HEAD: T3 warmup gate - gated hazard order + alpha 1.4 boost - CV 1418.3

While the training fleet's death ledger at the plan origin is below
WARMUP_RATE (0.0759/device = 35/461), Ft2RUL switches ORDER to a
censored-pool discrete-time hazard HGB and boosts appetite x1.4 via a
consume-once plan_tsroute.alpha_boost; past the gate the model is
byte-identical to the previous head (WARMUP_RATE=0 disables exactly,
verified 0.0 max cell diff vs 1541.3).

bgkf: 1418.3, -123.0 +/- 43.1 (2.9sigma), 14W/5L, Wilcoxon .0017.
COUNT-COUPLED (swaps 2.54 vs 2.27) - shipped on the user's CV-first
call for a public trigger. Scale harness: gate narrows as fleets grow
(8/4/0 active scenarios at 1x/2x/4x) - bounded downside. Rehearsal:
gate fired on exactly s_0-14, s_15+ byte-identical, 0 warnings.
Includes T3 lane reports (gm label-statistic axis closed: the exact
label is a lagging smoother; cold-start gating; scale validation) and
phases 65-66.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

EXPERIMENTS.md

@@ -3391,3 +3391,58 @@ 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.

agent_reports_bgkf/t3_coldstart.md

@@ -0,0 +1,121 @@
+# T3 COLD-START lane: order adapted to training-information richness
+
+Agent: T3 cold-start (c17). Date 2026-08-23 (deadline day). Work dir /tmp/c17/
+(driver run.py, analysis analyze.py, gate diagnostic gate_diag.py). Arm CSVs at
+/tmp/cv_bgkfcand_c17_*.csv. Reference /tmp/cv_bgkfcand_ft2a10.csv = deployed
+head 60d65da at 1541.3 (asserted). All arms: values/counts from primary
+logbag(vm7)@alpha1.0 untouched (order-only via run_bgkf blend hook at w 1.0,
+count-flat by construction; swaps 2.275 vs ref 2.267 on every arm),
+split_days=3, timing_weight=2.0.
+
+## The gate (plan-time observable, no forward reads)
+
+deaths_seen(origin) = number of unique TRAIN-fleet devices with recorded
+eol <= the plan origin date. Computed at fit time from the fold-train pool
+(dev_eol ledger, exactly walk_forward's / submit.py FLOOR_GATE's family),
+evaluated per predict() on the origin timestamp the rows carry. A death is
+observable from the moment it happens; the train fleet's labels are fully
+known when a foreign fleet is served -- deployment-legal both ways.
+Gate g = 1[deaths_seen < 20] (FLOOR_GATE parity). Firing pattern (fold
+crossings of 20 at origins [11,12,9,12,9]): all of warmup s_0-7 plus a
+2-4-origin bleed into early panel (s_8-11). deaths_seen at s_0 is 1-3,
+s_7 is 12-15. Hazard source = c14's DiscHazHGB (discrete-time hazard HGB,
+person-period buckets, censored pool via mc.censored_pairs at max origin
+filtered to train devices -- byte-identical to run_bgkf's own hazard branch).
+
+## Baseline calibration (matters for reading the table)
+
+cs_champ (my composite replication of the deployed order, champion weights,
+no gate) scores +18.5 +/- 15.0 vs the deployed ref with only 16/240 cells
+differing, dominated by 4 large sign-mixed jumps (+546, +280, +277, -248):
+the documented exact-rank-tie flips (submit.py's own comment: one flipped tie
+moved a cell by 539; the blend hook's extra rank layer resolves ties
+differently than Ft2RUL's internal sort). So:
+- "vs ref" carries ~15-cost-unit tie-flip noise unrelated to the mechanism;
+- "vs cs_champ" is the clean paired mechanism read (same code path, plans
+  bitwise identical at every non-gated origin -- deltas exactly 0 there);
+- "ship-proj" = delta vs ref restricted to gate-active origins (s_0-11),
+  exact 0 elsewhere: what a deployed-class implementation would score, since
+  the gate leaves non-gated origins byte-identical to the deployed head.
+
+## Results (48 paired all-building scenario totals)
+
+vs deployed ref 1541.3 | per-block = per-origin mean within block
+
+| arm        | SCORE  | vs ref        | vs cs_champ (clean)     | ship-proj      | W/L(ref) | wilc(ref) | swaps | warmup | panel | nearend |
+|------------|--------|---------------|-------------------------|----------------|----------|-----------|-------|--------|-------|---------|
+| cs_champ   | 1559.8 | +18.5+/-15.0  | --                      | --             | 6W/8L    | p0.391    | 2.275 | +34.6  | +15.5 | +13.6   |
+| cs_hz10    | 1556.5 | +15.2+/-15.9  | -3.3+/-4.0 (0.8s)       | ~0             | 7W/10L   | p0.548    | 2.275 | +38.8  | +10.0 | +13.6   |
+| cs_hz20    | 1547.0 | +5.7+/-18.3   | -12.8+/-9.5 (1.3s)      | -7.0+/-11.6    | 9W/10L   | p1.000    | 2.275 | -18.5  | +10.0 | +13.6   |
+| cs_hz35    | 1549.9 | +8.6+/-18.2   | -9.8+/-12.5 (0.8s)      | -4.1+/-11.5    | 9W/10L   | p0.738    | 2.275 | -27.0  | +16.1 | +13.6   |
+| cs_hz60    | 1520.0 | -21.3+/-23.1  | -39.7+/-20.1 (2.0s)     | -34.0+/-17.9 (1.9s) | 14W/10L | p0.331 | 2.275 | -149.0 | +2.6  | +13.6   |
+| cs_hz100   | 1514.7 | -26.6+/-28.0  | -45.1+/-25.5 (1.8s)     | -39.3+/-23.9 (1.6s) | 14W/11L | p0.474 | 2.275 | -177.9 | +1.9  | +13.6   |
+| cs_hz20t30 | 1540.6 | -0.7+/-15.1   | -19.2+/-15.6 (1.2s)     | --             | 9W/11L   | p0.985    | 2.275 | -18.5  | +1.0  | +13.6   |
+| cs_ramp    | 1564.9 | +23.5+/-15.8  | +5.1+/-4.4 (0W/3L)      | ~0             | 6W/10L   | p0.231    | 2.275 | +38.8  | +21.7 | +13.6   |
+| cs_wnup    | 1552.7 | +11.4+/-20.4  | -7.0+/-13.4 (0.5s)      | --             | 8W/10L   | p0.609    | 2.275 | +42.3  | +3.8  | +13.6   |
+| cs_lbdown  | 1552.1 | +10.8+/-17.6  | -7.7+/-8.5 (0.9s)       | --             | 8W/10L   | p0.671    | 2.275 | +38.6  | +3.8  | +13.6   |
+
+Arm rules: cs_hzNN = (1-g*w)*champ_order + g*w*rank(DiscHazHGB), w=NN/100,
+g=1[deaths_seen<20]; cs_hz20t30 same at thr=30 (fires to s_14-18);
+cs_ramp w_hz=.20*max(0,1-deaths_seen/20); cs_wnup gated weights .40/.40/.20;
+cs_lbdown gated .40/.36/.24 (no-hazard controls). Weights round(.,12).
+
+## Best arm per-block detail (clean pairing vs cs_champ)
+
+cs_hz60:  warmup -183.6/origin, panel -12.9/origin (ALL of it at the gated
+bleed s_8-11; s_12+ exactly 0), nearend exactly 0. 8W/3L on the 11 changed
+origins, wilcoxon p0.083.
+cs_hz100: warmup -212.4/origin, panel -13.7 (gated bleed only), nearend 0.
+8W/4L, p0.151. Per-origin vs ref: s_0 -422, s_1 -879, s_2 -107, s_3 +119,
+s_4 +8, s_5 +156, s_6 -234, s_7 -64, s_8 -17, s_9 +210, s_10 -190, s_11 -468.
+Gains are largest where deaths_seen is smallest (s_0/s_1: 1-3 deaths seen) --
+the dose matches the mechanism, cell by cell.
+
+## Mechanism evidence
+
+1. Dose-response monotone in gated hazard weight: warmup (vs ref) +39 / -19 /
+   -27 / -149 / -178 per origin at w = .10/.20/.35/.60/1.0.
+2. cs_hz100's warmup -177.9/origin REPLICATES the T2 hazard lane's
+   independently measured "warmup -178/cell for hazard-alone" (c14 h_alone,
+   old dials) -- the number was predicted before this campaign ran.
+3. Hazard-specific: the no-hazard gated controls (cs_wnup +7.7, cs_lbdown
+   +4.1 warmup) capture none of the warmup gain -- it is not "less logbag at
+   cold start", it is the hazard model.
+4. Concentration exactly as claimed: every point of gain sits inside the
+   gate-active origins; s_12+ and nearend are bitwise-identical plans.
+5. Count-flat: order-only by construction, swaps 2.275 vs 2.267 ref, every arm.
+6. Dead as predicted: cs_hz10/cs_ramp too weak (small hazard weight is
+   worse than none: at cold start the champ sources are the noise, so
+   diluting them slightly just perturbs ties); widening the gate to thr=30
+   (cs_hz20t30) adds nothing beyond its s_8-11 share.
+
+## Tier verdict
+
+New mechanism (per-origin gated order adaptation) => bar is 2sigma paired +
+concentration + controls. cs_hz60: 2.0sigma on the clean pairing
+(-39.7+/-20.1), 1.9sigma on the ship projection (-34.0+/-17.9, 8W/3L gated
+origins, wilcoxon p0.083), concentration textbook, controls pass.
+cs_hz100: bigger point estimate (-45.1 clean / -39.3 ship-proj) at
+1.6-1.8sigma. Raw vs-ref sig is only ~1sigma for both because tie-flip noise
+at non-gated origins contaminates that comparison; a deployed-class
+implementation would not carry it.
+
+VERDICT: cs_hz60 MEETS the bar at exactly 2.0sigma (clean pairing) with all
+three qualitative gates passed; cs_hz100 is the better point estimate but
+sits below 2sigma. Caveats for the main session: (i) the 2.0sigma is the max
+of a 5-point weight sweep -- mitigated but not eliminated by the monotone
+dose-response and the pre-registered T2 warmup number; (ii) the effect lives
+on 11-12 origins (effective n small, sign-mixed 8W/3-4L); (iii) warmup block
+public relevance depends on the held-out scenarios actually containing
+early-history origins. Recommend ship-candidate = gated hazard at w=.60
+(or the .60-1.0 shelf; the curve is flat-to-improving there), gate
+deaths_seen<20 on the train eol ledger, implemented inside the deployed class
+so non-gated origins stay byte-identical. Hazard model must ship with the
+pickle (Docker/sklearn 1.7.2 retrain lock applies).
+
+## Not retested (already dead) / files
+
+Global hazard 4th source, global slot replacement, warmup-gated ALPHA
+(count-coupled), per-building rank norm, boundary cascade, flat simplex --
+untouched. Driver /tmp/c17/run.py; gate diagnostic /tmp/c17/gate_diag.py
+(per-fold deaths_seen curves); tie audit in analyze output above.

agent_reports_bgkf/t3_gm.md

@@ -0,0 +1,92 @@
+# T3 gm campaign: exact EOL label statistic, replicated — no delivery wins
+
+LABEL-STATISTIC agent, 2026-08-23. Champion = head 60d65da (CV 1541.3,
+order .50 lb / .30 wiener / .20 gapHGB, split_days=3, timing_weight=2.0).
+Reference /tmp/cv_bgkfcand_ft2a10.csv (asserted 1541.3). All arms paired,
+per-origin, 48 origins. Work dir /tmp/c16/ (validate*.py, build_gmfeat.py,
+run.py, run2.py, gmfeat.parquet, simfeat.parquet, cv_bgkfcand_c16_*.csv).
+
+## 1. Mechanism replication: 81/82 exact death dates, 0 false positives
+
+The competition EOL label is deterministic from telemetry. Re-derived recipe
+(from dataset/train/battery_metrics.parquet vs dataset/train/eol_times.csv,
+82 non-null deaths — one more than the remembered 81):
+
+1. Gate readings to temperature strictly inside (10, 30) — exclusive bounds
+   (same as features.py daily_voltage).
+2. Daily median voltage per device; a day is VALID only if it has
+   **>= 5 gated readings** (minn=6 breaks 7 devices, minn=4 breaks 8).
+3. Reindex valid daily medians onto the full CALENDAR grid (crossings land
+   on no-data days as old high days fall out of the window — evaluating
+   only on observation days loses ~9 devices).
+4. Rolling 7-calendar-day median, **min_periods=3** (window needs >= 3
+   valid days; pandas interpolated median for even counts).
+5. EOL = first day where the rolling value <= 2.40.
+
+Score: **81/82 exact day-level matches, 0 date mismatches, 0 survivor false
+positives**. The 82nd (d_cfdd8b69ddd4) dies on the day its statistic equals
+exactly float32(2.4)=2.4000000954; comparing <= float32(2.4) fixes it but
+makes d_9ba303a913ad cross 37 days early at the same float value — the
+original generator's tie-at-threshold semantics are unresolvable from data
+(irrelevant at feature level). Journey: naive gated-daily-median cal7 recipe
+= 60/82; +calendar-grid eval = 61; +minn5/min_periods3 = 81. Pooled-reading
+median rejected (38/82). The dip-recovery "mid-life swap" mismatches
+(-72..-10 days) were all artifacts of low-count winter days polluting the
+window; the n>=5 validity rule removes them exactly.
+
+## 2. Feature cache (leak-free, repo day<=origin convention)
+
+/tmp/c16/gmfeat.parquet — 77,267 rows keyed (device_id, origin_idx), all 168
+origins (48 real + synthetic), 100% real-origin coverage: gm_last, gm_gap
+(=gm_last-2.40), gm_age, gm_slope_7/21 (polyfit on the rolling-median series,
+>=3 pts), gm_dtc (gap/max(-slope21,1e-4), clip [0,400]), gm_cov_30.
+/tmp/c16/simfeat.parquet — sim_k: exact days-to-label-crossing if the last
+valid daily-median level persists (window arithmetic simulation), else 400.
+
+Signal autopsy (real origins, n=19,890, 454 dying rows):
+- AUC dies42: gm_gap 0.939, gm_dtc 0.852 — vs v_corr_last 0.945, v_min_7
+  0.961. **The exact label statistic is a lagging smoother; our existing
+  proxies anticipate it.** Median dying-device within-origin rank: v_min_7
+  10 < gm_gap 14.
+- gm_dtc (slope extrapolation of a piecewise-constant median) is noisy:
+  corr with days_to_eol only 0.15, median abs error 13-18d. NOT
+  near-deterministic.
+- sim_k IS sharp where it fires: 99 flagged rows, 57% die in window (25x
+  base rate), median 3 days to death — but recall is only 12% of dying rows
+  and the champion already ranks them high.
+
+## 3. Arms (all kind='logbag' variant='vm7' blend=(arm,1.0) p_star=1.0,
+split_days=3, timing_weight=2.0; paired vs 1541.3; order-only, counts flat)
+
+| arm       | what                                            | SCORE  | delta±SE     | sig | W/L   | wilcoxon | swaps |
+|-----------|--------------------------------------------------|--------|--------------|-----|-------|----------|-------|
+| parity    | order3 re-implementation, no gm                  | 1541.3 | +0.00, maxabs 0.00 | — | — | — | 2.27 |
+| g_hgb     | (a) GM5 added to gap-HGB features                | 1567.6 | +26.3±18.6   | 1.4 | 9/8   | .705     | 2.27  |
+| g_wnr     | (b) WienerFHT with d = gm_gap (fallback v_corr)  | 1608.9 | +67.6±35.1   | 1.9 | 19/20 | .176     | 2.26  |
+| g_dtc405  | (c) 4th source rank(-gm_dtc) w=.05               | 1655.4 | +114.1±33.6  | 3.4 | 11/24 | .002     | 2.28  |
+| g_dtc410  | (c) same, w=.10                                  | 1734.5 | +193.2±47.9  | 4.0 | 16/27 | .001     | 2.25  |
+| g_both    | (d) = a+b                                        | 1639.4 | +98.1±35.0   | 2.8 | 18/24 | .034     | 2.26  |
+| g_ctl     | (e) control: GM5 in logbag order-primary         | 1541.3 | +0.0±35.0    | 0.0 | 17/12 | .670     | 2.26  |
+| g_boost05 | creative: +.05 surgical boost where sim_k<=6     | 1543.1 | +1.8±22.1    | 0.1 | 6/5   | .898     | 2.27  |
+| g_boost10 | creative: same, .10                              | 1549.3 | +8.0±22.8    | 0.3 | 5/7   | .733     | 2.27  |
+
+Per-block deltas: no arm wins nearend s_42-47 either (best g_hgb -1.0,
+noise); losses concentrate in the panel block. Parity control is exact, so
+deltas are pure gm effects. Swaps flat at ~2.27 everywhere.
+
+## 4. Verdict
+
+**Mechanism validated (81/82), delivery comprehensively negative. No
+candidate. Nothing meets any tier bar; do not ship.** The bold idea's
+premise half-holds: the label IS deterministic and now fully reproducible —
+but exactness at the barrier adds no ordering value because the statistic
+lags the raw voltage features the champion already uses (v_min_7 leads it by
+construction: min <= median-of-medians). The interesting control result:
+GM5 inside the logbag order-primary is a perfect wash (+0.0), unlike GAP4
+which cost +123 — gm features are benign but redundant.
+
+Reusable for future work (not this deadline): the exact recipe in section 1
+(e.g., deployment-side exact-crossing detection / label-consistent
+truncation). It is computable in submit.py's split pipeline from raw
+telemetry with ~15 lines (groupby day-median with count filter, reindex,
+rolling(7, min_periods=3).median()).

agent_reports_bgkf/t3_scale_counts.md

@@ -0,0 +1,116 @@
+# t3: count-coupled candidates on the synthetic scaling harness
+
+**Question.** Are the alpha-1.4 candidates' gains stable as fleets grow (more
+batteries, more failures), and does the death-count gate on gated14 bound its
+downside at scale? Robustness layer on top of the bgkf CV (champ 1541.3 /
+flat14 1501.9 / gated14 ~1483).
+
+**Setup.** Harness rebuilt from `agent_reports_bgkf/artifacts/{synth,scale_run}.py`
+into `/tmp/c21/` (the /tmp/teamrepo clone was wiped; team `features.py` now
+imported from `/Users/THIF/repos/BatterySwapAI2026-rul-mgt-zero`). Fleets at
+k=1/2/4 = 461/922/1844 devices via time-shifted building clones, seed 7,
+clone-0 = untouched original; EOLs derived from shifted telemetry by the
+official mechanism — clone-0 check 0 mismatches at every k (75/75 checkable
+deaths). Evaluation is deployment-faithful CURRENT head: one static
+`submit.Ft2RUL` fit on the packaged train table (3-way order .50/.30/.20),
+planner dials `split_days=3, timing_weight=2.0, ec_margin=0.8, ec_pmax=25,
+floor_scale=2.5` with `recent_deaths=0` (static path: train_uniq=82 >=
+FLOOR_GATE=20, floor inert exactly as in deployment). Arms differ ONLY in
+`plan_tsroute.alpha`; features/predictions shared per scenario, so deltas are
+exactly paired. All 48 scenarios per (k, arm). EARLY/LATE wired from each
+scenario's settings.
+
+- champ:   alpha 1.0 everywhere (shipped head)
+- flat14:  alpha 1.4 everywhere
+- gated14: alpha 1.4 while deaths_seen(T) < 20 else 1.0, with deaths_seen(T)
+  = # devices in the planned fleet with recorded eol <= T (plan-time
+  observable, FLOOR_GATE-family rule; no forward reads)
+
+**Anchor.** k=1 champ = 1489.8, far below the old-champion anchor ~1762.5 and
+in the current head's bgkf neighborhood (1541.3; the harness prices derived
+EOLs + a template travel matrix, so a ~50pt offset is expected). champ
+swaps/scen 10.8, swaps/death 1.16, recall 0.569 — deployment-plausible.
+
+## Results (mean fleet total per scenario; delta = candidate - champ, paired)
+
+### k=1 (461 devices)
+| arm | score | delta ± SE | W/L | swaps/scen | late-d/scen | early-d/scen | warmup Δ/scen |
+|---|---|---|---|---|---|---|---|
+| champ | 1489.8 | — | — | 10.8 | 95.4 | 640.4 | — |
+| flat14 | 1388.0 | -101.9 ± 52.9 | 20/23 | 14.3 | 69.8 | 939.9 | -537.9 |
+| gated14 | 1400.2 | -89.6 ± 38.5 | 7/1 (8 nonzero) | 11.5 | 85.3 | 672.3 | -537.9 |
+
+Block scores (champ/flat14/gated14): warmup s0-7 2314/1776/1776, panel s8-41
+1463/1442/1463, nearend s42-47 541/566/541. flat14 Wilcoxon p=0.27 (nonzero
+n=43); gated14 p=0.016 (n=8). flat14 panel Δ -21.8/scen, nearend Δ +25.5/scen.
+
+### k=2 (922 devices)
+| arm | score | delta ± SE | W/L | swaps/scen | late-d/scen | early-d/scen | warmup Δ/scen |
+|---|---|---|---|---|---|---|---|
+| champ | 3142.1 | — | — | 21.2 | 205.4 | 1306.2 | — |
+| flat14 | 2934.6 | -207.5 ± 82.8 | 27/21 | 27.6 | 154.1 | 1937.4 | -741.3 |
+| gated14 | 3054.9 | -87.2 ± 42.9 | 4/0 (4 nonzero) | 21.6 | 195.8 | 1324.9 | -523.0 |
+
+Blocks: warmup 3473/2732/2950, panel 3426/3284/3426, nearend 1092/1221/1092.
+flat14 Wilcoxon p=0.092; panel Δ -141.3/scen, nearend Δ +129.4/scen.
+
+### k=4 (1844 devices)
+| arm | score | delta ± SE | W/L | swaps/scen | late-d/scen | early-d/scen | warmup Δ/scen |
+|---|---|---|---|---|---|---|---|
+| champ | 7053.4 | — | — | 39.5 | 480.4 | 2602.2 | — |
+| flat14 | 6458.8 | -594.6 ± 167.9 | 26/22 | 52.2 | 356.1 | 3910.1 | -2080.1 |
+| gated14 | 7053.4 | 0.0 (identical) | 0/0 | 39.5 | 480.4 | 2602.2 | 0.0 |
+
+Blocks: warmup 8433/6353/8433, panel 7609/7169/7609, nearend 2062/2573/2062.
+flat14 Wilcoxon p=0.0077; panel Δ -440.0/scen, nearend Δ +510.6/scen.
+
+## Gate-closure trajectory (the KEY question)
+
+The gate closes strictly earlier as the fleet grows — deaths accumulate ~k
+times faster while the threshold stays at 20:
+
+| k | deaths_seen at s_0 | gate-active scenarios | gate closes | gated-vs-champ Δ (all-48 mean) |
+|---|---|---|---|---|
+| 1 | 3 | 8 (s_0-s_7) | s_8 (seen 20) | -89.6 ± 38.5 |
+| 2 | 10 | 4 (s_0-s_3) | s_4 (seen 25) | -87.2 ± 42.9 |
+| 4 | 21 | 0 (never opens) | pre-s_0 | 0.0 (byte-identical plans) |
+
+Per-scenario gated deltas on gate-active origins — k=1: s_0 +145.2, s_1
+-820.5, s_2 -215.5, s_3 -683.3, s_4 -754.4, s_5 -439.1, s_6 -1306.2, s_7
+-229.0 (7W/1L). k=2: s_0 -1357.9, s_1 -864.5, s_2 -1000.3, s_3 -961.1
+(4W/0L). Outside the gate window gated14 reproduces champ exactly at every
+scale (identical alpha on identical predictions => identical plans).
+
+## Verdict
+
+- **gated14: STABLE, with downside bounded by construction — converges to
+  champ at large fleets.** The gate is doing exactly what it was designed to
+  do: it fires only in the genuinely count-starved era, that era shrinks as
+  the fleet grows (8 -> 4 -> 0 scenarios at 1x/2x/4x), and at 4x gated14 IS
+  the shipped champ (zero exposure). Where it does fire it captures most of
+  flat14's warmup gain (7W/1L and 4W/0L; single worst scenario +145) while
+  taking none of flat14's panel/nearend exposure.
+- **flat14: mean-IMPROVING with scale, but tail-driven and regime-skewed.**
+  Per-device delta grows (-0.221 -> -0.225 -> -0.322 per device; Wilcoxon p
+  0.27 -> 0.092 -> 0.0077), and at scale the gain spreads from warmup into
+  the panel block (-21.8 -> -141.3 -> -440.0/scen). But the NEAREND block
+  degrades just as systematically (+25.5 -> +129.4 -> +510.6/scen), W/L
+  stays ~even at every scale (20/23, 27/21, 26/22), and it buys the gain
+  with +32% swaps and +50% early-days at 4x. Its mean edge is real in this
+  harness but concentrated in a minority of high-prevalence origins — a
+  higher-variance bet, not a uniform dominance.
+- Interpretation for the ship decision: the scaling axis does NOT expose a
+  hidden failure mode in gated14 — its worst case at any scale is "equals
+  champ". flat14's late-era over-inclusion (nearend +510/scen at 4x) is the
+  scale-amplified liability the bgkf mean hides.
+
+**Caveats.** Derived-EOL harness (ghost low-voltage clones shared equally by
+all arms), template travel matrix, single seed (7) for clone shifts; the
+harness overstates absolute deltas relative to bgkf even at k=1 (-102 here vs
+-39.4 on bgkf for flat14), so read trajectories, not magnitudes. Per the
+tiered adoption bar this is a report-only robustness layer, not an adoption
+adjudication.
+
+**Artifacts.** /tmp/c21/{synth.py,scale3.py,summarize.py},
+/tmp/c21/scale_k{1,2,4}_all.csv, /tmp/c21/fleet_k{1,2,4}.pkl,
+logs /tmp/c21/{synth,run}_k*.log. Repo artifacts untouched.

batteryswap_example/models_cv.py

@@ -1455,7 +1455,12 @@ def plan_tsroute(pred, included, locations, start_time, day_q,
         # (the harness itself drops eol <= origin devices) and estimate the
         # coming window's prevalence without trusting the probability scale.
         psum = max(psum, plan_tsroute.floor_scale * plan_tsroute.recent_deaths)
-    n = int(round(plan_tsroute.alpha * psum))
+    # consume-once appetite boost: armed by the model's predict (T3 warmup
+    # gate) for the plan that immediately follows, then self-resets so a
+    # fallback path can never inherit a stale boost
+    boost = plan_tsroute.alpha_boost
+    plan_tsroute.alpha_boost = 1.0
+    n = int(round(plan_tsroute.alpha * boost * psum))
     order_col = 'rank_score' if 'rank_score' in pred.columns else 'p42'
     chosen = pred[order_col].nlargest(n).index
     if plan_tsroute.p_be > 0:
@@ -1531,6 +1536,7 @@ def plan_tsroute(pred, included, locations, start_time, day_q,
 plan_tsroute.alpha = 0.8
 plan_tsroute.split_days = 0
 plan_tsroute.timing_weight = 1.0   # scales timing curves in optimize_days' objective
+plan_tsroute.alpha_boost = 1.0     # consume-once; see the count line above
 plan_tsroute.p_be = 0.0
 plan_tsroute.floor_day = 0
 plan_tsroute.p_conf = 0.5

batteryswap_example/submit.py

@@ -253,6 +253,47 @@ def _gap_features(frame):
     return out
 
 
+_HZ_BUCKETS = [(0, 7), (7, 14), (14, 21), (21, 28), (28, 42), (42, 63), (63, 84)]
+
+
+class _DiscHaz:
+    """Discrete-time hazard HGB on person-period (pair, bucket) rows from
+    the censored pool of whatever pair table fit() receives (durations
+    censored at its last origin); order source P(die<=42d) = 1 - prod(1-h_k).
+    Censoring is what the plain classifiers can't see - this model's edge is
+    cold-start origins, which is the only place the warmup gate lets it act."""
+
+    def fit(self, train):
+        pool = models_cv.censored_pairs(train, train['origin'].max())
+        self.cols = list(models_cv.FEATURES)
+        dur = pool['duration'].values
+        ev = pool['event'].values.astype(bool)
+        Xb = pool[self.cols].values.astype(float)
+        Xs, ys, mids = [], [], []
+        for lo, hi in _HZ_BUCKETS:
+            died = ev & (dur > lo) & (dur <= hi)
+            use = (dur >= hi) | died
+            Xs.append(Xb[use])
+            ys.append(died[use])
+            mids.append(np.full(int(use.sum()), (lo + hi) / 2.0))
+        from sklearn.ensemble import HistGradientBoostingClassifier
+        self.clf = HistGradientBoostingClassifier(
+            max_depth=3, learning_rate=0.06, max_iter=250,
+            l2_regularization=1.0, random_state=0, early_stopping=False,
+        ).fit(np.column_stack([np.vstack(Xs), np.concatenate(mids)]),
+              np.concatenate(ys))
+        return self
+
+    def predict(self, rows):
+        Xb = rows[self.cols].values.astype(float)
+        n = len(rows)
+        surv = np.ones(n)
+        for lo, hi in _HZ_BUCKETS[:5]:          # buckets covering 0-42d
+            X = np.column_stack([Xb, np.full(n, (lo + hi) / 2.0)])
+            surv = surv * (1.0 - self.clf.predict_proba(X)[:, 1])
+        return pd.DataFrame({'p42': 1.0 - surv}, index=rows.index)
+
+
 class Ft2RUL:
     """Faithful deployment of cv.run_bgkf(kind='logbag', variant='vm7',
     blend=('wiener', 0.20)): primary = SklearnRUL('logbag') fit with
@@ -292,6 +333,28 @@ class Ft2RUL:
                 max_depth=2, learning_rate=0.06, max_iter=300,
                 l2_regularization=1.0, random_state=0
             ).fit(X[self._gap_cols], train['dies42'])
+        self.hazard, self.dev_eol = None, None
+        if ('eol' not in train.columns and 'days_to_eol' in train.columns
+                and 'origin' in train.columns):
+            # the cached parquet ships days_to_eol but not eol (cv.load_table
+            # adds it at load time); the derivation is exact (verified 0 diff
+            # on all 8206 dying rows)
+            train = train.copy()
+            train['eol'] = train['origin'] + pd.to_timedelta(
+                train['days_to_eol'], unit='D')
+        if WARMUP_RATE > 0 and 'eol' in train.columns and 'device_id' in train.columns:
+            try:
+                self._gate_thr = WARMUP_RATE * train['device_id'].nunique()
+                self.dev_eol = (train[['device_id', 'eol']].dropna()
+                                .drop_duplicates('device_id')
+                                .set_index('device_id')['eol'])
+                with self._swap():
+                    self.hazard = _DiscHaz().fit(train)
+            except Exception:
+                # warmup gate is an overlay: a broken hazard fit degrades to
+                # the plain (previous-head) behavior, never to a crash
+                traceback.print_exc()
+                self.hazard = None
         return self
 
     def predict(self, rows):
@@ -304,9 +367,28 @@ class Ft2RUL:
             rows = rows.copy()
             rows['vmin7_rank'] = (rows['v_min_7'].rank(pct=True)
                                   if 'v_min_7' in rows.columns else np.nan)
+        gate = False
+        if (self.hazard is not None and self.dev_eol is not None
+                and 'origin' in rows.columns):
+            # T3 warmup gate: TRAIN-ledger deaths recorded <= this plan origin
+            seen = int((self.dev_eol <= rows['origin'].max()).sum())
+            gate = seen < self._gate_thr
         g = _gap_features(rows) if self.gap is not None else None
         with self._swap():
             pred = self.primary.predict(rows)
+            if gate:
+                try:
+                    rk = self.hazard.predict(rows)['p42'].rank(pct=True)
+                    models_cv.plan_tsroute.alpha_boost = WARMUP_ALPHA
+                    vals = np.sort(pred['p42'].values)[::-1]
+                    out = pred.copy()
+                    out['p42'] = pd.Series(
+                        vals, index=rk.sort_values(ascending=False).index
+                    ).reindex(out.index)
+                    return out
+                except Exception:
+                    traceback.print_exc()
+                    models_cv.plan_tsroute.alpha_boost = 1.0
             rk = pred['p42'].rank(pct=True)
             w3 = FT2_GAPW if g is not None else 0.0
             # round the derived primary weight: 1-0.30-0.20 is 0.4999...94 in
@@ -390,6 +472,24 @@ TIMING_WEIGHT = float(os.environ.get('BATTERYSWAP_TIMING_WEIGHT', '2.0'))
 # the champion. Warm-up CV: 4256 -> 2380 at k=2.5, winning 8/8 origins.
 FLOOR_SCALE = float(os.environ.get('BATTERYSWAP_FLOOR', '2.5'))
 FLOOR_GATE = int(os.environ.get('BATTERYSWAP_FLOOR_GATE', '20'))
+# T3 warmup gate (2026-08-23 night, user CV-first directive): while the
+# TRAINING fleet's ledger shows < WARMUP_GATE recorded deaths at the plan
+# origin (plan-time observable, FLOOR_GATE family; the full train ledger
+# crosses 35 at s_15), the ORDER switches to a censored-pool discrete-time
+# hazard HGB and the appetite is boosted x WARMUP_ALPHA via a consume-once
+# plan_tsroute.alpha_boost. bgkf: gated alpha 1.4 alone -125.4 (4.0sigma,
+# width bowl clean 6..48); + gated hazard order -143.3 (3.8sigma, 16W/0L,
+# Wilcoxon <.001) = CV 1398.0 vs the 1541.3 head. Scale harness: the gate
+# narrows as fleets grow (8/4/0 gate-active scenarios at 1x/2x/4x) so the
+# downside is bounded - on fleets past the gate this is byte-identical to
+# the previous head. WARMUP_GATE=0 disables both channels exactly.
+# The gate threshold is a RATE (deaths per training-fleet device), anchored
+# at 35 deaths / 461 devices: a fixed count fires wider on the CV's 4/5-size
+# fold ledgers than on the full deployment ledger (measured: fixed 35 ->
+# 22 changed origins, 15W/7L, -111.6; rate form matches the width-16 bowl
+# optimum on both). WARMUP_RATE=0 disables the gate entirely.
+WARMUP_RATE = float(os.environ.get('BATTERYSWAP_WARMUP_RATE', '0.0759'))
+WARMUP_ALPHA = float(os.environ.get('BATTERYSWAP_WARMUP_ALPHA', '1.4'))
 # Phase 28: cliff-gated appetite for the static held-out path. The static
 # foreign fit undercounts exactly where a knee is already visible in the raw
 # hourly readings (>=35% of the last 7 days' readings below 2.45 V, >=20
@@ -703,6 +803,9 @@ def adaptive_split_plans(split, split_path, log=print, deadline=None):
                 log(f'  WARNING {scenario["name"]}: ft2 wiener params failed '
                     f'({e}); wiener source dropped this scenario')
         feats = feats.reindex(locs['battery'])
+        # plan-origin timestamp for the T3 warmup gate (the cv table carries
+        # this column natively; the split pipeline adds it here)
+        feats['origin'] = T
 
         pred = model.predict(feats)
         if static_model is not None and cliff_counts is not None: