag (uniform)	+90.5	-
mbag_devw (1/n)	+479.7	+389
mbag_devsq (1/sqrt n)	+288.9	+198
mbag_dev25 (1/n^0.25)	+91.0	+0.5

Clean dose-response: the more device weighting, the worse. With the double bootstrap removed the harm is unchanged, so the implementation was not the cause. Weak weighting is simply inert.

H2: quar's threshold (2.48) is over-broad - REFUTED

The filter is meant to remove the ~26 "ghost" trajectories (below the 2.45 death threshold, no EOL recorded). At s_30 it actually drops:

v_thr	rows	devices
2.40	169 (0.3%)	24
2.45	642 (1.3%)	53
2.48	1066 (2.2%)	77

So 2.48 removes 3x the genuine ghost count - mostly real slow-decliners that survived. Tightening it toward the ghosts should have kept the cold-start gain and dropped the panel harm. It does not. On s_8-47 (base 1518.0):

v_thr	2.40	2.44	2.45	2.46	2.47	2.48	k=0.5	k=0.8
delta	+36.7	+163.0	+181.3	+198.9	+458.8	+509.5	+138.6	+78.3

Monotone, every value positive, and the limit of the trend is "do not filter" (base = 0). Softening (keeping 50%/80% of the poison rows instead of deleting them) walks the same curve back toward zero without crossing it. And at v_thr 2.40 the warm-up block is 2379.7 - identical to base - i.e. the tight filter that spares the panel also delivers no cold-start gain.

Conclusion

The benefit and the harm come from the SAME rows. Devices that look dead and survive are label noise when the model has 3-20 positives and informative hard negatives when it has 800+; the effect does not attenuate with data, it changes sign. No threshold, no soft-weighting and no sample-weight scheme separates the two. dq_g20 is intrinsically a cold-start mechanism and cannot be made always-on - which, given static_uniq=82, means it cannot be made to matter in deployment at all.

Phase 38 - the label process reverse-engineered (2026-08-21)

User question ("the voltage continues past the EOL line - if the battery fails we should not have voltage anymore") exposed a documented error and, chasing it, the exact rule that generates eol_times.csv.

Correction: post-EOL telemetry is the SAME battery, not a replacement

Phase 16 recorded: "All 82 dead slots keep reporting past EOL (mean 107d, max 305d): replacement batteries under the same device_id", marked human-verified. The durations are right (measured: median 85 d, mean 105, max 305). The interpretation is wrong. Median voltage tracked against days-since-EOL over all 82:

window	-90..-61	-60..-31	-30..-8	-7..0	+1..14	+15..30	+31..60	+91..300
median V	2.562	2.510	2.454	2.398	2.379	2.363	2.352	2.295

Smooth through the line, no discontinuity, and only 3 of 78 devices ever reach fresh-battery level (>2.9 V) afterwards. Phase 16's own named example d_57e99ff00802 reports to 2026-05-13 as claimed and never exceeds 2.559 V.

Genuine swaps DO exist and are unmistakable - 34 censored devices jump to

=2.95 V and hold (caveat 1 verified, ~30 claimed). They just never occur at recorded EOLs. Consequence: a feature window spanning a recorded EOL does not mix two batteries; segmenting history at recorded EOLs is unnecessary. The only genuine two-battery contamination is those 34 censored resets.

The rule: gated 10-30 C, threshold 2.40 V, sustained crossing

smooth_series's 10-30 C band is the ORGANISERS' default (batteryswap_public/utils.py:42) but is never called by the evaluator - we reimplemented it and inherited the constants. Never swept until now. Which (band, threshold) best reproduces the recorded EOL dates:

band	median lag at 2.40	coverage
0-50	6.0 d	100%
5-35	6.0 d	100%
8-32	4.0 d	100%
10-30	3.0 d	100%
12-28	3.0 d	98%
15-25	5.5 d	88%

2.42 is far worse everywhere (10-23 d); 2.38 loses coverage. (10-30, 2.40) is the unique pair with 100% coverage at minimum lag. The gate and the threshold are one definition, not two choices that happen to agree - a voltage reading is meaningless without a temperature. Independently, 10-30 is also the tightest band retaining all 82 deaths (12-28 drops to 77, 15-25 to 70) and peaks voltage AUC at 0.9604.

EOL marks the SUSTAINED crossing. Median lag 3 d (IQR 3-6), but 12 of 81 devices lag >21 d, max 163. Those are not measurement gaps: during the lag they have a median of 23 gated readings and 83% go back above 2.40 (mean 57% of readings above). They wobble. The lag is NOT predictable from origin-time features: RF leave-out Spearman +0.006 (p=0.95), R2 -0.193, and flagging the tail hits base rate exactly. Volatility runs the wrong way (vstd_30 rho -0.194). Characterisation, not a lever.

No calendar structure: EOL day-of-week is flat (weekend share 26% vs 29% chance), so no human/administrative recording process; month is confounded with fleet aging by window exposure; 68 distinct dates for 82 deaths is random collision. Calendar features rejected as a memorisation trap - fleet-history facts cannot transfer to splits at other dates.

Consequence for relabelling (user-asked): do not. The evaluator scores against eol_times.csv verbatim, the label is tight (100% within 7 d of a sub-2.40 gated reading, median 0), and the nearest relabelling - adding the ~26 ghost deaths - was measured harmful at six thresholds in phase 37. Residual error is entirely trajectory-forecast error with no label-process component, which retro-explains phase 17's label-arrival-hazard rejection: arrival is near-deterministic, so there was no hazard to model.

Artifacts: notebooks/all_batteries.ipynb (all 461 devices, categorised DEAD-SILENT 18 / DEAD-VISIBLE 64 / RESET 34 / GHOST 17 / ADMIN-GONE 67 / HEALTHY 261), notebooks/silent_deaths.ipynb.

Phase 39 - four hypotheses closed, four measurement laws (2026-08-21)

Five parallel sessions. Every line closed negative; the durable output is the laws.

hypothesis	verdict
install-batch death clock	CLOSED on 3 implementations. The famous "AUC 0.91 on silent pairs" rests on 4 positives (phase 10 already knew: all-covered 0.584, dark bucket 0.426-0.410 - worse than chance). Coverage was never the blocker: 83-95% monthly at late origins. Gated rank-only form measured +97.9 harm.
temp_mean (30 d, pre-filter)	Mechanism replicates (deaths 18.67 C vs survivors 20.18) but absorbed: -1.1pp recall@12 on production. Fifth temperature rejection; the beta correction already occupies the channel. Gains +6.4pp on a champion with no temperature handling.
battery-type clustering	Types are real (13x risk spread surviving level-removal) but miss rate is FLAT across types. Cannot buy recall.
dv42 voltage-drop regression	CLOSED on 4 implementations. Out-of-time R2 is negative everywhere (-0.117 / -0.029 / -0.172): the drop cannot be forecast. The score was a differently-parameterised linear classifier anchored on voltage - ties the incumbent at N=12. Censoring-aware variant exactly inert (3% of rows); quantile variant worst on a full panel. Phase 10's wall from the other side.

The four measurement laws

All the same shape: pick the denominator that matches the question.

1. Report recall at the operating N, not AUC. With ~400 negatives/origin, AUC 0.93-0.95 says nothing about a rule taking 12. Silent deaths measured AUC 0.947 ("p42's strongest region") and recall@12 0.000 of 24, median rank 60. Same fact, one currency.
2. A candidate must beat the incumbent ranker ON ORDERING before any architecture can make blending it safe. "Gated + rank-only" excludes calibration and count damage but NOT ordering damage: reordering a good list with a chance-level signal costs in proportion to the weight (+97.9 at w=0.05). No safe non-zero weight exists for a ~0.5-AUC feature.
3. Condition on the outcome, not row shares. Any partition (building, cohort, type, room) looks significant against row shares because missed deaths concentrate where deaths are. chi2 13.00 -> 2.67 on conditioning.
4. Always report the test panel. Requiring a recent in-band reading on test rows deleted 18% of deaths and swung v_now recall@12 from 0.800 to 0.382
   • larger than every modelling effect measured that day, and it reversed the ranking of every arm. A recency filter deletes the silent population by construction.

Caution on the blind-harness champion lineage

Two sessions independently found raw current voltage BEATS their champion's own ranker at the operating N (.721 vs .668; .703 vs .672). On production the reverse holds (.572 vs .429, +14.3pp). That lineage was ablated on COST under a threshold rule where every feature change moves depth as well as order, so a 4-feature set losing to last_voltage on ordering is more likely over-pruned than optimal. Do not port anything from it without re-running the ablation on recall at fixed N.

Phase 40 - the research loop's CV discipline, ported into this repo

Every prior phase re-derived its harness in /tmp and lost it. The three measurement disciplines the blind loop paid leaderboard points to learn are now repo-resident in batteryswap_example/cv.py, with the champion config itself promoted out of /tmp/all48.py into cv.CHAMPION.

venv/bin/python -m batteryswap_example.cv all48  [base|vcorr|zone|vrel]
venv/bin/python -m batteryswap_example.cv static [variant] [n_seeds]
venv/bin/python -m batteryswap_example.cv screen

1. Ranking precision at the operating N, reported beside the cost. models_cv.rank_stats now runs inside walk_forward, so dev/panel/test and the new protocols all emit rec12, rec20, operating-point recall/precision, the count actually bought, and the median rank of missed deaths. summarize() prints them, pooling caught/deaths over origins rather than averaging per-origin ratios (a 1-death origin must not outvote a 12-death one).

Base reproduces 1661.7 / 2379.7 / 1679.2 / 604.8 exactly, so the additions are diagnostic only. What they show that eight phases of cost-only reporting could not:

block	n	total	rec12	rec20	recall	prec	sched	median miss rank
warmup s_0-7	8	2379.7	0.526	0.735	0.763	0.570	19.8	30.5
panel s_8-41	34	1679.2	0.572	0.675	0.571	0.458	11.3	57.0
nearend s_42-47	6	604.8	0.837	0.935	0.893	0.250	16.7	50.8
ALL-48	48	1661.7	0.597	0.718	0.641	0.453	13.4	49.2

The panel's missed deaths sit at median rank 57, 45 ranks past the cut. That is the closed appetite axis stated as a measurement rather than as an argument: no value of N reaches rank 57, which is why +395 swaps on the real fleet caught zero. It also localises the warm-up floor - rec12 0.526 there but recall 0.763, i.e. the floor buys catches with depth (19.8 swaps), not with order.

2. screen: leave-BATTERY-out vs leave-SCENARIO-out, side by side. 66503 closed pairs, 461 batteries, 1130 positive rows from 80 batteries. The effective n is 80, not 1130.

model	battery AUC	battery rec12	scenario AUC	scenario rec12
logistic	0.9669	0.541	0.9716	0.639
logbag	0.9669	0.535	0.9718	0.641
rf	0.9745	0.564	0.9946	0.832
gbm	0.9749	0.540	0.9921	0.786

Scenario grouping inflates rf's margin over logistic from +0.008 to +0.023 AUC and from +0.023 to +0.194 rec12 - an 8x exaggeration of the tree advantage, because holding out a scenario leaves the same batteries in train and test. Levels are not comparable across columns (a battery fold ranks ~92 devices, a scenario fold ~430); the margin within a column is. Selection reads the battery column. This is the offline-screen counterpart of what all-48 already knew: the harness rejected rankavg (+55.3) after cross-fleet CV had confirmed it at -18.4, p=0.0002.

3. static: the deployment gate, made measurable. Fit ONE model on half the buildings' fully-closed pairs, plan every origin of the other half with it - no refits, no accumulating own history, cold-start floor off (82 foreign deaths keeps gate 20 permanently shut in deployment). That is submit.py's held-out branch. 11 s a seed, so variants are screened across 5 building splits, not one.

Its levels are NOT comparable to all48 (half the fleet, half the deaths, and the count rule scales with the fleet) and its appetite readings are known-bad - this construction manufactured the sum(p42) undercount that phase 20 refuted and phase 26 believed (public 1661 -> 2189). Ordering only, variant vs variant. Both facts are enforced in the printout.

What the two instruments said about three feature variants

variant	static gate (5 seeds, paired)	mechanism on the gate	all48	verdict
vrel (+vrel, vmin7_rank)	-24.6 +/- 9.3, W143/L66	caught 460 -> 463, sched 1176 -> 1060, rec12/rec20 flat	+138.7	reject
vcorr (+v_corr_last, thresh_dist)	-24.8 +/- 13.5, W115/L98	caught 460 -> 478, sched 1176 -> 1191, rec20 +0.007	+32.7	reject
zone (+z_days, z_in, z_rate)	+12.3, prec 0.391 -> 0.349	sched 4.9 -> 5.6	+104.9 (phase 34)	reject

Champion unchanged. Two things are worth keeping from the exercise.

vcorr's +32.7 reproduces the phase-34 measurement of +32.8 to within 0.1 - a third independent confirmation that the all-48 harness is deterministic and sound.

vrel is the discipline working. It won the gate by the widest margin of the three (W143/L66, 4 of 5 seeds) and it won it entirely through the count: 116 fewer swaps for 3 more deaths, with rec12 and rec20 flat to three decimals. The gate's own docstring says a count reading from this protocol is invalid, and the native path then charged +138.7 for it. Had the gate been read the way phase 26 read its ancestor, this would have shipped.

Latent deployment bug fixed: unobserved_eol_days

cost_curves and plan_tsroute hardcoded the evaluator's imputed-EOL offset as + 30.0 while the evaluator reads settings.unobserved_eol_days per-scenario - and submit.py imports plan_tsroute from this module, so the hardcode shipped. All 48 train scenarios use 30.0, which is precisely why no local protocol could ever catch it: the same shape as the hardcoded 42-day park that repark_for_window exists to undo. Now a module global that plan_tsroute overwrites from the settings it is handed; verified to propagate (30 -> 90) and to leave all-48 byte-identical at 1661.7.

Phase 41 - the phase-39 directive discharged: fleet-relative port REJECTED

Phase 39 left a standing rule: "Do not port anything from [the blind-harness champion] lineage without re-running the ablation on recall at fixed N." The vrel / vmin7_rank port in features.py was exactly the case it describes, and it was sitting half-landed - computed by features.py and shipped in the packaged parquet, absent from models_cv.FEATURES.

Why the earlier cost measurement did not discharge the directive. Under the topsum rule a feature moves p42, p42 moves sum(p42), and sum(p42) moves N. So a cost delta is the sum of an ordering change and a depth change and cannot separate them - which is how the blind lineage concluded a 4-feature set beat its own champion. cv.ablate holds N fixed and varies only the score. No planner and no evaluator are involved: recall at fixed N and the rank distribution of misses are functions of the score alone.

Deaths caught at fixed N, paired per origin, bootstrap 95% CI

variant	native N=12	native N=20	static N=12	static N=20
vrel alone	-6 [-12,-1]	+4 [-1,+9]	+6 [+0,+13]	-1 [-9,+7]
vmin7_rank alone	-1 [-5,+2]	+4 [+0,+9]	-5 [-11,+1]	+0 [-4,+4]
both	-7 [-15,+0]	+5 [-1,+11]	+4 [-2,+10]	-1 [-9,+7]
v_last alone	-39 [-57,-22]	-8 [-20,+5]	+7 [-10,+25]	+5 [-10,+20]

454 deaths over 48 native origins; 976 over 240 origin-seeds on the gate.

Verdict: reject. The sign flips with N on both instruments and every CI for the port straddles zero. The single CI that excludes zero is vrel alone, native, N=12 - and it is negative. Decisive corroboration: the ordering is byte-identical at 36-41 of 48 native origins and at 228-230 of 240 static ones. These features barely enter the ranking. The +138.7 cost penalty measured in phase 40 was therefore almost entirely a depth effect, exactly as the directive predicted, and the -24.6 the static gate liked was the same depth effect with the opposite sign.

The phase-39 claim reproduces - and it is protocol-specific, not over-pruning

v_last alone was the specific claim: two blind sessions found raw current voltage beating their champion's own ranker at the operating N.

• Static gate: v_last alone is statistically tied with the full 13-feature champion (+7 [-10,+25] at N=12), with a better mean rank (11.9 vs 13.3) and a much better missed-death rank (48 vs 65).
• Native: v_last alone is -39 [-57,-22] at N=12, rec@12 0.478 vs 0.564 - -8.6pp. Phase 39 measured +14.3pp on production; same sign, same magnitude class, independently reproduced.

So the lineage's finding was real for its protocol - one static fit on a foreign fleet - and does not survive the refit path. The cause is not an over-pruned feature set and not the threshold rule phase 39 suspected: the champion's 12 features beyond v_last are worth ~8.6pp recall@12 when the model is refit on the native fleet and worth approximately nothing to a single static fit on a foreign one.

That is a statement about training data, not about features, and it read at the time as the first independent argument for what phase 40 shipped: in-fleet pooling gives the static branch 164 own deaths against the foreign 82. The public leaderboard refuted that reading within the hour - see phase 42. Tripling the static fit's death count bought no recall at all, so training volume is not the binding constraint the argument assumed.

features.py keeps computing both columns (deleting them means regenerating the packaged parquet on submission day for ~1s inside a 360s build budget) but now carries the rejection and these numbers inline, so the port cannot be re-attempted by inspection.

Phase 42 - PUBLIC: in-fleet pooling is a wash (1665.89 vs 1661.45)

Submission d5b03a1 (phase 40, HELDOUT_POOL=1). This push also carried phases 34-37, which had never reached the remote - but they touch submit.py by one deleted line, so the held-out path's behaviour change is pooling.

	pooled (d5b03a1)	champion (3c7b883)	delta
total	1665.89	1661.45	+4.44
late_swap	1222.92	1212.29	+10.63 (+51 late-days)
early_swap	133.69	137.47	-3.78 (-363 early-days)
ops	309.28	311.69	-2.41
swaps	571	580	-9

All four inversions land on integers (571 swaps, 5870 late-days, 12834 early-days, 54 penalised days), so public is still 48 scenarios at default settings - which also confirms the phase-40 unobserved_eol_days fix was inert here, as predicted.

Verdict: a wash, marginally negative. Pooling made the model slightly less eager - 9 fewer swaps, 363 fewer early-days, 2.4 fewer ops hours - and paid 51 late-days for it. Net +4.4 on a 1661 baseline is 0.27%, inside any reasonable noise band, but the sign is wrong and there is no recall gain hiding in it.

The finding is not "pooling is neutral", it is where the neutrality comes from. The static branch went from 82 foreign deaths to 246 (82 foreign + 164 own, in-fleet, same-fleet, correctly embargoed). That is 3x the training signal, on the target fleet itself, and the ordering did not improve. Put beside phase 41 - where v_last alone ties the full 13-feature champion on a held-out fleet - the two results say the same thing from opposite directions:

On a fleet the model has not seen, the ranking is essentially determined by raw voltage, and neither richer features nor 3x more in-fleet training data moves it. The information needed to rank better is not in what we are feeding the model.

This closes the training-data axis the way phases 1/6/28 closed the appetite axis. The remaining late cost is the voltage-silent population (recall 0.000 at the operating N, median rank 60 of ~450), and it is invisible to every signal in the current feature space - which is exactly what the phase-38/39 silent- death campaign concluded before it was closed.

Standing champion remains 3c7b883 at 1661.45.

Phase 43 - innovation sweep: three new axes opened and closed, one leak found

Directive: challenge assumptions, look for exploits in the setup, do not stop early. Five things were tested; four are new.

(a) The setup leaks the answer (VERIFIED, NOT USED - see the caveat)

Scenarios are spaced exactly 7 days apart and the split ships one timeseries file. For 42 of 48 scenarios the entire 42-day death window is already present in that file; mean window coverage is 93.9%. Combined with the reverse-engineered label rule (gated 10-30 C daily median voltage sustained below 2.40 V), reading forward reconstructs the labels directly:

	recall	precision
future-voltage reconstruction (10-line rule, all 48)	0.841	0.874
production model at its operating point	0.641	0.453

382 of 454 in-window deaths identified with 55 false positives, day error median -3 d (the confirmation lag). A tuned version reaches the 81/82-exact level check_labels.py already demonstrates, i.e. approximately the oracle (186 vs our 1661.7).

This circumvents the sanctioned accessor. iterate_scenarios explicitly truncates (cut = flat.loc[:scenario_start]), so past-only is the intended contract and reading the raw file past the origin breaks it. Recorded because it bounds what the leaderboard can mean, not adopted. Our own path is clean: features_at_origin filters day <= origin (features.py:73) and every physics/beta path does the same.

(b) Late cost decomposed: 85% ranking, 15% timing

Of 454 deaths: 163 never swapped (ranking failure -> emergency queue), 95 swapped AFTER EOL (timing failure, 150.8/scenario, mean 7.6 d late), 196 swapped before EOL at a median 13 d early. So 84.7% of late cost is inclusion, 15.3% is placement. First time these have been separated.

(c) The emergency queue is convex - and unexploitable

From evaluate_plan: misses are swapped one per day in alphabetical battery order from end_day, each with a dedicated round trip and its own end_previous_day. A miss at queue position k therefore pays an extra 10k, and including a battery shifts every alphabetically-later miss one day earlier. That externality is real and unpriced - the true marginal cost of a miss is ~330, not the ~220 previously used, which moves break-even from 25% to ~16%.

Tested as a rank bonus. First form (additive) scored +1007: an additive bonus is just alphabetical sorting, because the shift only occurs if the battery actually dies. Corrected to the p42-conditional form p42 * (1 + qw * 10 * E[misses after b] / 54):

qw	0.25	1.0	3.0
all-48 delta	+60.0	+160.1	+275.1

Monotone loss, rejected. The mechanism is a ~9% pricing correction but it reorders by alphabet, a variable with zero predictive content - measurement law #2 again: ordering damage beats pricing gain at the boundary.

(d) Lead-time bias on day placement - real trade, break-even

(b) says the day model is systematically optimistic, and late/early trades at 20:1. Two implementations:

k days early	0	2	4	7	10
shift the OUTPUT	1661.7	+36.2	+33.6	+53.0	+87.9
shift the BELIEF (quantiles, re-solve)	1661.7	-12.0	-0.3	-9.7	-11.3

Shifting the output works on late (987 -> 878) but compresses all work into earlier weeks and costs +157 ops in limit penalties. Shifting the predicted death-day quantiles instead lets optimize_days re-solve with routing and the weekly limit intact: ops stays at 206-227 and late falls to 922-962. Net best is -12.0 at k=2, on 23W/25L - a coin flip, non-monotone across k. Rejected, but it establishes that day placement is already near-optimal: the 95 late-swapped catches cannot be recovered by earlier scheduling because the early penalty on the 196 correctly-timed ones plus the ops cost consumes the whole gain.

Standing after this sweep

Every axis now has a measurement against it: appetite (public), features (phase 41 fixed-N), training volume (phase 42 public), routing-aware inclusion (phases 7/22, "even zero-travel extras lose"), expected-cost ranking (phase 27b), queue pricing (c), day placement (d), silent deaths (phases 6-13, 38-39).

The honest residual: the leader is at 1477.48, so ~184 points are demonstrably extractable by someone. We cannot account for them from any mechanism measurable here, and the only path this sweep demonstrated is the one in (a), which we are not taking.

Phase 44 - why our late_swap is the field's highest (full leaderboard inversion)

The public board exposes all 14 teams' components. Inverting each team's late_swap through the emergency-queue cost model (each miss = one dedicated day at (24 + k) * 10, solved for k self-consistently) gives, per scenario:

team	swaps	misses	caught	precision
J2W	14.9	2.45	7.65	0.51
YassY	17.4	2.37	7.73	0.44
CarlAlbert	15.5	2.80	7.30	0.47
sackriti	14.5	3.21	6.89	0.48
us	12.1	4.69	5.41	0.45
kalsach	28.2	2.14	7.96	0.28

Deaths solve to ~10.1/scenario (local all-48 is 9.5 - good agreement).

Our per-swap precision, 0.45, is mid-pack in a field of 0.28-0.51. The ranking is not broadly worse. We simply buy the fewest swaps in the competition - 580, against 695 for the next-lowest and 1355 for the highest - so we catch 5.41 where everyone else catches 7.3-8.0, and each of the ~2 extra misses costs ~285.

Appetite is genuinely closed - now with numbers, not an argument

From phase 28's own measurement: +8.2 swaps/scenario moved early_swap +444, so a marginal false swap costs 54. Our ops (309 at 4.69 misses) against J2W's (252 at 2.45) prices the emergency day at 25 per miss, so a miss costs 260 + 25 = 285. Break-even is therefore 15.9%, and our local precision in ranks 13-20 is (0.705-0.564)*9.5/8 = 16.7%. The marginal pick sits exactly on break-even. Raising alpha cannot pay.

The cliff-contamination escape is also closed: cliff rows die at 39.4%, 17x the base rate, at 12.0 rows/scenario. They are good picks, so phase 28's "zero extra catches" was not an artifact of the boost's reordering.

The real defect, and why no local instrument can see it

Our ranks 13-20 are 16.7% precise locally and ~0% on public (phase 28: 395 extra swaps bought late_swap +13). J2W's 13th-15th picks run at 80%. The head transfers to a foreign fleet; the tail is noise.

Where the misses live (native, 454 deaths):

rank band	n	frac stale > 7d	median n_30
1-12 (head)	256	0.04	30
13-30 (tail)	104	0.03	30
31+	94	0.30	30
fleet baseline	-	0.16	-

Rank 31+ is the closed silent-death channel (staleness at 2x the base rate). The tail is fully-observed devices - 3% stale, 30 readings in 30 days - that we simply order below rank 12 and J2W orders above it.

Two things were tested against it and both failed:

• Fleet calibration shift. The half-fleet gate splits ONE fleet, so it has no calibration offset by construction. Injected one (+/-0.03, +/-0.06 V on the target fleet only, with tau moved along the curve at 572 d/V). Recall moved by at most 0.005 at every N. The reason is elementary and closes the hypothesis: a uniform offset shifts every device's linear predictor equally and cannot reorder a linear model at all. It also explains why vrel does nothing - it protects against the one perturbation that was already harmless.

• Regularisation. Effective n is ~80 death-batteries, not 66503 rows, so logbag's C=1.0 is tuned for a sample ~800x too large. Swept C over 100x:

  	C=1.0	C=0.3	C=0.1	C=0.03	C=0.01
  gate, tail gain 13-30	0.154	0.155	0.157	0.158	0.153
  native, tail gain 13-30	0.229	0.231	0.247	0.244	0.240
  native rec@12	0.564	0.562	0.551	0.557	0.555

  Flat on the gate. Natively it trades head for tail (C=0.3: +5 at N=16 and +6 at N=20, both 5W/0L, but -1 at N=12). At our operating N of ~12 it is negative. Rejected.

The blocker is instrumental. The gate reports a perfectly healthy tail - rec@12 0.766, tail gain 0.154 - while public reports ~0% marginal precision there. No local protocol reproduces the defect, so no local protocol can validate a fix for it. That is the honest reason this stops here rather than a claim that nothing is left: J2W demonstrably finds 2.2 deaths per scenario, in well-observed data, that our ordering puts below the cut.

Standing champion remains 3c7b883 at 1661.45.

Phase 45 - the objective axis: pairwise ranking loss REJECTED (last open axis)

Rationale: log-loss on a 2.3% positive rate spends its capacity on the 97% that obviously survive; the decision is a top-3% ordering, and the entire 638-point gap to the leader sits at ranks 12-20. A pairwise loss (RankNet- style, linear scorer, within-origin death-vs-survivor difference vectors, no intercept) optimises exactly that boundary. Two samplers: uniform survivors, and hard negatives (survivors the incumbent ranks highest). Scored on recall at fixed N, both instruments; score drops into rank_score so p42/count/timing would stay byte-identical if shipped.

variant	native rec@12	native tail 13-30	gate rec@12	gate tail
logbag (base)	0.564	0.229	0.766	0.154
pair-uniform	0.509	0.264	0.743	0.154
pair-hard	0.518	0.271	0.746	0.149

Paired at N=12: native -21 (4W/15L), gate -20 (19W/36L). Loses everywhere at the operating point.

The apparent tail win (0.271 vs 0.229 native) motivated a head/tail hybrid: logbag's top-K byte-identical, pairwise re-orders only below it. Result: the hybrid loses at every K on both instruments (gate K12: -18 at N=16, 8W/24L; native K12: -12 at N=16, 4W/12L). That exposes the tail win as an artifact: through rank 30 the two objectives catch the same deaths (native rec@30 0.793 vs 0.789, gate 0.920 vs 0.894) - pairwise does not discover new deaths in the tail, it demotes base's head deaths INTO the tail. When the hybrid keeps base's head and asks pairwise to improve the tail with its own picks, there is nothing there.

Conclusion: the pairwise scorer extracts no information the log-loss fit does not already have. Same features, same data, same deaths found, lower positions. Combined with phases 41 (features), 42 (training volume), 44 (regularisation, calibration shift), the information ceiling of this feature space is the champion's current ordering. Every axis is now measured:

axis	closed by
appetite / count	public (phase 28), break-even arithmetic (44)
features	fixed-N ablation (41)
training volume	public pooling (42)
regularisation	C sweep, flat/negative at N=12 (44)
calibration robustness	offset cannot reorder a linear model (44)
objective	this phase
timing / day placement	lead-time bias, coin flip (43)
queue pricing	alphabetical ordering damage (43)
routing-aware inclusion	phases 7/22
silent deaths	phases 6-13, 38-39

What would beat 1661 from here is information we do not have locally: either a signal J2W computes that is outside this feature space entirely, or the forward-reading of the split file that phase 43(a) verified and declined.

Standing champion: 3c7b883 at 1661.45. Deadline 2026-08-23.

Phase 47 - the blind champion itself, as a whole: REJECTED (lineage fully adjudicated)

(Numbered out of order: written before the phase-46 commit landed from the parallel session; the work predates it, the number does not.)

Question (user, 2026-08-22): the blind-repo multi-agent loop's shared champion (9a969a6a7dfe.py, blind full-40 1566.7) - reusable here? The headline number is not on our scale (its own 40-origin harness; ~55 of its margin was extra cut depth), and most pieces were already adjudicated: CV discipline ported (40), vrel/vmin7_rank as ADDITIONS rejected (41), in-fleet pooling a public wash (42), trip-merge already native as building batching + group-day search (phases 1/14). One piece had never been tested: the champion's exact 4-feature set as a replacement ranker - FEATURES=[last_voltage, vmin7, n_obs, vrel], mapped here to [v_last, v_min_7, n_30, vrel] (blind4 in cv.VARIANTS).

Fixed-N ablation: a real but small head-ordering gain

Paired vs base, deaths caught, bootstrap 95% CI. v_last_only cross-check reproduces phase 41 (-39 [-56,-21] native N=12 vs published -39 [-57,-22]).

variant	native N=8	native N=12	static N=8	static N=12
blind4	+11 [-1,+23]	+9 [-8,+26]	+27 [+7,+47]	+14 [-2,+31]
blind3 (no vrel)	+14	+10	+22	+12

One CI of eight excludes zero (static N=8), signs consistent at N<=12 on both instruments, and rec@30 / miss-rank get WORSE (57 -> 65 native): the small set trades deep tail for head. blind3 matching blind4 everywhere settles the phase-41 question from the other side: vrel is not load-bearing; the gain is the small set itself - 4 features against effective n ~80 death-batteries, where 13 features overfit slightly (phase 44's C-sweep said the same thing through the regularisation knob).

Full harness: the ordering gain does not survive the planner

all48	total	sched	caught/454	early	late
base	1661.7	13.4	291	473.1	987.5
blind4, alpha 0.8	1765.4	14.6	298	584.0	948.3
blind4, count-matched alpha 0.7342	1740.7	13.7	290	530.1	972.3

Raw: +103.8/origin - the 4-feature calibration inflates sum(p42), the exact depth confound phase 41 diagnosed, in mirror image. Count-matched: +79.1 [-65,+231], 22W/26L, catches identical (290 vs 291) - the +9 fixed-N catches vanish under the real inclusion rule (topsum + break-even top-up + floor is not top-N), and the whole residual is placement: +57 early-days/origin, because day placement integrates the p-curve and a 4-feature curve is cruder than a 13-feature one. Ordering is only half the score's job here; timing is the other half, and the small set is worse at it.

Verdict: reject; nothing in the blind lineage is left unported or untested. The blind champion's margin decomposes fully into (a) its own harness's depth and threshold tuning - non-transferable, appetite closed; (b) mechanisms we already have natively; (c) a head-ordering effect real at fixed N but worth zero catches and negative cost through this planner. Runnable via blind4 / blind3 in cv.VARIANTS.

Standing champion: 3c7b883 at 1661.45. Deadline 2026-08-23.

Phase 46 - the overfitting question, audited; and the certainty-first adoption

User hypothesis: the leaders are overfitting the public board. Audited with a 4-way measurement (certain-pool size, our capture of it, the fix on both instruments, luck bounds). Verdict: partially right, and the audit found a real local gain on the way.

How much of the board is luck

The board keeps each account's best public score, so it best-of-N selects every team's luckiest config. From our own per-origin variance (sd 838 over 48 origins -> sd of the 48-scenario mean = 121), the expected luck of keeping the best of N equal-skill submissions is 217/260/296/338/367 points for N=5/10/20/50/100.

• J2W's 210-point lead over rank 2 needs only N~4.5. The ordering WITHIN the leading pack is plausibly board luck.
• The pack's 500-point lead over us would need N5300. Not luck.

The pack's edge is real - and it is not legitimately reachable by us

The certain pool (devices whose gated daily median has sustained < 2.40 V for 3 consecutive gated readings BEFORE the origin, i.e. the label rule already fired but the EOL has not landed - the wobble/lag population):

rule at origin	flagged/scen	die in-window	die by +90d	never die
2.40 sustained-3 confirmed	1.06	94.4%	100%	0
2.45 zone (days_below_245>=0)	21.8	24.6%	-	62%

So the free-certainty pool is ~1/scenario, of which our champion already ranks 42/51 in the top 12. The leaders' ~2.2 extra catches/scenario cannot be built from any past-only certainty rule we can construct: the implied J2W marginal precision is 0.69-0.88 (robust across all 9 inversion assumptions), on picks that for us measure ~0% transfer precision. Either they hold a tail-ranking signal outside this feature space entirely, or they are reading outcomes (the phase-43a forward-read, whose ceiling - 84% recall, 1.6 misses/scen - they sit comfortably below at 2.45 misses; or battery-toggling against public, ~1 battery learned per submission pair). Given phases 41-45 measured every legitimate axis at ~zero, the balance of evidence is that the top of the board is exploiting the setup, not merely lucky. Unprovable from here; recorded as an assessment, not a fact.

ADOPTED: certainty-first ordering (cert-first, k=3)

The audit's by-product. Force confirmed-crossing devices to the top of the ranking (rank_score = 2 + p42; count and timing untouched):

	all-48	panel	W/L/T	mechanism
first-reading rule (LEAKS 2 future readings)	-88.2	-124.5	6/0/42	struck
confirmed rule (past-only)	-37.9	-53.6	3/0/45	adopted
k=1 / k=2 confirmation	+290.7 / +199.5	-	9/30 / 9/31	recoveries flood the head

The confirmation depth is load-bearing and equals the label rule's own sustain parameter - k=3 is the unique safe point. Fixed-N instruments agree: native +9 catches at N=12 (6W/0L), deployment gate +6 (6W/0L, ties elsewhere) - loss-free on every instrument at every N tested. The wins are label-lag devices our model under-ranks BECAUSE they recovered (v_last back up) - exactly the wobble population phase 39 characterised.

Deployment note: at runtime iterate_scenarios hands truncated data, so the rule is past-only by construction there; the local strict measurement is the honest predictor. Wired into submit.py (BATTERYSWAP_CERT_FIRST, default on, both branches; stacks on top of cliff's rank_score if that is ever enabled; p42_sum untouched so the count rule is byte-identical).

Local champion is now cert-first at 1623.7 (from 1661.7). Public look not yet spent; submission is the user's call. Expected public value if the wobble rate transfers: ~-35 +/- (fires 0.46 devices/scenario locally).

Phase 49 - four new ordering-signal families opened and closed in one sweep

Directive (user, 2026-08-22): test new ideas against the champion, be innovative, try brand-new approaches. Everything below is a fixed-N ordering screen (cv.ablate discipline, both instruments where applicable, paired per origin, bootstrap 95% CI), so no depth confound; costs on the 1661.7 scale (cv.run_all48 does not route through submit.py's cert-first). Screen code: exp_x29_new_signal_sweep.py [long|tau|tavg]. Phase 48 is reserved for the cert-first public adjudication (parallel session).

(A) Longitudinal cross-origin features - the first untouched signal family

The feature table is a 168-origin snapshot grid (~3-4d apart) per device; every existing feature reads one 30/90d window at ONE origin, and no phase had ever read ACROSS origins. Three features, all strictly backward-looking (merge_asof, snapshots < T): mom28/mom56 = within-fleet rank change of v_last vs 28/56d ago; wolf = count of snapshots >= 42d ago in the bottom decile (device provably survived - it is alive at T).

The raw signal is REAL: bottom-decile chronics (wolf > 2) die at 20.7% vs 15.5% for newcomers, and dying devices' fleet rank slides (median mom28 -0.025 vs -0.007 for survivors). The model cannot use it:

variant	native N=12	gate N=12	verdict
mom	-5 [-13,+3]	-7 [-14,+0]	reject (gate N=16/20 significantly negative)
wolf	-24 [-42,-7]	-14 [-27,-1]	reject
momwolf	-27 [-47,-7]	-15 [-30,-1]	reject

Mechanism: the 30d-window slope/drop/accel features already price the trajectory in; the cross-origin rank version duplicates them with extra noise, and wolf duplicates the level (chronics ARE low). Effective n ~80 punishes every redundant column.

(B) Interaction terms - the linear champion offered its first products

z(v_last)*z(slope_30), z(v_last)*z(vstd_30), z(v_min_7)*z(accel) (within-origin z-scores, leak-free). Flat everywhere: native +2/-2/+5/+2 across N, gate +0/-1/-5/-1, every CI straddling zero. There is no exploitable interaction structure - the phase-45 ceiling holds against nonlinearity in feature space, not just against nonlinear learners.

(C) tau descent-band re-rank - phase 43's open candidate, closed

The knn screen left one open item: logbag underweights tau_projected in the descent band where its misses live (24/36/44 of 132 panel misses in tau's top-3/5/10); EV-negative as extra swaps, never tested as a pure REORDER at fixed N. Tested three transforms (band rank-blend, global rank-blend, promotion-only): native -45 of 454 at N=12 (7W/29L), -53 at N=8. The gate is mildly positive (+3..+11) - the same protocol-specificity signature as v_last in phase 41: a single static foreign fit underuses the timer, the weekly refit already extracts it. The gate's sign must never override a native loss. Candidate closed for the native path.

(D) Temporal score ensembling - the variance question, settled directly

Rank-average the current origin's fit with the previous 1-2 origins' fits applied to the same rows (adjacent origins are near-identical draws of the ~80-death training set). Dead flat: tavg2 -3 at N=12 (0W/3L), tavg3 -4 (1W/5L) - refit-to-refit orderings barely differ. With phase 44's flat C sweep (regularisation not binding) and logbag never separating from plain logistic (bagging variance not binding), this is the third independent leg:

The marginal-rank defect is missing information, not estimation noise. Any future ordering candidate must bring a NEW signal, not a better estimate of the current ones.

(It also moots the deployment-gate objection - there is no fit noise to smooth, in any branch.)

Standing

Four families opened, four closed, all with mechanisms. The information ceiling of phase 45 now covers: re-weightings (45), replacement sets (47), cross-origin dynamics, interactions, timer blends, and temporal ensembles (49). What remains is what phase 44 said: a signal outside this feature space, or the declined forward-read. Local champion unchanged (cert-first 1623.7 / base 1661.7); public champion 3c7b883 at 1661.45. Deadline 2026-08-23.

Phase 46 addendum: the rule is one point wide in BOTH dimensions

Threshold micro-sweep at k=3 (cost harness): 2.40 -> -37.9 (3W/0L); 2.41 -> +93.6 (8W/32L); 2.42 -> +383.9; 2.43 -> +314.1. With the earlier k-sweep (k=1 +291, k=2 +200), cert-first works ONLY at the label process's exact parameters (2.40 V, 3 consecutive gated readings). The mechanism is label-rule replay on past data, not prediction; any deviation reverts to guessing at sub-break-even precision. Direction closed.

Phase 48 - PUBLIC REFUTES cert-first (1704.64); the third and sharpest CV failure

Submission 9438aac (cert-first isolated, pooling default off): 1704.64 vs champion 1661.45, +43.19. Inversion (all integers - settings unchanged):

	swaps	late-days	early-days	total
champion 3c7b883	580	5819	13197	1661.45
cert-first 9438aac	598	5894	15203	1704.64
delta	+18	+75	+2006	+43.19

+18 swaps at ~111 early-days each. The diagnosis writes itself: on train, every device that sustains <2.40 V dies within 90 days (54/54, phase 46), so the boost was loss-free locally. The public fleet contains devices that sustain-cross and NEVER receive an EOL label - the ghost population phase 15 found 26 of in train (at 2.45-2.48; train has none at 2.40, public evidently does). A ghost is force-swapped in every scenario it survives, and with no EOL on file the evaluator prices each swap against end_time+30: ~111 early-days apiece. The count rule could not protect us - a boosted device displaces a marginal pick, and when the displaced pick sits in the break-even top-up set it re-enters, so the union grows.

The premise that failed is not the label rule (81/82 exact) but its CONVERSE on an unseen fleet: "crossed implies labeled-dead-soon" is a train-fleet fact, not a label-process fact.

What this makes three of

Locally-validated interventions refuted by public: ALPHA_STATIC=1.5 (holdout -240 -> public +527), cliff boost (adjudicated -55 -> refuted stacked), and now cert-first (-37.9 at 3W/0L/45T -> public +43.2). Cert-first is the sharpest of the three: mechanistically derived, past-only, loss-free on every local instrument, refuted anyway - by a fleet-composition fact no local protocol can exhibit because train contains zero counterexamples. When the local evidence for a class of intervention is this good and it still fails, the class is uninsurable, not under-validated.

Action: CERT_FIRST default off (code retained behind the env var); champion remains 3c7b883 at 1661.45 on the board (best-per-account row unaffected). A recency-guarded variant (trust the crossing only while fresh) is measured locally in this phase's sweep, but shipping it would repeat the same bet with a smaller stake and no way to price it - the ghost rate at 2.40 on the scored fleets is unobservable from here.

Phase 48 addendum: the guard is measured, and the late component kills it

Recency-guarded sweep (flag only crossings confirmed within the last X days): g14 -18.7 (2W/0L), g28/g42/gInf all -37.9 identical - locally every catch happens within 28d of confirmation, so a 28d guard is free AND caps a public ghost at ~4 scenarios. But the public inversion settles whether to ship it: late-days went UP (+75). If cert-first had caught even one real wobbler death per few scenarios on public, late would have dropped by ~240 days per catch. It caught ZERO. The local -37.9 was wobbler catches that the public fleet does not offer below the cut; a guard cannot create catches that are not there. Cert-first family closed entirely - the label-lag population either does not exist on the scored fleets or is already inside the head.

Phase 52 - building-grouped CV adopted; the contamination gradient quantified

Trigger: a sweep of every public competitor artifact (HF repos, hidden HF branches, GitHub repos, and deleted-HF-repo cards recovered via the 686f6c61/radar-huggingface-dataset archive) established that the scored public/private fleets hold FOREIGN BUILDINGS: the Attention-Heads card states it outright, and the FAQ transcription in ziadnasif77-debug's repo puts the full pool at ~200 buildings vs train's 24. Both serious hidden teams (YassY-The-AlchemYst's RobustHazardPlanner lineage, ziad) validate and calibrate building-grouped. Organizer-shaped conclusion, user directive: the trusted offline fold must be leave-BUILDING-out.

Change (cv.py): screen gains a third grouping column, whole buildings assigned to 5 folds greedily balanced by unique observed deaths (deaths are lumpy - the top building holds 15/82, nine hold exactly one; random building assignment starves folds). static/ablate already split by building (fleet_halves), so this closes the last battery-grouped selection surface. Selection guidance in the module now reads: building > battery > scenario.

Result, 3 fold-assignment seeds - rec12 margin over logistic (AUC margins agree in sign; levels are NOT comparable across columns, margins are the decision quantity):

model	scenario	battery (s0/s1/s2)	building (s0/s1/s2)
logbag	+.003	-.006 / -.004 / +.001	-.002 / +.000 / +.000
rf	+.194	+.023 / +.031 / -.008	+.005 / -.005 / -.014
gbm	+.148	-.000 / -.016 / -.026	-.022 / -.039 / -.030

The contamination gradient is monotone: rf's +.194 phantom margin on scenario folds collapses to ~+.02 on battery folds and to zero-or-negative on building folds; gbm goes outright negative. The residual rf margin in the battery column (sign-unstable across seeds) is sibling leakage - a held-out device's building/batch mates stay in train. logbag == logistic (margin ~0) in every column and every seed: the champion's model choice never rested on fold luck, and the deployment-shaped fold CONFIRMS it rather than reversing anything. No selection changes; the protocol does.

Standing rule upgrade (supersedes the blind-loop "group by battery" rule): group by BUILDING when the deployment is cross-building. Battery grouping is necessary but not sufficient; it still leaks the environment/ install-batch channel that phase 10/13 closed on the feature side.

Phase 52 addendum - the building split extended to the COST harness (all48oof)

Directive (user, 2026-08-22): "building split not only for the AUC but for the scenario split as well." New protocol cv all48oof: the walk-forward cost harness with BUILDING-CLEAN predictions - each device's p42 at every origin comes from a champion-config logbag fitted on pairs from every building except its own (5 greedy death-balanced folds, OOF-stitched), then the FULL fleet is planned as usual (planning must stay full-fleet: the count rule scales with fleet size - the phase-26 static-gate lesson). Paired self arm through identical machinery, cold-start floor OFF in both, so the per-origin delta isolates the self-building fit effect. Read the DELTA, not the level. Validity check: the self arm reproduces the canonical panel 1679.2 and near-end 604.8 exactly (floor only fires in warm-up).

block	self	oof	delta/origin	oof-better W / self-better L
warmup s_0-7	4256.1	4180.0	-76.1	6W / 2L
panel s_8-41	1679.2	1813.6	+134.4	8W / 26L
nearend s_42-47	604.8	757.3	+152.5	1W / 5L
ALL-48	1974.4	2075.9	+101.6	15W / 33L

Ranking moves with the cost: panel rec12 0.572 -> 0.538, recall 0.571 -> 0.542, ALL-48 rec12 0.597 -> 0.560. Warm-up's negative delta is noise (floor off, 3-20 train deaths). Reading: self-building fitting flatters the walk-forward by ~+134/origin on the panel; the transfer-shaped cost estimate for the champion on foreign buildings is the oof column. This is the cost-denominated size of the contamination the phase-52 screen showed as AUC margins - it prices the OPTIMISM of the harness, not a defect of the champion (both arms use the same model family; there is no building-free challenger that closes the gap, per the phase-52 screen). No selection change. Results /tmp/cv_all48oof.csv, seed 0.

Phase 50 - the forecast axis: fleet-learned conditional dynamics (construction and screens)

Directive (user, 2026-08-22): innovate on the FORECAST part; reinforced with "benefit from more training data" and "a forecast as a model ensemble". The phase-49 wall requires a NEW signal; this candidate brings one by construction - auxiliary supervision on 70k UNLABELED device-week transitions from the 168-origin grid, so the effective-n80 wall does not apply. Machinery: exp_x30_dynamics_forecast.py; session-55 built an independent implementation in parallel (HGBR rate model, refit cadence 8, MC 120x6, generation-time ghost guard; artifacts /tmp/dyn/).

Construction. One-step (7d) voltage forecaster - HGBR conditional drift on state + HGBR sigma + standardized empirical residual pool (plunge tails survive sampling) - rolled 6 weekly steps, 200 paths; p_dyn = P(cross 2.40 V within 42d). Dynamics AND residuals refit per origin on transitions ending <= T (corpus 22k -> 64k across the walk). Differs from rejected WienerFHT (phase 9): fleet-pooled nonlinear conditional drift, not per-device linear drift; fixed-N delivery, so calibration never enters. GHOST GUARD from the phase-48 lesson: below-2.40 devices keep base rank/probability - NOT a local no-op (342 below-threshold rows at scenario origins, 236 never-labeled).

Screens - the strongest fixed-N profile of the campaign (paired, bootstrap 95% CI, ghost-guarded): native +11 at N=16 (13W/6L), +13 [+5,+22] at N=20 (12W/1L); gate +24 [+11,+38] at N=8, +23 [+11,+36] at N=12 (28W/7L). Weight sweep w=0.1..0.5 positive in 39/40 cells - plateau, not knife edge. Tail diagnostic: 30 of 104 champion misses at ranks 13-30 promoted into p_dyn's top-12. Fuel: descent-band AUC 0.8005 vs v_last 0.8024 (891 positives), Spearman(p_dyn, v_last) -0.683. Runtime: 4.0s per split at 2x fleet vs the 1800s budget. Checks passed: no leakage (per-origin refits), weight robustness, runtime. What follows is why none of it survived.

Phase 51 - the dynamics family: CLOSED (eight deliveries, two implementations, and the building-split autopsy)

Cost harness, all arms, 1661.7 scale (count pinned where stated):

delivery	total	caught	mechanism of failure
rank_score blend37	1744.5 (+83)	-	inclusion reordered; placement/pricing believe base p42
global p42 quantile-blend, pinned	1717.1 (+55)	287	every device's day placement perturbed
band-limited (ranks 9-30), pinned	1690.8 (+29)	290	catch parity; +13 late of timing noise
B: p_dyn as 14th feature, retrain (55)	1710.1 (+48)	295	sched 13.4->13.9 - appetite in disguise
B pinned, 55's table	1707.4 (+46)	288	count pinned -> catches DROP below base
B on c6 table	1702.5 (+41)	288	same
B on c6 table, pinned	1706.8 (+45)	286	same
rank-mean ensemble	moot	-	both parents dead on both tables

Post-hoc blends fail as a class (beliefs inconsistent with count/timing machinery); retrains fail as a class (the only "conversion" was calibration inflation that pinning exposed as appetite).

Cross-implementation check. Two independent implementations agree at Spearman only 0.369 on above-threshold nonzero rows; on max-disagreement rows the ground-truth referee splits (0.680 vs 0.628 - both mediocre). The MC tail probability is implementation-fragile; only the coarse ordering is robust.

The building-split autopsy (user question: "are you using the new building-split CV here?"). The gate's logbag was always building-split, but the DYNAMICS corpus pooled both fleets' transitions - deployment-legal (label-free; the scored split ships its own telemetry), yet untested under phase 52's trusted fold. Refit on FOREIGN-building transitions only:

gate blend37	N=8	N=12	N=16	N=20
pooled dynamics	+24 [+11,+38]	+23 [+11,+36]	+7	+5
foreign-only dynamics	-14 [-30,+2]	+1 [-13,+15]	-12 [-25,+1]	-10 [-22,+1]

The entire screen gain was own-fleet telemetry memorization: each device contributes ~160 near-duplicate transitions, HGBR memorizes device-specific drift signatures, and p_dyn partially encodes "this device has been declining" - which the level features already price. This is the phase-40 CV-contamination lesson recurring one level down, in a LABEL-FREE model: grouping discipline applies to self-supervised corpora too, not just label fits. It also explains the implementation fragility (different configurations memorize different signatures - fine values diverge, real descent agrees) and why the harness never converted what the screens promised.

Verdict: family closed for this competition. No submission. The signal looked real (band AUC 0.80, -0.68 to v_last) but the transferable component is ~zero under the building fold; the apparent independence was memorized device identity. Corollaries, binding on the next candidate: (a) the phase-52 building fold applies to auxiliary/self-supervised training corpora, not only label models; (b) a screen gain that the cost harness cannot convert is now a two-time pattern (47, 51) - suspect the screen's population before suspecting the delivery.

Standing: local champion base 1661.7 (cert-first off), public champion 3c7b883 at 1661.45. Deadline 2026-08-23.

Phase 53 - the user's re-adjudication directive: promising ideas vs the building-split CV

Directive: retry the most promising rejected ideas against the new building-split instruments. Inventory first: everything from phases 41-51 had ALREADY faced a building-split instrument (the gate splits by building since phase 40), so the genuinely untested cells were the phase-15 ghost mechanisms (rejected pre-gate, on the native panel only) and the feature variants through the Phase-52 building screen column. Four parallel measurements, base reproduced at 1661.7 in every arm.

candidate	cost harness (ALL-48)	building gate	verdict
GHOST_ZERO	+472.7 (10W/37L; late +573, early -109)	rec@20 0.866 -> 0.681, -180 caught (4W/136L)	REJECT
QUARANTINE	+395.0 +/- 68.2 (13W/35L; late -189, early +539)	flat: -10/-6/-10/+4/-4, all noise	REJECT
both	+682.9 (9W/39L)	rec@20 -> 0.674, -187 caught	REJECT
vcorr / zone / vrel	(native refs +32.7/+104.9/+138.7)	screen margins negative at rec@12 in nearly every fold draw	CONFIRMED REJECTED

Two mechanisms worth the record:

• GHOST_ZERO's mask is not ghost-specific. days_below_245 is sticky, so the zeroed pool GROWS across the walk (2-6 devices early, 33-42 by s_24+, mean 22.4/scenario) and ~12% of it (128 devices over 48 scenarios) are real in-window deaths - each becomes a guaranteed miss. The foreign-fleet hope is quantified away: a target fleet would need a ghost-to-death ratio in the below-2.45 pool far beyond anything the building split simulates before the ~111-early-days-per-avoided-force-swap saving covers the misses.
• QUARANTINE's "label noise" is calibration ballast. Dropping the looks-dead-labeled-alive negatives barely changes ordering (p42 Spearman 0.9905 at s_20) but nearly DOUBLES the p42 level (0.0187 -> 0.0345): those mislabeled rows are the only counter-evidence that plunge signatures can survive, and removing them inflates every probability the count rule and the ec-margin logic consume. Late improves (-189), early explodes (+539).

The building screen column itself AGREED with the native cost harness on all three feature variants - tiny uniformly-positive AUC margins with negative operating-point recall, the familiar mid-list-reshuffle signature. The transfer-shaped fold offers no grounds to revisit anything.

The re-adjudication is complete and the record is now closed on every axis under the deployment-shaped instruments. Champion 3c7b883 / current head (static @ 0.8, cert and pooling off) stands. Endgame per the user's mechanics (private = user-selected past versions): score the current head publicly; if ~1661, select the head for private (robustness-hardened); else select 3c7b883.

Phase 54 - full re-adjudication complete + the CV settles (user-driven)

(a) The exhaustive re-try of every rejected idea (user directive)

10-agent sweep over the entire EXPERIMENTS.md candidate history on the then-current instrument pair, base reproduced at 1661.65 in every arm. All planners (14 arms, +343..+954, all REJECT), all model families (4 slices, ~50 arms, zero ALIVE - logbag==logistic now confirmed on four instruments), all feature re-adds (calendar/vproj/ranks/fht_days/temp30/dbz/stale2 - REJECT; temp_mean upgraded from 'absorbed' to actively harmful +54.5), timing target days-to-crossing (+35.2 REJECT, with the label-process refinement that the administrative EOL is stamped ~AT the confirmation day, not +3d after), and the 87-config timing-knob grid whose single nominal survivor cfg59 (ec_margin 0.6 / emergency_day 48, -6.61 +/- 3.10, 6W/0L/42T) was killed on the building-OOF surface (+14.97 +/- 22.05) - textbook best-of-87 selection noise with the cert-first W/T profile. The peer session's parallel building-OOF candidate sweep confirmed 23 of 24 rejections and parked its one coin-flip survivor (rankavg25fht: wf-OOF panel -48.6 at 17W/16L, self-fit +17.8 with catches -2, static multi-seed never run).

(b) THE SCORE - the CV settles (user decisions, in order)

1. Building split is the score, always (not the self-fit 1661.7).
2. Foreign buildings' FULL history (before and after the origin) is legal training data - deployment itself hands the model future-of-origin foreign-building knowledge, and it removes the cold-start distortion.
3. Per-building planning (cv lobo) was specified, built, and superseded the same day: it breaks the count rule by construction (129 zero-swap-death building-scenarios, recall 0.367 vs 0.542; per-device thresholds fix the recall - 0.766 at p*=0.10 - confirming the break was count granularity, not prediction). Kept as a diagnostic.
4. Final protocol: EXACT leave-one-building-out predictions (24 models, each trained on the other 23 buildings' full closed history), ONE full-fleet plan per scenario, single value = ALL-48 mean.

Champion on the final CV: cv all48oof SCORE = 1877.3 (72 s per run; 5-fold approximation was 1913.3; panel 1857.7). Reference chain for level sanity: self-fit 1661.7 < exact-LOBO 1877.3 < 5-fold-OOF 1913.3 - more foreign training data helps, self-building identity flatters by ~200.

Everything is judged against 1877.3 paired per scenario from here, sched-flat rule in force. Public best remains 3c7b883 at 1661.45; private selection recommendation unchanged (score the hardened head publicly, then select).

Phase 55 - THE SCORE final: temporal-purge CV, sole validator (user directive)

The CV settles for good, three user decisions in one afternoon session arc:

1. Protocol (implemented by session 59): for each scenario, ONE model fitted on ALL buildings' pairs whose 42d label windows lie fully outside the scenario window (origin+42 <= T or origin >= T+42 - label-overlap purge), whole-fleet prediction, ONE full-fleet plan, cost per scenario. SCORE = ALL-48 mean, floor off. Rationale (user): admits every forecasting method (no fold plumbing), one-sentence protocol, no leak within the scenario, and the temporal shape matches deployment (training data spans before and after most scored origins - phase 43a measured 93.9% window coverage). The building axis is deliberately not validated (user judgment: same battery type across buildings).
2. Sole validator (user, verbatim intent): "we only use purge score going forward, nothing else to validate a new model or optimization method." Documented as standing rule. LOBO's chimera-plan critique (24 stitched calibrations, no per-building cost attribution, seams inflate the transfer penalty) was the user's decisive argument against building-holdout cost.
3. Consolidation: cv.py now contains ONLY the purge harness + oofcand/oofreport. all48/static/screen/lobo/ablate/wf/self removed; recover from git 2000281. exp_x29/exp_x30 no longer run as-is.

Champion on THE SCORE: 1972.0 (warmup 3560.3 / panel 1839.6 / nearend 604.8). Level chain, never cross-compare: self-fit 1661.7 (floor on) < exact-LOBO 1877.3 < 5-fold-OOF 1913.3 < purge 1972.0.

Paired purge-vs-LOBO diagnosis (59's run; LOBO reproduced 1877.3 exactly before removal): the purge is HARSHER overall (+94.7) - entirely warm-up (3560.3 vs 2875.3: purging the calendar-local +/-42d window starves early scenarios more than building-holdout does) - and modestly softer on the panel (-18.1/origin; sched 8.9 vs 9.2 flat, recall 0.468 vs 0.461, rec12 0.576 vs 0.509). Reading: the same-device-fate channel exists but is ~-18/origin for the LINEAR champion - an order of magnitude below the +134/+216 self-fit flattery - because the label-overlap purge removes the direct channel and logbag cannot memorize identity. The known, accepted residual risk: flexible challengers (rf/gbm/dynamics) are flattered by same-battery training; the receipts are phases 51-54 and they are not to be re-argued unprompted.

The leaderboard closures (appetite, cert-first, pooling) rest on public measurements and survive any CV change. Public best 3c7b883 at 1661.45; private selection recommendation unchanged.

Phase 55 addendum (2026-08-22, user): harness de-tuned from the champion's output shape. "cv.py is tuned towards our current model (rec12 rec20 etc), remove those to make sure we can use the cv on any method we try." Removed the rank_stats call and every column built on it (rec12/rec20/recall/prec/ miss_rank/n_sched/caught/n_deaths) - all assumed the method emits a per-device p42 ranking. The harness now requires only that a method produce a PLAN; the one diagnostic kept beside the cost blocks is swaps (in-window unique batteries read from the plan itself), preserving the phase-39/51 appetite-in-disguise guard method-agnostically. Cost path untouched: SCORE reproduced 1972.0 exactly post-change (106s). models_cv.rank_stats remains for ad-hoc ordering diagnostics on methods that do rank. CSV SCHEMA CHANGE: /tmp/cv_all48oof.csv and cv_oofcand_* now carry n_swaps instead of the rank columns; the pre-change stale candidate CSVs were deleted (they were denominated in dead instruments regardless - any candidate must re-run fresh through cv oofcand <label>).

Phase 56 - full re-adjudication under the purge + radical-pipeline campaign; NEW CHAMPION logbag_rank+gbm30 (1630.3)

User directive: re-run ALL improvement ideas on the purge instrument, including previously public-refuted axes ("remove the closed line... we can re-try all of that"), spin off agents, be radical ("change the whole pipeline method"). Execution: 11 parallel lanes, ~250 candidates, every full-48 run paired per scenario against the champion baseline. Harness verified DETERMINISTIC (champion repeat = bit-identical on all 48), so paired SE is scenario-sampling noise only; survivor bar = delta < 0 beyond ~2SE at ~flat swaps. Live per-lane records: agent_reports/.md (kept in-repo).

NEW CHAMPION (adopted): logbag_rank + gbm value-preserving rank blend, w=0.30

cv.run_all48_oof(kind='logbag_rank', blend=('gbm', 0.30))
THE SCORE 1630.3 vs 1972.0  (-341.7 +/- 103.8, t=-3.3, 37W/11L,
swaps 9.23 vs 9.21 flat; wins all three blocks: wu 2276/panel 1677/ne 504;
gain entirely late_swap: -305 of -342)

Components, independently discovered and individually significant:

• logbag_rank (champion + within-origin fleet-rank features r_vlast/r_tau): -225.6 +/- 74.0, 34W/11L. Linear, rank-based, no same-device channel. This kind was REJECTED on the old self-fit instrument (phase 8); purge reverses.
• gbm order increment on that base: -116.1 +/- 53.6, 31W/13L. Broad weight plateau (.25/.30/.35 all <= -335). CAVEAT (md lane): a monotonic-constrained HGB partner DESTROYS the increment (+142..+269) - the gbm contribution is specifically non-monotone response, i.e. real interactions or residual same-battery flattery. Accepted risk per standing directive; public arbiter. Local optimum confirmed by 5 lanes: hyperparams (C=3 coin-flip -15.0+/-13.2, W/L against - kept C=1), features (12 full runs, all subsumed by r_vlast/r_tau; r_tau load-bearing), training weights (proximity/dedup/hard-example all neutral-to-worse on the frontier), neural rankers (RankNet/ListNet lose head-to-head +99..+120), physics blends (sim-rank 1.9sigma on old champ, ghost- promoting). vrel subsumed by ranks (worse in both combos).

DEPLOYMENT: submit.py FrontierRUL (BATTERYSWAP_FRONTIER=0 reverts to logbag). Fidelity test: bit-exact vs the CV candidate on s_10/s_30 (max p42 diff 0.0, ordering identical). Fit ~3.3s incl. gbm; budgets and degradation paths unchanged. cv.py CHAMPION now carries the blend (candidates with explicit kind/blend do NOT inherit it).

INSTRUMENT CALIBRATION - the campaign's most important negative result

The purge CANNOT adjudicate two classes, proven by public anti-correlation:

1. COUNT-COUPLED changes: purge alpha curve is monotone DOWN to 1.5 (-202)
   • the exact config the public refuted at +527 (phase 28). Same shape via p_be top-ups, EC-inclusion counts, dedup class-rebalance (tr: the -117 dedup win is 80% a smuggled +1.7-swap appetite shift; its count-matched twin is -24 n.s.), gamma flattening, cohort gates.
2. LABEL-PROCESS exploits: wd re-derived cert-first on the frontier at -26.2 (3W/0L/45T) - the EXACT local signature (-37.9, 3W/0L/45T) that preceded the public +43.2 (phase 48, tested ISOLATED on the clean stack; the lane's exculpation claim was checked against the record and is FALSE). Ghosts (crossed-never-labeled devices) exist publicly, are zero in train, and are invisible to ANY local instrument. SHIPPING BAR going forward: ordering-only, label-process-free, swaps flat.

Report-only results (real on the purge; fail the shipping bar)

• pl_stack_hgatez (frontier + arrival-triggered p_be=0.2 top-up): 1559.4 overall best raw number, but the gate's marginal on the frontier is -70.8 +/- 42.4 at +3.2 swaps with panel n.s. - count-coupled. The frontier self-repairs warmup (its wu swaps 9.9 vs 5.0), removing the gate's niche.
• pl_stack_shiftp4 (visit day 4*p42 days earlier): 1614.0; swap-neutral and reproducibly negative on two bases (champ -13.2+/-9.4, frontier marginal -16.3+/-10.9) but sub-2sigma each. Optional public A/B candidate.
• tr_hard_prox60 (hard-band x3 + proximity hl60 weights): -107.5 +/- 41.2 on the PLAIN base only (2.6sigma, swaps flat) - subsumed by the frontier.
• sv_dhgate (censored-pool discrete-time hazard when train deaths < 20): -51.4 +/- 22.1 - real, but the gate NEVER fires in deployment (full-history fit; cf. phase 36's inert warm-up mechanism, same effect size).
• tj_blend25e3 (MC voltage-trajectory sim rank blend 25%): -86.1 +/- 45.5 (1.9sigma) on the old champion base; promotes crossed-unlabeled devices (cert class) - report-only. Standalone physics pipeline: honest negative (labels carry unrecoverable info; count-matched it catches fewer deaths).

Closed by this campaign (with numbers, on the purge)

Planner: day placement CONVERGED (e2 surrogate local search -1.7+/-1.6 for 4x effort; ed/sd/blend day variants reject); piggyback/second-chance swaps dead (+8.8/+52); EC-economics count dead (inferior appetite); CVaR/portfolio dead (deaths do not cluster by building). Models: survival family dead (aft +337, cox +546, wiener +1100); conformal = byte-identical to logbag; building fixed effects dead (+423/+48); MoE/staleness gates inert; calibration null-to- harmful (Platt +21, isotonic -16.8 n.s.); MLP noise. Features: raw levels fail where ranks transfer (whit +43 raw vs -49 ranked); aux families dead (+95..+124); v_last_only +453 (blind-repo phase-39 claim dead on this instrument); top-12 slot squatters are label-process ghosts that still die at 19% - no flag can demote them profitably. Training: recency dead, proximity real-but-subsumed; label windows d35/d21 count-starve (+72/+105); building- balanced +40; thinning +19..+32; past-only pool +19 (future-side purge pairs stabilize fits). Candidate CSVs: /tmp/cv_oofcand_*.csv (~180 full-48 files).

Phase 56 addendum - PUBLIC REFUTES the unanchored frontier (698401a: 1877.08); mechanism found and fixed same-day

	swaps	late-days	early-days	total
champion 3c7b883	580	5819	13197	1661.45
frontier 698401a	517	7076	8011	1877.08
delta	-63	+1257	-5186	+215.63

The failure is a COUNT collapse, not (necessarily) an ordering failure: the value-preserving blend preserves the PRIMARY's (logbag_rank's) p42 values, and under deployment's PAST-ONLY fits that model's calibration collapses as the pool grows - measured sum(p42) ratio vs plain logbag: 1.02 (s_5), 1.01 (s_15), 0.95 (s_25), 0.80 (s_35), 0.78 (s_45). Public scenarios live in the late-origin regime -> -11% swaps -> ~+5 extra misses/48. The purge CV could not see it: its fits train on BOTH sides of the window, where the two models calibrate identically (swaps 9.23 vs 9.21). Fourth public refutation of a locally-validated intervention (alpha, cliff, cert, unanchored frontier) - but the first with the failing channel measured and separable.

FIX (same commit): FrontierRUL now VALUE-ANCHORS to plain logbag (BATTERYSWAP_FRONTIER_ANCHOR=1 default): the donor logbag supplies p42 values - count/pricing/timing byte-equivalent to the old champion's measured 580-swap calibration - and the rank+gbm frontier supplies ONLY the order. Ordering-only in deployment by construction, not just in CV. Verified at s_35/s_45 (worst collapse origins): anchored sum(p42) == plain logbag exactly; order == frontier blend exactly (top-13/top-25 sets identical; tie-aware rank comparison zero mismatches). What the anchored head tests publicly, cleanly for the first time: is the frontier ORDER better on the scored fleet? Downside is bounded to ordering (count machinery is the measured champion's).

Phase 56 addendum 2 - full decomposition: the frontier's purge gain was a hidden per-scenario appetite dial; head reverted to legacy default

Anchor decomposition on the purge (all paired, 48 scenarios):

arm	values (count source)	order	SCORE	vs old champ
unanchored frontier	logbag_rank	frontier	1630.3	-341.7 (publicly +215.6)
scale-anchored	lbr shape, donor sums	frontier	1925.1	-46.9 +/- 42.6 (33W/14L, n.s.)
value-anchored	plain logbag	frontier	1934.8	-37.2 +/- 44.1 (n.s.)

The smoking gun: mean swaps are FLAT (9.21 vs 9.23) but per-scenario counts swing |d| = 2.56 on average (23/48 scenarios differ by >= 2), and corr(count delta, cost gain) = 0.759. The 15 scenarios where logbag_rank's sum(p42) swelled the budget carry +956 gain each (~= the entire -342); the 25 where it shrank carry -2. MECHANISM: in both-sides purge fits, future-side pairs from a deadly era raise the fitted base rate exactly in deadly scenarios - era-level hazard leaking through CALIBRATION. The label-window purge cannot block it; deployment's past-only fits cannot replicate it (measured collapse 0.78-0.80x at late origins); the public board refuted it (-63 swaps, +215.6). PURGE LIMITATION #3, and it retroactively weakens every "swaps flat at the mean" clearance in this campaign: the shipping bar is now COUNT-MATCHED (value-anchored) evaluation, where the frontier's true ordering value is -37..-47 +/- 43 - real-looking W/L (33/14) but sub-bar.

ACTIONS: BATTERYSWAP_FRONTIER default OFF (head = legacy logbag champion, the measured-1661.45 behavior, plus all hardening); frontier dials retained for explicit A/B (ANCHOR=scale|value|off, W2). No model-side campaign candidate survives count-matched adjudication. Instruments going forward: deployment-mirror walk-forward (past-only fits) as the public predictor; building-GroupKFold with test-group-planned-as-fleet (user proposal, 2026-08-22) as the held-out/private-fleet predictor.

Phase 57 - CV replaced (user directive v3): building-GroupKFold, test group planned as a standalone fleet

The temporal purge is RETIRED as THE SCORE (its three deployment blind spots are the phase-56 record). New sole harness in cv.py: buildings -> 5 folds, stratified deterministically on failure count (snake assignment; fold deaths 22/18/14/15/13 over 5/5/5/5/4 buildings) and travel centrality as the location proxy. Per fold: ONE fit on the TRAIN buildings' closed pairs (full history - deployment-legal for foreign fleets), then for each of the 48 scenario times the TEST buildings are planned as a COMPLETE standalone fleet and costed. SCORE = mean over folds of fold-mean cost. This is a measured simulation of submit.py's held-out (static foreign fit) path - the path the private split's unseen fleets will take - with no per-building attribution problem (the fleet IS the test group) and no partial-fleet count bias (phase 26). Fits never see the test buildings; nothing sees the future of a label it trains on beyond what deployment itself would know. CLI: cv bgkf / cv cand <label> / cv report. Purge harness recoverable at f5f3fda; walk-forward family at 2000281. Candidate rule strengthened by phase 56: wins must hold at ~flat swaps PER CELL, not just at the mean.

Phase 57 baseline (champion = deployed plain logbag): SCORE 352.0 (folds 489.0/322.6/365.9/289.1/293.7; blocks wu 523.2 / panel 354.7 / nearend 108.8; swaps 1.9/scenario on ~90-device test fleets - the count rule scales with fleet size as designed). NEW DENOMINATION: per-test-fleet cost, never cross-compare with the full-fleet instruments (1972.0 purge / 1661.7 self-fit). Harness runtime 22s - cheap enough to run per candidate without triage subsets.

Phase 57 alignment check vs recorded public scores (user request; pooling arm killed mid-run): champion 352.0 reference; cert-first bgkf -7.3+/-4.3 vs public +43 (flip); frontier -5.3+/-7.4 vs public +216 (flip, but far less fooled than the purge's -342, and it reproduces the public swap-count-drop direction); alpha 1.5 -28.8+/-13.6 at swaps 3.15 vs public +527 (flip - still rewards appetite; note 65W/118L, the count-asymmetry signature). Verdict: better than the purge, still sign-flips all three public killers - fleet-composition facts (uncatchable public tail, train-absent ghosts) are invisible to ANY local split. The structural shipping bar (flat swaps per cell, no label-process exploits) remains the only local defense. One coherent cross-fleet signal: vrel -13.4+/-7.9, 104W/61L, swaps flat.

Phase 58 - definitive local-vs-public coherence check on the v5 harness (deployment-faithful GroupKFold)

All five publicly-measured configs reproduced AS SHIPPED (the alpha point as the true ALPHA_STATIC=1.5 + CLIFF_SCALE=2.2 stack driving the count through p42_sum; cert stacked on rank_score; pooling with time-masked own pairs; deployment-exact planner state incl. the gated floor; v5 baseline = 352.0, byte-identical to v3, floor confirmed inert):

config	local delta +/- SE	swaps	public delta	decision at the 2-sigma bar
alpha15+cliff22	-15.4 +/- 16.8	4.18	+527.3	NOT adopted (0.9 sigma) - correct
cert-first	-7.3 +/- 4.3	1.90	+43.2	NOT adopted (1.7 sigma) - correct
pooling	+10.2 +/- 3.6	1.92	+4.4	rejected - SIGN AGREES
frontier	-5.3 +/- 7.4	1.79	+215.6	NOT adopted (0.7 sigma) - correct
(alpha15 solo, no public point)	-28.8 +/- 13.6	3.15	-	2.1 sigma BUT count-coupled (swaps +1.25) -> quarantine rule catches it

VERDICT: the v5 CV cannot reproduce public MAGNITUDES (ghosts and the uncatchable tail are absent from every local fleet - unfixable), but as a DECISION INSTRUMENT the full protocol - v5 harness + 2-sigma adoption bar + count-flat rule + no-label-process rule - makes the correct call on all five public measurements: zero false adoptions, one correct rejection with sign agreement. Notably the faithful reproduction itself mattered: alpha-solo reads a confident-looking -28.8 (2.1 sigma) while the actually-shipped alpha+cliff stack reads a statistical zero - earlier alignment checks that approximated configs overstated the misalignment. Residual sharpest risk: cert-class (1.7 sigma local win, public +43) - the label-process quarantine is what protects against it, not the harness.

THE CV going forward: v5 building-GroupKFold (deployment-faithful held-out simulation, one instrument for public AND private - both are ~2x held-out fleets), used for ordering/model questions under the quarantine rules.

Phase 58 addendum (user): SCORE aggregation switched to the familiar full-fleet magnitude - per scenario the 5 fold costs are SUMMED (every building is out-of-fold exactly once per scenario -> the sum is the all-24-building out-of-fold cost), SCORE = mean of the 48 totals. Champion baseline: SCORE = 1760.2 (warmup 2616.1 / panel 1773.4 / nearend 544.0; swaps 9.5/scenario). Identical information to the fold-mean form (352.0 x 5), same paired decisions; stale pre-v5 candidate CSVs purged from /tmp (only the phase-58 faithful runs remain valid).

Phase 59 - campaign 2 (bgkf instrument, 7 lanes) closes; frontier ft_vmin7r_wnr20 ADOPTED and deployed (3f9bf12)

Deadline-day morning. Campaign 2 re-tested every prior method plus radical non-p42 pipelines on THE SCORE (bgkf v5 summed, baseline 1760.2), with a coordinator-maintained FRONTIER board for cross-lane composition (agent_reports_bgkf/). Overnight outage (repo re-cloned, /tmp+venv+dataset lost) fully recovered: dataset copied from rul_mgt_zero, venv symlinked to the blind repo (versions identical), baseline and every headline number reproduced BIT-EXACT (the harness is deterministic end to end).

ADOPTED - ft_vmin7r_wnr20 (commit 3f9bf12, BATTERYSWAP_FT2 default ON): logbag on FEATURES+vmin7_rank, wiener .20 value-preserving order blend. 1688.4 vs 1760.2 = -71.8 +/- 23.5 (3.05 sigma, Wilcoxon .037, swaps 9.58 flat, count-clean by every control, components additive -50.3 wiener / -23.2 vmin7r, reproduced exactly by ft, md and rk lanes independently). Deployment review that preceded adoption:

• Reference-shift worst case (the vrel killer): vmin7_rank recomputed WITHIN each ~5-building test fleet -> still -60.8 +/- 26.3 (2.31 sigma); consistent small degradation vs table-native (+11.0 +/- 17.9, 9W/35L). Deployment ranks over a full ~922-device fleet (features_at_origin own-fleet semantics), closer to the shared-reference case.
• Ft2RUL parity: submit.py implementation run through the CV harness reproduces the frontier CELL-EXACT (max diff 0.0, swaps identical).
• Deployment data path: wnr_m/s2/n merged into the shipped train table (97% coverage; WienerFHT shrinks through NaNs); split-path wiener params computed past-only from the scored split's own telemetry (1.3s once + 0.1s/scenario). Soft degradation: missing wnr -> pure vm7 ordering (values identical by construction); any fit error -> legacy champion. BATTERYSWAP_FT2=0 = the measured-1661.45 config.

NOT adopted (the two bigger nominal numbers):

• rk_4w (vm7 values; order .35 logbag + .20 wiener + .20 RankNet + .25 aux_ho) = -111.5 +/- 27.7, 4.02 sigma, Wilcoxon .001 - the campaign's strongest fully-anchored number, seed-block replicated. Head-to-head vs the frontier: -39.7 +/- 20.3 but 15W/15L/18T, Wilcoxon .131 - mean-only support; plus a torch dependency (MLP 32-16 x 7 seeds) the Docker is not known to carry. SHELVED (agent_reports_bgkf/artifacts/).
• wd raw vm7+aux_ho+wiener.20 = -126.4 +/- 43.8: rk's 4-way control run settled the adoption form - aux_ho as an ORDER source adds +6.1 n.s. on the frontier, so the raw excess is VALUE-CHANNEL (the calibration channel publicly refuted in phase 56). Report-only. aux_ho's anchored ordering core (-59.7 on champion base) is real but subsumed.

Lane verdicts (all closed, findings files final): pl - ranking is the wall (~99% of late cost = exclusion misses; routing axis +/-1). tr - keep the training pool exactly as-is; value-anchored twin flipped every count-smuggle. bl - wiener .20 the one piece; bagging/local-blend/calibration refuted. md - logbag(C=1,b=15) locally optimal; monotone HGB real (-44.7 anchored) but subsumed (+1.0 as third source); MORE BAGS HURT. ft - vmin7_rank the piece; entire vrel family quarantined (anchored twin +21.4, within-fleet recompute +89). rk - see above. wd/radical - physics loses to the 82 labels at matched count (+604); EC planner = count knob; direct death-day rejected both ways.

Private-selection matrix as of this phase: candidate head = 3f9bf12 (FT2 frontier); fallback = 3c7b883 (measured 1661.45); never 698401a (1877.08). The user manually triggers public scoring; a public look at 3f9bf12 prices the residual ghost/composition risk the CV cannot see (phase 58).

Phase 60 - campaign 3: exhaustive post-frontier exploration; frontier confirmed as the pipeline ceiling

User directive: improve the champion CV further, radical options included, review all assumptions (podium context: team-repo scale ~1400). ~45 candidates over 7 waves + 2 device-level autopsies + full mining of the team repo (rul-mgt-zero, 1493.65 public vs our 1661.45). Full detail: agent_reports_bgkf/c3_findings.md.

KEY FINDINGS:

1. AUTOPSY overturns the silent-death assumption: frontier recall is 48%; 237 missed deaths at median rank 5 (gap quartiles 2/3/8 below a ~rank-2 cutoff) carry 70% of remaining cost; only 45/237 are voltage-silent. False includes are plateau devices at 2.41-2.44V for 40+ days (36% die, median 83d — just outside the window). The prize is top-band ordering.
2. No candidate reaches it: every feature/order/model/count axis measured; best survivor is a -6-point crossing-label micro-piece (xlb, Wilcoxon .0475, shelved). ft2 (1688.4, 3f9bf12) is the ceiling of this pipeline under the count-clean bar.
3. Appetite re-audited and re-closed: alpha 1.0 reads -64.5 mean but 17W/24L, Wilcoxon .375 — warmup-carried count redistribution; gap-band precision ~11% < ~16% break-even.
4. Team repo decoded: their edge = multi-level self-supervised targets + poskeep_target sampling (-155.6, not portable — we already train on all rows) + CP-SAT planner over all nine cost components. Their CV is in-fleet (1412.6 not comparable to our held-out 1688.4). Direct feature transplant fails in our model form (+15..+167). Cross-pollination candidates TO them: WienerFHT order blend, vmin7_rank, building-held-out transfer check; their gen-10 state-dependent dV/dT (resistance channel) is the strongest open lead anywhere.

STANDING RECOMMENDATION unchanged: public-trigger 3f9bf12/head; private = 3f9bf12 unless publicly refuted, else 3c7b883; never 698401a.

Phase 61 - team solution evaluated ON our bgkf; operating-point finding; alpha-1.0 A/B prepared

User directive: run the team repo's stack under our CV and adopt structural changes if they improve it. Full port (panel rebuilt on our dataset with 81/81 exact EOL reproduction; model retrained building-OOF, h42 AUC 0.961; CP-SAT planner) plus four mix-and-match decompositions. NO ADOPTION: their full stack +228 vs ft2 (p .017, over-buys on ~90-device fold fleets); their ordering loses decisively anchored (+412, monotone dose-response at every weight); our ft2 model driving their CP-SAT lands at -96.3 mean but 16.4 swaps/scen, 25W/23L, Wilcoxon .72 = count-carried, mean-only. Instrument caveat recorded: their machinery targets 460-920 device fleets; our folds are ~90 - public is the only cross-pipeline arbiter (them 1493.65 vs us 1661.45).

OPERATING-POINT FINDING (leaderboard decomposition): the entire podium swaps 15-17.6/scenario with ~60-70 late-days/scen vs our 12.1 swaps and 121 late-days; the CP-SAT mix independently lands at the podium's operating point on our predictions. The alpha-1.5 public failure (6ea03da, zero catches) was on the OLD ordering; whether ft2's ordering converts extra appetite publicly is the one open question a public trigger can answer. PREPARED: commit 296f8f2 = alpha 1.0 on ft2 (single variable, selectable for public scoring); head 7d7632c keeps the safe 0.8.

Phase 62 - synthetic scaling study (user directive): CV tested on fleet-size scaling; fixed-reference vmin7 dial shipped

Synthetic fleets at 1x/2x/4x (461/922/1844 devices) from time-shifted building clones with OFFICIAL-mechanism-derived EOLs (0 mismatches on the checkable set at every scale); deployment-faithful static-path runs; anchor k=1 champ = 1762.5 ~= bgkf 1760.2 (same recall .471). Verdicts: count rule scales cleanly (swaps/death 0.98-1.07); live-rank vmin7_rank decays and inverts at 4x while a train-ECDF fixed-reference form is scale-neutral (shipped as BATTERYSWAP_FT2_FIXEDREF, default 0 - the Docker rehearsal measured 922-device splits = the safe 2x regime); alpha 1.0 best at EVERY scale with recall +7pts held (third independent support for the 296f8f2 public A/B); runtime 3.4s/scenario at 4x - no budget risk. Full tables: agent_reports_bgkf/c3_findings.md.

Phase 63 - PUBLIC CONVERSION: head 4c93883 measured 1575.31 (deadline day); wave T all dead

The user triggered public scoring on head 4c93883 (CV 1545.4; values lb(vm7)@alpha1.0, order .50 logbag + .30 wiener + .20 gapHGB, value-preserving): public 1575.31, -86.1 vs the 3c7b883 champion's 1661.45. Components invert to exact integers (clean run, default settings, 48 scenarios): 644 swaps (13.4/scen, was 580 at alpha .8), 5009 late-days (104.4/scen, was 121), 19716 early-days. CV->public transfer +29.9 (2%): bgkf v5 priced the whole stack (ft2 ordering + alpha 1.0 + gap-HGB + weight tweaks) correctly - the first locally-validated intervention stack to survive public after three pre-bgkf refutations. PRIVATE SELECTION: 4c93883 (hedge slot, if any: 3c7b883).

Wave T (post-ship, 13 arms, paired vs 1545.4): ALL DEAD - the champion is a local optimum on every axis touched. Wiener THRESH 2.40 beats 2.38/2.42 (+39/+45 - the labeled barrier is also the ordering-optimal barrier); K_SHRINK 6 vs 3/12 flat/+9; gap-weight .15 shelf +30/+29; gapHGB seed-bag +34 at 2.0sig (seed-0 is favorable; do NOT "stabilize" it away) and lr .03 +12; rank-shape transforms pow2/sqrt/logit +4/+42/+58 (linear pct-rank blending confirmed); GAP4 inside the logbag order-primary +123 at 2.3sig (gap features only work as a separate voice); RankNet 4th source at .10 +22. Driver /tmp/c9/run.py (volatile).

Phase 64 - T2 five-lane subagent campaign on the 1545.4 champion: one adoption (lam2sp3 planner objective), four axes closed with structure

Five parallel lanes, each paired vs the champion on bgkf (reports: agent_reports_bgkf/t2_{timing,planner,order,counts,hazard}.md):

• TIMING: closed. Early cost decomposed - 74% imputed-EOL false positives (never die in-window), 19% real deaths beyond the window, 7% genuine early timing; planner packing is 0.6%. Every date shift trades early for late at slope >=0.86; best arms 1.0sig AND count-coupled. EARLY is a selection problem, not a dates problem.
• PLANNER: ADOPTED lam2sp3 - timing curves x2 inside optimize_days' objective (new TIMING_WEIGHT dial) + SPLIT_DAYS 7->3. -4.11 +/- 1.41 (2.9sig), 25W/4L, Wilcoxon 8.8e-6, count-flat with battery sets cell-identical; dose-response plateau on both dials; additivity checks; survives ~35-arm multiplicity (sign p 1e-4). Mechanism: refuses cheap trip-merges that drag swaps off timing-optimal days (late -3.5, early -1.2, mobility +0.8). Also: plan-phase scheduling is SATURATED past its own surrogate optimum (exact enumeration is +5.1 WORSE - shallow search is an accidental regularizer; closes the CP-SAT question for good); of 232/scen "ops" cost only 78 is schedulable, 154 rides on missed swaps (recall in disguise).
• ORDER: closed. Boundary cascade fails structurally (cutoffs are n~2-3, the band IS the confident top; +147 at 2.6sig); per-building rank normalization catastrophic (+549/+939, 7-8sig) - global pooled ranking is load-bearing; prob-space blending loses to rank-space (+76). The rank-gap-1-10 frontier is unreachable by aggregation on this table.
• COUNTS (all report-only): flat alpha 1.1-1.5 dead (podium-posture 1.4 is a 1.0sig coin-flip, median-worse); calibrated counts REFUTED with a count-matched control (+67.5 at 3.4sig - allocation failure); risk-N refuted. Headline: warmup-gated alpha 1.4 (s_0-7 only) -> 1488.0, -57.4 +/- 22.0 (2.6sig nominal, post-hoc block, 8-scenario support, count-coupled) - candidate for a public trigger, user's call.
• HAZARD: closed for the deadline. dth3 crash fixed; discrete-time hazard HGB is a standalone statistical peer of the 3-way blend (h_alone 1537.0, 25W/22L) with its edge at COLD-START origins (censored-pool training; warmup -178/cell) - converges with the counts lane on warmup as the only exploitable block; unresolvable at 2sig on 8 cells of +/-800 scatter. AFT-lite dead on arrival.

SHIP (lam2sp3): submit.py TIMING_WEIGHT (default 2.0, env BATTERYSWAP_TIMING_WEIGHT) + SPLIT_DAYS default 7->3; models_cv.py plan_tsroute.timing_weight attr (default 1.0 = old path bit-exact, verified: ft2a10_sp7 control cell-diff 0.0 vs pre-T2 reference); cv.py run_bgkf split_days/timing_weight hooks, ft2a10/head track the deployed dials. Verified: cand ft2a10 = 1541.3, cell-exact (0.0) vs the lane's lam2sp3 measurement; paired -4.11 +/- 1.41 (2.9sig) 25W/4L p8.8e-6, swaps flat 2.27. New champion CV 1541.3.

Phase 65 - CV frontier 1398.0 (user CV-first directive): warmup-gated alpha 1.4 x gated hazard order, both at deaths_seen<35

User directive: CV is the reported/optimized metric; count-coupled candidates are KEPT (not auto-shipped) as champions-in-waiting. Frontier config st_hz100t35_g16: at origins where the planned fleet's recorded deaths <= T number fewer than 35 (~s_0-15 on bgkf), ORDER = discrete-time hazard HGB (censored-pool trained, c14/c17 machinery) replacing the 3-way blend, and planner alpha = 1.4; everywhere else byte-identical to the 1541.3 head. CV 1398.0, -143.3 +/- 37.4 (3.8sig), 16W/0L, Wilcoxon <.001, swaps 2.53/cell (vs 2.27; count-coupled by design).

Decomposition of the -143.3: gated alpha 1.4 width-16 alone -125.4 (4.0sig, clean bowl: width 6/8/10/12/14/16/20/24/32/48 -> 1501/1483/1462/1436/1429/1416/1414/1438/1475/1502; level bowl 1.3/1.4/ 1.45/1.5/1.6 -> 1476/1436-1403/1416/1427-1449/1490 with 1.4 optimal at every width); hazard order adds -13 at thr20 and -18 at thr35 (16W/0L final). EC break-even expansion (teammate mechanism through our pricing) REFUTED: margin 1.0/1.2/1.5 monotone worse (+17/+29/+46), so the count posture belongs to gated alpha, not EC top-ups. Teammate public analysis (1486.57 vs our 1575.31): +191 swaps -> late -353, early +316, ops -53; pure operating-point difference, no luck involved either way.

Caveats attached to 1398.0: gate widths selected on the shared 48 scenarios (multiplicity); deployment rule re-expressed as deaths_seen<35 (plan-time observable, FLOOR_GATE family); the composite harness carries exact-rank-tie noise vs a deployed-class implementation (16/240 cells, documented in t3_coldstart.md) - a real ship needs the gate inside the deployed class with non-gated origins byte-identical. Pending: t3_gm label-statistic lane; synthetic-scale validation of the count-coupled family (user-commissioned; key question = does the gate close early on large fleets, bounding the downside). Drivers /tmp/c19, /tmp/c20 (volatile); arm CSVs /tmp/cv_bgkfcand_c19_*, c20_*.

Phase 66 - SHIPPED the warmup gate (user call): head CV 1418.3

Deployed form of the phase-65 frontier inside Ft2RUL (user approved the ship for a public trigger): rate-form gate WARMUP_RATE=0.0759 deaths per training-fleet device (= 35/461; a fixed count fires wider on the CV's 4/5-size fold ledgers than on deployment's full ledger - measured 1429.7 fixed vs 1418.3 rate), TRAIN-ledger observable seen = deaths recorded <= plan origin; while gated: ORDER = censored-pool discrete-time hazard HGB (_DiscHaz, person-period buckets, depth3/lr.06/250it), appetite x1.4 via a consume-once plan_tsroute.alpha_boost. eol column derived from origin+days_to_eol when absent (the cached parquet ships no eol; exact, 0 diff on 8206 dying rows - this bug made the first rehearsal silently inert). VERIFIED: cand ft2a10 = 1418.3, -123.0 +/- 43.1 (2.9sig, 14W/5L, Wilcoxon .0017, 19 changed origins, swaps 2.54); WARMUP_RATE=0 control bit-exact vs the 1541.3 head (0.0 max cell diff); rehearsal: gate fired on exactly s_0-14 (+2..+5 swaps each), s_15+ byte-identical, 0 warnings, 41s. Delta vs the phase-65 composite 1398.0 (~+20) = the composite's harness-only advantages (full-table hazard pool, literal-16 gate, tie noise); the deployed number is the honest one. COUNT-COUPLED - public trigger arbitrates; previous head remains 60d65da (1541.3/public lineage 1575.31) via WARMUP_RATE=0.

Phase 66 postscript — PUBLIC REFUTATION. e4745d2 publicly scored 1594.89 (+19.6 vs the 1575.31 head). Mechanism executed exactly as designed: +38 swaps in the cold scenarios, late-days -263 (-54.8) — but early +59.8 and ops +14.4 made the trade ~1.4:1 against, vs ~6:1 favorable on bgkf warmup. Fourth count-coupled public refutation (alpha_static, cliff, cert-first, warmup-gate); alpha 1.0 remains the lone converter. bgkf's warmup block over-prices cold-scenario catches relative to the public fleet. PRIVATE SELECTION: 60d65da (count-flat lam2sp3, expected ~1571) or measured 4c93883 (1575.31); not e4745d2.

Phase 67 - T4 count-flat campaign (6 lanes, ~50 arms): ZERO adoptions - 60d65da confirmed as the count-flat ceiling of this feature space

User-directed sub-1400 count-flat push. Every lane paired vs 60d65da (1541.3) at 2sigma, parity-verified, swaps flat. All six REJECTED with mechanisms (reports agent_reports_bgkf/t4_*.md):

• STACKING: learned metas over-weight the best-AUC source (gapHGB); plan cost is decided at the cutoff, which the cost-tuned .50/.30/.20 already encodes. Loss monotone in meta dose (+16.5 at 25% ... +139 HGB-meta).
• EXTERNAL data/models: NASA transfer collapses to a monotone transform of the current voltage gap (Spearman .999) - lab aging corpora have near-uniform fade slopes; foundation models unloadable in the runtime stack (no transformers/safetensors). Measured +32.5 / wash.
• SURVIVAL/LTR: Cox, cost-weighted listwise, martingale-GBM all lose the gap slot by >=2.6sig WITH identical features - the dies42 classification objective itself is the value. Cost-aligned listwise is the worst (gradient too diffuse at ~1.2k positives/157 groups).
• LITERATURE features (knee/Bacon-Watts, sample entropy, monotonicity, relaxation, variance): all lose or wash inside the gapHGB; best standalone-AUC features lose MOST (dilute the depth-2 splits on the validated quartet). Anchored logbag control failed on cue (+61.6).
• TRANSPLANT (rank-4 MnesisLab, full source read): geo-aware ordering tilt +135..+452 (their compactness gate passes on all our fold-fleets, so gating does not rescue it); seasonal temp layer = plan no-op (gated devices already top-ranked); left-truncated FP-AFT faithful (params match theirs) but moves ~1 unit; their EB wiener +128. THEIR 1407.63 IS OPERATING POINT + their untransplantable base pipeline, not a readable layer: 743 swaps converting at ~16.8 late-days/swap vs our 6.0 marginal.
• SEQUENCE (GRU-D on 60d daily telemetry, 5-seed, fold-safe): encoder genuine (OOF AUC 0.94-0.96) but fully redundant with the snapshot stack; monotone-against dose response (+21.6 at w=.05).

CONCLUSION: 60d65da (CV 1541.3) is the measured count-flat ceiling of this feature space across our own features, the literature's, and the readable competitors' - the ordering axis is closed from every direction tried. PRIVATE SELECTION stands: 60d65da first, 4c93883 (measured 1575.31) as the no-inference alternative.

Phase 68 - T5 global-optimum probe (6 structural lanes): ZERO adoptions - 60d65da is the honest optimum of this framing, search closed

User-directed final campaign ("it's a local optimum but probably not the global one") - four commissioned angles plus the two rank-3-derived prototypes, every arm paired vs 60d65da (1541.3), parity-verified, count-flat unless noted (reports agent_reports_bgkf/t5_*.md):

• BASE REBUILD: competitor-style wide-space stacked base (78 leak-verified cols, 4-learner stack) IS a better dies42 classifier than logbag (per-fold OOF AUC 0.965 vs 0.938, 4/5 folds, biggest where labels are scarce) and it does NOT price: |delta| <= 9.2 (<=0.3sig) at every mixing weight. AUC gains live in separating obvious cases; cost lives in the fine ordering near the inclusion cutoff, which logbag already has. Base-model capacity is not the binding constraint.
• JOINT SEARCH: 8-dim differential evolution, 90 evals, tune/validate origin split. Best tune delta -104; 100% of it sign-reverses on the held-out origins (top-3 all flip). The CV's resolution is exhausted at this scale - any future multi-dial "win" of this size is presumed noise unless it passes the same split.
• SELECTION ARCHITECTURE: expected-gain (cost-based) selection at pinned count: best arm -5.0+/-28.8 (0.2sig); the geography term is a measured LOSS +27..+31/scen (2sig, robust across three geo-on/off contrasts). The rank-4 team's -14.6 selection edge exists only count-floating.
• DATA WALL: learning curve 40->100% is shelf-not-slope (local slope ~0, p=.68); a hypothetical DOUBLED corpus prices at ~-44 [-100..0]; label audit recovers zero positives. The surface rewards building/failure-mode COVERAGE, not volume - the optimum is not gated on labels nobody has.
• SEASONAL MARGIN (rank-3 recipe): the fleet seasonal factor is REAL here (p2p 0.0585 V in demeaned fleet-median voltage, matches their 0.048) but injecting it into the wiener barrier distance loses at every strength on both climatology maps (best arm +1.0, and the mild setting is significantly harmful, wilcoxon p=.031; one Dec origin +546 from a 2 mV shift). Their -104 lived in calibrated-probability expected-cost selection; in a rank-only frozen-count slot the factor only reorders, and it reorders worst where deaths cluster. Wrong lever, axis closed.
• DRAWDOWN REFRAME (rank-3 core method, faithful rebuild): 1.72M (h,m)-conditioned censored drawdown rows from 71,878 battery-days, their exact HGB (450/.07/63/200), 107 causal features, fold-safe. Monotone- against dose response: alone +255.6 (3.9sig), as 4th source @.10 +14.1 (0.7sig loss, best arm); survival-curve read strictly worse. OOF AUC 0.958 vs raw v_min_7's 0.961 - it learns first-passage of the LAGGING smoothed statistic while the champion sits on leading raw-voltage indicators. Their 1350.7 is their whole stack (expected-cost scheduler, seasonality, timing), not an order signal that transplants.

CONCLUSION: with T2-T5 plus peer transplants (~150 arms) the ordering, base-capacity, selection, data, seasonality and reframing axes are ALL closed with mechanisms. 60d65da (CV 1541.3, count-flat) stands as the genuine optimum of this framing at this instrument's resolution.

FINAL SELECTION (deadline 2026-08-23 23:59): private pick 60d65da first; 4c93883 (publicly measured 1575.31) as the no-inference alternative. e4745d2 publicly refuted (1594.89). Never 698401a. Count-coupled CV frontier (gated hazard+alpha 1398.0) remains report-only: the family is 0-for-4 on public conversion.

Phase 69 - PUBLIC read on the restored head (71bfff2 = 60d65da config): 1583.11, lam2sp3 does NOT transfer - private pick flips to 4c93883

User-triggered public scoring of the gate-off head: total 1583.109 vs 4c93883's measured 1575.307 (+7.80). Component inversion (all integers, clean run): swaps 644 = IDENTICAL to 4c93883 (count-flat exactly as designed); late-days 5005 (-4); early-days 19638 (-78) -> timing components -1.64; ops/packing residual +9.45 (daily_limit 112.5, overtime 89.2, weekly_limit 64.6, travel 49.1, bldg/room 17.2). MECHANISM: timing_weight=2.0 + split_days=3 re-packs swap days - it bought the intended timing gain but pays +9.4 in schedule/ops constraints on the public fleets. The bgkf priced the same trade at -4.1+/-1.4 (2.9sig, 25W/4L) on train buildings: the ops cost of tighter packing is fleet-composition-sensitive (building sizes/geography), which building-OOF CV on train buildings cannot see. First COUNT-FLAT CV->public inversion (prior four were count-coupled); the CV remains accurate on ordering/inclusion but does not price planner-packing trades across fleets. FINAL SELECTION (updated): private pick 4c93883 (measured 1575.31, best public we own) FIRST; 60d65da/71bfff2 (measured 1583.11, CV favorite at 1541.3) as the second slot / CV-side hedge if two picks exist. Never e4745d2 (1594.89) or 698401a (1877.08).

Phase 70 - Fleet-composition probe: the CV's lam2sp3 win is a LOW-PRESSURE artifact - pick locks to 4c93883

Post-refutation diagnostic (report agent_reports_bgkf/t5_fleetcomp.md, raw /tmp/c35): A = tw2.0/sd3 (60d65da) vs B = tw1.0/sd7 (4c93883), identical model, paired, count-flat in all 1536 stage-2 cells.

• Fold fleets: A wins 5/5 but pays an ops tax in 4/5 - the CV win is entirely late/early timing on small sub-fleets.
• 32 random subset fleets (3-12 bldgs): 20W/8L/4T, median -0.14; the dials only ACTIVATE under pressure - P(cell differs) 5.5%->85% from <50 to 300-400 devices, corr(delta, n_dev) -0.71.
• PRESSURE fleets (18 bldgs, 250-384 devices): 3/6 flip to A-loses; the two big losses replicate the public signature almost exactly (+7.90 with dOps +10.82; +9.31 with dOps +9.35; public +7.8 with ops +9.45, timing slightly better in all three).
• CAPSTONE: the full 24-building universe planned as ONE fleet flips A to +0.97 (24W/20L, dLate -4.4, dOps +6.47). The CV's 2.9sig win exists only because bgkf test fleets are five small sub-fleets where daily/weekly limits never bind. VERDICT: the -4.1 CV delta is not transferable to large fleets; at public-like pressure A-vs-B is a coin flip locally and A measured worse publicly. INSTRUMENT FIX (for future work): adjudicate planner/packing dials on pressure-matched fleets (whole-universe-as-one-fleet + stress fleets at 250-400 devices), not the pooled small-sub-fleet mean. FINAL SELECTION locked: 4c93883 (measured 1575.31). 60d65da/71bfff2 only as a second slot. Never e4745d2 / 698401a.

Phase 71 - Pressure-matched CV instrument: CALIBRATION FAILED - no instrument change, no final round; selection stays 4c93883

User-approved attempt to upgrade the CV to pressure-matched evaluation (whole-24-building-universe as one fleet, OOF-stitched per-device predictions from the 5 fold fits) and re-run a packing-dial grid on it. Calibration gate: retrodict the public pair (B=tw1/sd7 1575.31 < A=tw2/sd3 1583.11, dOps +9.45). Result: PRESSURE(B)=1619.8, PRESSURE(A)=1615.9, A-B = -3.88+/-5.17 (22W/19L/7T), dOps -1.71 - WRONG SIGN on both total and ops; grid auto-stopped per protocol. Diagnosis (t6_pressure.md): at binding pressure the A-B delta is a daily/weekly-limit COLLISION LOTTERY - +/-100/200 quanta ~13x the +7.8 target signal, landing on different origins under OOF-stitched vs leaky-fit predictions (per-origin deltas uncorrelated, r=-0.08) - with only 48 origins the one-fleet mean cannot be calibrated; any grid ranking would fit the lottery. Corollary: the public +7.8 itself is likely 1-2 collision events, i.e. the packing channel may be UNRESOLVABLE by any instrument at this scale; the measured public number is then the only defensible ranking, and it favors B. Redesign direction for future work (not attempted tonight): sign-frequency over a many-fleet stress ensemble rather than a one-fleet mean; stress validator ready at /tmp/c36/stress_winner.py. FINAL: cv.py UNCHANGED (an uncalibrated instrument would mislead); selection stays 4c93883 first, 60d65da/71bfff2 second slot only.

Phase 72 - Final count-flat closure (c28/c37/c38) + alpha 1.4 public A/B ship (user call)

Deadline corrected to 02:00 CEST (midnight UTC) - time for one more public A/B. Count-flat space closed with three sweeps (all paired vs 1541.3, parity exact in each driver):

• c28 hyperparams (logbag C/bags, gapHGB l2/leaf): 8 arms, ALL lose or wash (+5.6..+34.2). Champion hyperparameters locally optimal.
• c37 blend-weight jitter (+/-.05 around .50/.30/.20, 6 directions): ALL lose +23..+50 (several ~2sig AGAINST). Sharp optimum.
• c38 averaging ensembles (leaders' style, order-only: device-bootstrap x3, drop-fifth x5, time-snapshot x3, feature-set x2): best is a 0.3sig wash (devboot -6.2+/-24.9), timebag a 2.2sig LOSS. Averaging destroys the favorable seed-0 configuration (same lesson as wave T seed-bag). ALPHA 1.4 SHIP (62e9d91, count-coupled, explicit user call): fresh confirm on head config bgkf 1498.3 (-43.0+/-43.0, 1.0sig, 19W/24L, swaps 2.27->2.97/fleet = podium operating point 15-18/scen; mean carried by death-heavy origins - the public read arbitrates). Static path governs held-out fleets (submit.py:836 ALPHA_STATIC), both knobs set. Rehearsal: 520->688 swaps (+32.3%), 43/48 scenarios; ship-code counts byte-identical to env-override run. Old-ordering alpha 1.5 failure (2189) predates ft2 - this is the operating-point hypothesis' first test on the current ordering. DECISION RULE: public < 1575.31 -> select 62e9d91; else select 4c93883. Revert = both alpha defaults 1.0.

Phase 73 - PUBLIC CONVERTS alpha 1.4: 1480.85 (-94.5) - both pre-registered gates pass; SELECTION = 62e9d91

User-triggered public read of 62e9d91: total 1480.85 vs 1575.31 (4c93883). Pre-registered decision rule (phase 72): margin >= 15 AND late-swap fingerprint (late down >= 100). Both pass decisively:

• margin -94.46 (6x threshold); components invert to integers (clean).
• swaps 747 (+103, 15.6/scen - INTO the podium band 15-17.6);
• late-days 4006 (-1003; component -208.96 = 2x the required fingerprint) - the marginal-rank batteries DO contain real deaths on held-out fleets: ~0.8 extra catches/scenario at a strongly profitable exchange (early +115.35, ops -1.40 flat). FIRST large count-coupled public conversion (family was 0-for-4; the old "ranks 13-20 convert at ~0%" verdict was an OLD-ordering fact - the ft2 ordering's marginal band converts). The public-overfit objection was pre-answered: single pre-registered hypothesis from external evidence (podium operating points + bgkf -43), component fingerprint predicted before the read, margin rule against collision noise. By sibling-split exchangeability the effect transfers to private. SELECTION: 62e9d91 (public 1480.85) PRIMARY; 4c93883 (1575.31) fallback slot. Dose-response CV (alpha 1.6/1.8/2.0) running to decide whether ONE more A/B is justified - stopping rule: only if the CV curve is still falling at 1.6+; iterating alpha values on public beyond one more point = the probing pattern we reject.

Phase 73 postscript - dose-response complete: alpha 1.4 -> 1498.3 (-43.0), 1.6 -> 1562.1 (+20.8), 1.8 -> 1573.7 (+32.4), 2.0 -> 1589.0 (+47.7). Clear minimum at 1.4; monotone rise beyond. NO second public A/B per the stopping rule. FINAL: select 62e9d91 (1480.85); fallback 4c93883 (1575.31).

Phase 74 - Re-tune at alpha 1.4 (user-directed): dials orthogonal EXCEPT the wiener weight - basin floor .20 (CV 1476.0, -22.3, 2.0sig)

22+10 arms paired vs the alpha-1.4 reference (1498.3), parity exact in all drivers, count-flat (swaps 2.95-2.98):

• Alpha fine grid: 1.4 is a SHARP minimum (1.3 +36; 1.35 +2.7 wash; 1.45 +22.9 at 1W/21L; 1.5 +47.6 at 3W/29L; 1.6/1.8/2.0 monotone up). Public-confirmed operating point = the exact CV bowl bottom.
• Packing (tw x sd) at 1.4: flat (< +4.6) - independent of alpha.
• Hyperparams (logbag C/bags, gapHGB l2/leaf) at 1.4: all within +-4.3.
• ORDER WEIGHTS SHIFT: at alpha 1.0 the .50/.30/.20 mix was a sharp optimum (phase 72: all jitters +23..+50); at alpha 1.4 the wiener->gap direction is a coherent basin: .25/.25 -13.9 (p.013), .23/.27 -3.7 (p.042), floor at wiener .20 - .50/.20/.30 = 1476.0 (-22.3, 2.0sig, p.013) and .55/.20/.25 = 1475.2 (-23.1, 1.6sig, 26W/8L p.004) - then rising (.15/.35 -8.0, .10/.40 +6.2). Mechanism: the wiener barrier distance ranks imminent deaths well but discriminates poorly in the deeper 15-19/scen band that alpha 1.4 buys; the gapHGB is stronger there. The dial's CV optimum is OPERATING-POINT-DEPENDENT. Winner's-curse haircut (best of ~30 arms at this alpha): true effect likely -12..-18. NOT shipped as of this entry: the selection stands on the publicly MEASURED 62e9d91 (1480.85); re-selecting an unmeasured re-weight is a user-level risk decision (no public triggers remain).

Phase 75 - PUBLIC REFUTES the order re-weight: 1495.92 (+15.1) - FINAL selection stays 62e9d91 (1480.85); head reverted

User-triggered read of 354fd2d (.50/.20/.30): 1495.92 vs 62e9d91's 1480.85 (+15.08). Counts flat as designed (752 vs 747 swaps); the loss is spread evenly: late +5.00 (4030 vs 4006 late-days), early +3.67, ops +6.39. CV said -22.3 (2.0sig, basin-confirmed, p.004/.013 floor arms); public says +15.1 - a SIGN FLIP in the bgkf's home channel (ordering). Post-mortem read: best-of-~30 winner's curse plus a real fleet effect - the wiener barrier-distance source earns MORE weight on held-out fleets than building-OOF folds suggest (its physics transfers; the gapHGB's depth-2 splits are more fold-idiosyncratic). Cost of the experiment: zero (selection rule pre-locked, portal unchanged). LEDGER OF THE NIGHT'S A/Bs: packing CV-win refuted (+7.8), warmup gate refuted (+19.6), alpha 1.4 CONVERTED (-94.5, amplified 2.2x), order re-weight refuted (+15.1). The bgkf prices the count/appetite channel and long-standing ordering STACKS, but individual 2sig dial moves in any channel are coin flips publicly - only public A/Bs adjudicate. FINAL: selection 62e9d91 (public 1480.85); head reverted to its config.