docs: phase 67 - T4 count-flat campaign (6 lanes, ~50 arms) all rejected with mechanisms; 60d65da = count-flat ceiling

EXPERIMENTS.md

@@ -3456,3 +3456,40 @@ favorable on bgkf warmup. Fourth count-coupled public refutation
 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.

agent_reports_bgkf/t4_ext.md

@@ -0,0 +1,85 @@
+# T4 external-resources lane — final report (2026-08-23)
+
+Question: can EXTERNAL battery datasets or pretrained time-series models
+improve the count-flat ORDER vs head 60d65da (CV 1541.3 on bgkf)?
+
+**Verdict: NO — well-argued and empirically measured negative.** Nothing on
+the external axis reaches the 2σ new-mechanism bar tonight; the one candidate
+that could be built (NASA PCoE transfer GRU) degenerates to a redundant
+voltage-level rank and measures null-to-negative.
+
+## Rules caveat (applies to everything below)
+
+Whether competition rules permit external data / pretrained weights is
+UNKNOWN. Any external candidate — including the NASA-trained GRU measured
+here — would need an explicit rules check before any ship. None qualified on
+merit anyway, so the question is moot tonight.
+
+## Triage table
+
+| Resource | Verdict | Why |
+|---|---|---|
+| NASA PCoE battery aging (phm-datasets S3, 200 MB, no registration) | ACQUIRED + MEASURED — null/negative | 34 cells .mat v5 (scipy.io reads; 26 usable ≥40 discharge cycles). Capacity-fade-per-cycle mapped to our gap-to-2.40V scale; transfer signal collapses to a level detector (see below). |
+| CALCE (CS2 etc., direct zip, no registration) | DEAD in stack | Zips contain .xlsx; openpyxl absent from venv and competition runtime → unparseable. Same per-cycle structure as NASA anyway. |
+| Toyota/Stanford Severson (data.matr.io) | DEAD in stack | .mat v7.3 = HDF5; h5py absent → unreadable. Files ~GB each. Same capacity-per-cycle structure. |
+| Oxford battery degradation (ORA) | TRIAGED OUT | n=8 cells (too tiny to train anything), .mat retrieval path unclear inside budget; cycling structure again. |
+| Sandia / batteryarchive.org | TRIAGED OUT | Download flow not scriptable in budget (account-gated cell downloads); per-cycle capacity structure → same degeneracy as NASA. |
+| Calendar-aging sets (Samsung INR21700 279-cell "multi-stage", NMC/C-SiO 228-cell, Nature Sci Data 2024) | TRIAGED OUT | Closest structurally, but they publish periodic lab check-ups (capacity/OCV curves), NOT continuous daily float-voltage drift; multi-GB (over 2 GB budget); mapping to our task reduces to the same fade-level trajectory. |
+| HF time-series foundation models (amazon/chronos, timesfm, moment, moirai, TTM) | DEAD ON ARRIVAL | `transformers` and `safetensors` absent from venv AND system python; no pip installs allowed; competition runtime = sklearn/torch/lifelines/scipy only. Models cannot even be LOADED, let alone shipped. Not measurable, not shippable. |
+| Competition's own unlabeled telemetry | SKIPPED per mission | Closed, contaminated axis. |
+
+## What was built and measured (NASA transfer)
+
+Pipeline (all in /tmp/c26/, nothing tracked touched):
+- `prep_ours.py` — leak-free lookback windows: per (device, origin 0–47),
+  last 64 calendar days of temp-gated (10–30 °C) daily-median voltage minus
+  2.40, interpolated on the day grid. 22,128 windows, 461 devices.
+- `nasa_train_score.py` — 26 NASA cells' discharge-capacity trajectories
+  mapped to our scale (gap = 0 at 70% initial capacity, fresh gap 0.6 V);
+  time-stretch augmentation f ∈ {1,1.5,2,4,6,8} + 0.015 V noise; windows
+  L=64, label = crosses threshold within 42 units; GRU(1→16)+linear
+  (~1k params, torch-only → runtime-shippable in principle). Trained ONLY on
+  NASA → frozen, fold-safe by construction. Held-out-cell AUC 0.93.
+- `run.py` — head-exact driver (order .50 logbag / .30 wiener / .20 gapHGB,
+  vm7, p*=1.0, split_days=3, timing_weight=2.0; ref asserts 1541.3).
+
+**The decisive diagnostic**: on our windows, Spearman(GRU score, −final gap)
+= 0.999, and slope-probe ramps at equal level get identical scores (fast
+decline 0.031 vs slow 0.031 at gap 0.05). The GRU is a pure LEVEL detector.
+Cause is structural, not a bug: within any time-stretch class of NASA
+capacity trajectories, per-unit fade slope is nearly uniform, so
+"crosses within 42 units" is fully determined by current level — the corpus
+never forces slope-dependence. Broadening the stretch mix (f=1…8, 8× slope
+range) did not change this. So the maximum transferable signal ≈ a monotone
+transform of v_corr_last — the single most-mined feature in this competition
+(v_last, vmin7_rank, wiener FHT all already encode level and better).
+
+## Measured arms (bgkf, paired vs head 1541.3, 28 scored origins)
+
+| Arm | SCORE | Δ ± SE | sig | W/L | Wilcoxon | swaps |
+|---|---|---|---|---|---|---|
+| ext4 — GRU as 4th order source, weights .45/.27/.18/.10 | 1573.8 | +32.5 ± 21.1 (worse) | 1.5σ | 10W/12L | p 0.236 | 2.26 |
+| extemb — 8 GRU hidden dims added to gapHGB feature cols | 1540.2 | −1.2 ± 8.1 | 0.1σ | 4W/2L | p 0.688 | 2.27 |
+
+Count-flatness vs ref: extemb swap counts identical on all origins; ext4
+differs by 1 swap on 1 origin (reorder moved one battery across the planning
+window edge — same magnitude as prior c9 arms at 2.28).
+
+Reading: ext4 actively hurts — a redundant level rank at w=.10 dilutes the
+richer 3-way order (mirrors t_rk4's +22.3 from the c9 wave). extemb is a
+pure null — HGB absorbs/ignores the level-transform columns (which are also
+NaN on the 72% synthetic training rows). Neither is close to the 2σ +
+controls bar for a new mechanism. Nothing to ship.
+
+## Tier verdict
+
+REJECT the external axis for this competition. The failure is structural,
+not effort-bounded: every accessible public dataset is cycle/calendar aging
+whose mapping onto "daily float voltage drifting to a hard 2.40 V threshold"
+collapses to fade-level — information the order already exploits to
+saturation. Pretrained TS foundation models are unloadable in the allowed
+stack. Recommend no further spend here before the deadline.
+
+Artifacts: /tmp/c26/{prep_ours.py, nasa_train_score.py, run.py,
+ext_scores.csv, gru_nasa.pt, cv_bgkfcand_ext4.csv, cv_bgkfcand_extemb.csv,
+ours_windows.npz}; NASA raw in /tmp/c26/nasa/.

agent_reports_bgkf/t4_lit.md

@@ -0,0 +1,115 @@
+# T4 LITERATURE lane — literature-derived count-flat order features
+
+Date: 2026-08-23 (deadline day). Baseline: head config 60d65da = CV 1541.3
+on bgkf (`/tmp/ref_head_1541.csv`, per-origin paired, 48 origins).
+Work dir `/tmp/c25/` (build_lit.py, run.py, lane.log). No repo files touched.
+
+## Literature sweep (~20 min, 5 searches)
+
+What the battery-RUL/PHM literature suggests that is computable from a daily
+median-voltage series in <=30 lines, mapped to our two known failure modes
+(misses ranked 3-4 too low with signal present; plateau devices at
+2.41-2.44 V, 40+ days, only 36% die):
+
+1. **Knee-point / broken-stick detection** — Bacon-Watts and two-segment
+   fits are the standard for locating the transition from slow to fast
+   degradation; robust to noise, no capacity-fade model needed.
+   - Fermin-Cueto et al. (knee-point identification, Bacon-Watts),
+     via [MDPI Batteries 11(6) 221](https://doi.org/10.3390/batteries11060221)
+     and [knee-onset ID, IOPscience](https://iopscience.iop.org/article/10.1088/3049-4761/ae5067)
+   - [Segmented regression](https://en.wikipedia.org/wiki/Segmented_regression)
+   - Implemented: `bw_tau` (days-before-origin of best two-segment break on
+     the last 120 d, O(n) prefix-sum SSE grid), `bw_slope2` (post-knee
+     slope), `bw_dslope` (knee sharpness). Targets misses whose recent
+     regime change is diluted in the fixed-window slopes.
+
+2. **Sample entropy of voltage** — SampEn of discharge voltage is a
+   recurring SOH indicator (Widodo et al.; RVM prognostics).
+   - [Sample-entropy prognostics with RVM, J. Manuf. Syst.](https://www.sciencedirect.com/science/article/abs/pii/S0278612521000741)
+   - Implemented: `sampen_30` (m=1, r=0.2*std, last 30 daily values).
+
+3. **Monotonicity metric (PHM prognostic-parameter suitability, Coble &
+   Hines family)** — Spearman rho of voltage vs time; a plateau device is
+   flat (rho ~ 0) while a real decliner is monotone (rho ~ -1). Directly
+   attacks the plateau false-include mode.
+   - [Data-driven prognostics based on health indicators (HAL)](https://hal.science/hal-00869489/document)
+   - Implemented: `mono_60` (Spearman rho over last 60 d).
+
+4. **Relaxation/recovery statistics** — Zhu et al., Nat Comm 2022: variance,
+   skewness, maxima of the voltage-relaxation curve predict capacity; the
+   recovery-from-minimum idea transfers as rebound features.
+   - [Zhu et al. 2022, Nature Communications](https://www.nature.com/articles/s41467-022-29837-w)
+   - Implemented: `reb_30 = v_last - v_min_30`, `reb_frac` (rebound as a
+     fraction of the drop from baseline), `skew_30` (skewness of last 30
+     daily values).
+
+5. **Variance-of-degradation** — float-voltage spread/variance rises toward
+   EOL in stationary lead-acid fleets.
+   - [Lead-acid EOL field discussion](https://forum.solar-electric.com/discussion/354321/learning-to-identify-end-of-life-for-lead-acid-batteries),
+     [multi-phase Wiener RUL for lead-acid](https://www.sciencedirect.com/science/article/abs/pii/S0957582025002411)
+   - Implemented: zero-cost arm, `vstd_30`/`rng_30` already on the table
+     (never tried inside the gap-HGB; dead only as GLOBAL logbag features).
+
+Considered and skipped: label-statistic smoothers (banned — lagging),
+cohort/batch clocks (already adjudicated), CNN/GPR/RVM model swaps (not
+count-flat order deliverables in the time budget), ranking-under-censoring
+methods (method-level, no <=30-line feature).
+
+## Delivery vehicle
+
+All features delivered ADDED to the gap-HGB order source of the shipped
+3-way order (.50 logbag(vm7) + .30 wiener + .20 gapHGB), champion HGB
+architecture identical (depth2 lr.06 300it seed0), only cols change.
+Feature pass strictly as-of-origin (`day <= origin`), same in-band
+daily-median series as features.py, all 168 origins (real + synthetic),
+cached to /tmp/c25/lit_features.parquet (77 299 rows).
+
+Parity arm `t_head` re-runs the exact head config against the reference.
+
+## Feature diagnostics (closed pairs, before any CV run)
+
+Direction-free single-feature AUC vs dies42: bw_slope2 0.75 (reversed),
+mono_60 0.65, reb_frac 0.66, vstd_30/rng_30 0.82/0.83, sampen_30 0.53
+(dead), skew_30 0.56. On the plateau cohort (v_corr_last 2.41-2.44,
+n=636, death rate 0.37): bw_slope2 0.63 and mono_60 0.63 are the best
+discriminators — the literature signals DO exist in this data, on the right
+cohort. They just do not convert to plan cost (below).
+
+## Results (paired vs head 1541.3 per origin, 48 origins; positive = worse)
+
+| arm    | delivery              | SCORE  | delta±SE     | sig | W/L    | wilcoxon | swaps |
+|--------|-----------------------|--------|--------------|-----|--------|----------|-------|
+| t_head | parity (exact head)   | 1541.3 |  +0.0± 0.0   | 0.4 |  1W/1L | nan      | 2.27  |
+| t_bw   | gapHGB + bw_tau/slope2/dslope | 1573.4 | +32.1±17.2 | 1.9 | 8W/9L | 0.159 | 2.28 |
+| t_ent  | gapHGB + sampen_30    | 1555.9 | +14.6±19.7   | 0.7 |  8W/4L | 1.000    | 2.27  |
+| t_mono | gapHGB + mono_60      | 1576.2 | +34.8��21.3   | 1.6 |  7W/9L | 0.348    | 2.27  |
+| t_reb  | gapHGB + reb_30/reb_frac/skew_30 | 1549.7 | +8.3±12.5 | 0.7 | 7W/6L | 0.893 | 2.27 |
+| t_var  | gapHGB + vstd_30/rng_30 | 1567.5 | +26.1±21.7 | 1.2 | 10W/11L | 0.432   | 2.28  |
+| t_lb_reb | logbag primary + reb family (anchored control) | 1602.9 | +61.6±26.3 | 2.3 | 13W/18L | 0.079 | 2.27 |
+
+Anchored control behaves exactly as predicted: the least-bad HGB family
+(reb) moved into the logbag order-primary loses +61.6 at 2.3sig — same
+failure shape as GAP4 in logbag (+123). The instrument is measuring what it
+should.
+
+Count-flat check (per (origin, fold) n_swaps vs reference, 240 cells):
+t_head/t_ent/t_mono cell-identical; t_bw/t_reb/t_var differ in 3-4 cells by
+max ±1 (break-even top-up wiggle). All arms are order-only as designed.
+
+Combos (t_bwmono, t_all, ...) were pre-registered but NOT run: every
+single-family arm lost, so stacking them cannot clear the bar; deadline
+budget went to the anchored control instead.
+
+## Verdict
+
+REJECT — tier: no adoption, report-only. Zero of five literature arms wins;
+best is t_reb at +8.3±12.5 (noise-level LOSS), worst are the two arms whose
+features had the best plateau AUC (t_bw +32.1 at 1.9sig, t_mono +34.8 at
+1.6sig — significant in the WRONG direction). The pattern repeats the GAP4
+logbag lesson from the other side: features with real cohort-level AUC
+dilute the depth-2 HGB's splits on the already-validated gap quartet and
+push the marginal ranking around at exactly the rank-gap 1-10 positions
+where the cost lives. The gap-HGB at its shipped weight (.20) appears
+saturated; the head config's feature set survives another five-way attack
+unchanged.
+

agent_reports_bgkf/t4_seq.md

@@ -0,0 +1,86 @@
+# T4 SEQUENCE lane: GRU temporal encoder on raw daily telemetry (NEGATIVE)
+
+Date 2026-08-23. Baseline: head 60d65da (order .50 logbag(vm7) + .30 wiener
++ .20 gapHGB; values logbag @ alpha 1.0; planner tw 2.0 / split 3), reference
+per-cell CSV /tmp/ref_head_1541.csv, CV 1541.3 on bgkf. All arms paired vs
+that reference on the 48 all-building scenario totals.
+
+## Hypothesis
+
+The whole stack is snapshot features; degradation SHAPE (knees, dips,
+recovery, plateau texture) in the raw daily series is unexploited. A small
+recurrent encoder (CMAPSS / battery-SoH literature) on the last 60 days of
+daily telemetry as a new order source.
+
+## Method
+
+- Sequences: per (device, origin) row of the feature table, the 60 calendar
+  days ENDING at the origin (leak spot-checked). 4 channels: daily median
+  v_corr (v - 0.005*(t-20), last-obs-carried-forward, fixed physical
+  normalization - nothing fitted), observed mask, days-since-last-obs
+  (GRU-D-style delta, /90), temp-gated in-band coverage mask. Built once for
+  all 71,878 rows (no cross-device stats, so one cache serves every fold).
+  Builder: /tmp/c22/build_seqs.py -> /tmp/c22/seqs.npz.
+- Encoder: GRU(4 -> hidden 48, 2 layers, dropout 0.1) -> linear logit on
+  dies42. BCE, pos_weight = neg/pos, positives urgency-weighted (RankNet
+  'days' precedent). Adam lr 3e-3, wd 1e-3, 30 epochs, batch 1024.
+- Fold safety: exact replication of run_bgkf's split (cv.build_folds) and
+  train subset (closed rows, fold-train buildings only). NO self-supervised
+  pretraining, no test-fold telemetry in any fit (phase-51 lesson honored).
+  Seed ensemble = 5 seeds, combined as mean of within-fleet pct ranks.
+  Fitter: /tmp/c22/fit_seq.py -> /tmp/c22/seq_scores_f{0..4}.parquet.
+- Measurement: cv.run_bgkf(kind='logbag', variant='vm7', blend=(arm,1.0),
+  p_star=1.0, split_days=3, timing_weight=2.0) - value-preserving order
+  mapping, counts fixed. Driver /tmp/c22/run.py; per-cell CSVs
+  /tmp/cv_bgkfcand_c22_*.csv.
+- Parity control: re-implemented champion order reproduces the reference
+  EXACTLY (+0.0, 46/48 cells identical), so arm deltas are clean.
+
+## Encoder quality (out-of-fold, test-building closed real-origin rows)
+
+| fold | seq AUC | -v_last AUC | pos |
+|------|---------|-------------|-----|
+| 0 | 0.951 | 0.944 | 123 |
+| 1 | 0.942 | 0.946 |  91 |
+| 2 | 0.939 | 0.964 |  80 |
+| 3 | 0.908 | 0.779 |  71 |
+| 4 | 0.956 | 0.954 |  61 |
+
+The encoder is real (it generalizes across buildings; fold 3 shows it beats
+the raw level where v_last fails) but overall only ~matches a single strong
+snapshot feature.
+
+## Results (paired vs head 1541.3; 48 scenarios)
+
+| arm | order | SCORE | delta +/- SE | sig | W/L | wilcoxon | swaps |
+|-----|-------|-------|--------------|-----|-----|----------|-------|
+| champ (parity) | .50/.30/.20/- | 1541.3 | +0.0 +/- 0.0 | 0.4 | 1W/1L | nan | 2.27 |
+| (d) seq_alone | 0/0/0/1.0 | 1935.7 | +394.4 +/- 71.3 | 5.5 | 9W/39L | 0.000 | 2.26 |
+| (a) seq_g20 | .50/.30/0/.20 | 1654.3 | +113.0 +/- 36.5 | 3.1 | 14W/21L | 0.017 | 2.27 |
+| (c) seq_w30 | .50/0/.20/.30 | 1613.8 | +72.5 +/- 27.1 | 2.7 | 20W/22L | 0.089 | 2.25 |
+| (b) seq_4th | .45/.27/.18/.10 | 1584.1 | +42.8 +/- 20.4 | 2.1 | 9W/14L | 0.140 | 2.27 |
+| dose ctrl | .475/.285/.19/.05 | 1562.9 | +21.6 +/- 16.9 | 1.3 | 6W/7L | 0.508 | 2.27 |
+
+All arms count-flat (2.25-2.27 vs ref 2.27). Dose-response is MONOTONE
+against the source: 0 -> +0.0, .05 -> +21.6, .10 -> +42.8, .20 -> +113.0.
+Every insertion point (gap slot, wiener slot, 4th source, any weight) makes
+the order worse; the losses are significant in the wrong direction.
+
+## Verdict
+
+REJECTED (new-mechanism tier needs >= 2 sigma IMPROVEMENT; we have 2-5 sigma
+degradation). The temporal-shape signal the GRU extracts is already captured
+by the snapshot stack (slopes, accel, drop_recent, wiener drift, gap
+features); at the margin the encoder adds noise, not recall. Standalone
+anchor (1935.7) confirms it is far from champion-grade on its own.
+
+Do NOT retest: GRU/temporal encoder on daily (v_corr, mask, delta, in-band)
+sequences as an order source at any slot/weight in the 3-way order. A
+materially different attempt would need a different signal (e.g. sub-daily
+texture), not different weights.
+
+## Deployment cost (moot, for the record)
+
+Fit: ~7 s/epoch uncontended at 2 torch threads -> ~3.5 min/seed, ~17.5 min
+for the 5-seed ensemble (deployment fits once on full train); inference on
+the whole table <10 s. Inside the 30-min budget, but not shipped.

agent_reports_bgkf/t4_stack.md

@@ -0,0 +1,86 @@
+# T4 stacking lane: learned meta-combiner over the 3-source order — REJECTED
+
+Date: 2026-08-23. Lane: STACKING (T4 count-flat campaign).
+Baseline: head config 60d65da, CV 1541.3 on bgkf (`/tmp/ref_head_1541.csv`,
+parity re-verified through this lane's own driver). Champion order =
+fixed-weight rank blend .50*rank(logbag vm7) + .30*rank(WienerFHT) +
+.20*rank(gapHGB depth2 on FEATURES+vmin7_rank+GAP4).
+
+## Question
+
+Can a LEARNED fold-safe meta-combiner (stacked generalization) beat the
+fixed .50/.30/.20 weights by knowing per-device WHEN to trust which source
+(context: wnr_n, staleness, gap_ratio, vmin7_rank; disagreement stats)?
+
+## Method
+
+Driver `/tmp/c24/run.py` (diagnostic `/tmp/c24/diag.py`). Inside each outer
+fold's fit(train): fold-train split by BUILDING into K=4 inner folds
+(deaths-sorted snake, deterministic; the established contamination rule —
+never by row); out-of-fold predictions of all 3 sources on every train row,
+pct-ranked WITHIN origin_idx groups (the deployment-time analog: the test
+fleet is ranked at one origin, ~5 buildings — inner-test is also ~5
+buildings, so rank distributions match); meta fit on
+[r1,r2,r3, spread, |ri-rj|, wnr_n, staleness, gap_ratio, vmin7_rank] ->
+dies42; sources refit on full fold-train; meta combines their test ranks.
+Order-only: measured via cv.run_bgkf(kind='logbag', variant='vm7',
+blend=(arm,1.0), p_star=1.0, split_days=3, timing_weight=2.0) — blend w2=1.0
+reassigns the primary's sorted p42 values onto the meta's order, so the
+count machinery is untouched. Paired vs the 48 per-scenario reference
+totals. cv.py/models_cv.py md5-pinned per run (race guard); no repo files
+touched.
+
+## Results (paired vs head 1541.3, 48 scenarios)
+
+| arm | meta | SCORE | delta±SE | sig | W/L | wilcoxon | swaps |
+|---|---|---|---|---|---|---|---|
+| par (parity) | fixed .50/.30/.20 | 1541.3 | +0.0±0.0 | 0.4σ | 1W/1L | – | 2.27 |
+| a_log | logistic, ranks+disagree+context | 1604.5 | +63.2±44.8 | 1.4σ | 19W/25L | p0.208 | 2.26 |
+| b_hgb | HGB depth2, same features | 1680.5 | +139.2±61.6 | 2.3σ | 17W/31L | p0.020 | 2.28 |
+| c_ranks | HGB depth2, 3 ranks only | 1613.2 | +71.9±46.5 | 1.5σ | 17W/26L | p0.113 | 2.30 |
+| d_cvx | convex weights, simplex grid by OOF AUC | 1598.6 | +57.3±41.0 | 1.4σ | 21W/21L | p0.620 | 2.28 |
+| e_logrk | logistic, 3 ranks only | 1619.1 | +77.8±44.0 | 1.8σ | 17W/24L | p0.257 | 2.26 |
+| e_log50 | 0.50*champ + 0.50*rank(a_log) | 1594.8 | +53.4±27.5 | 1.9σ | 13W/26L | p0.065 | 2.26 |
+| e_log75 | 0.75*champ + 0.25*rank(a_log) | 1557.8 | +16.5±19.6 | 0.8σ | 16W/11L | p0.838 | 2.27 |
+
+All deltas positive = WORSE than head. Every learned arm loses; b_hgb is a
+2.3σ significant LOSS. Swaps flat everywhere (2.26-2.30 vs 2.27) — the
+mechanism is count-flat as designed; the losses are pure ordering damage.
+
+## Why it fails (diagnostic, per outer fold)
+
+Per-source inner-OOF AUCs: lb 0.932-0.944, wiener 0.919-0.947, gap
+0.941-0.955 — gapHGB is the BEST dies42 classifier in every fold. So any
+classification-trained meta over-weights it:
+
+- d_cvx learned weights (lb, wn, gap): (.15,.45,.40), (.10,.25,.65),
+  (.05,.30,.65), (.30,.30,.40), (.25,.20,.55) — gap at 2-3x its champion
+  weight, logbag starved. The champion .50/.30/.20 is IN the grid and the
+  learned points beat it on inner-OOF AUC in all 5 folds (+0.001 to +0.008)
+  — yet cost says the opposite. Objective mismatch, not estimation noise.
+- a_log coefficients: r3 (gap) largest in 4/5 folds; context features get
+  small/unstable coefficients (vmin7_rank negative only because it double-
+  counts with the sources) — context adds nothing (a_log ~ e_logrk ~ c_ranks).
+- Dose-response is monotone: 25% meta +16.5, 50% +53.4, 100% +63.2. The
+  learned order contributes no positive increment at ANY mixing weight.
+
+Root cause: dies42-AUC/logloss over the closed-pair corpus is dominated by
+separating obvious survivors from obvious deaths, where gapHGB excels. Plan
+cost is decided by the fine ordering near the swap-inclusion cutoff and its
+timing interplay, where logbag's calibrated top ranks matter — exactly what
+the fully-mapped cost simplex (.50/.30/.20 optimal) already encodes. A
+fold-safe meta cannot see cost, so it reliably walks AWAY from the cost
+optimum toward the classification optimum (~.15/.30/.55 region).
+
+## Verdict
+
+Tier: REJECTED, mechanism dead — do not adopt, no arm to escalate.
+Stacked generalization over these 3 sources is structurally mispriced under
+any label-classification objective; the fixed .50/.30/.20 cost-mapped
+weights stand. Do NOT revisit with a different meta-model family (HGB,
+logistic, constrained-convex, ranks-only, context, shrinkage all measured
+here); the only untried fix — per-fold weights tuned on plan cost — is the
+already-dead static-simplex search wearing fold noise.
+
+Artifacts: /tmp/c24/run.py, /tmp/c24/diag.py, /tmp/c24/*.out,
+/tmp/cv_bgkfcand_c24_*.csv.

agent_reports_bgkf/t4_surv.md

@@ -0,0 +1,90 @@
+# T4 survival/ranking campaign: literature-grade survival & LTR as order sources
+
+**Date:** 2026-08-23 (deadline day). **Agent:** SURVIVAL/RANKING lane, /tmp/c23/.
+**Verdict up front: comprehensive REJECT — all 13 arms lose to the head, most at 2.6–3.3σ. No adoption candidate. All arms count-flat (swaps 2.27±0.02).**
+
+## Setup
+
+- Baseline: head config 60d65da, CV 1541.3 on bgkf; reference per-cell CSV
+  `/tmp/ref_head_1541.csv` (assert passed: groupby-origin mean within 1.0).
+- Measurement: `cv.run_bgkf(kind='logbag', variant='vm7', blend=(name,1.0),
+  p_star=1.0, split_days=3, timing_weight=2.0)` — value-preserving rank blend,
+  counts untouched. Paired per-origin deltas over 48 all-building totals.
+- Champion order = .50·rank(logbag vm7) + .30·rank(WienerFHT) + .20·rank(gapHGB).
+- **Parity control:** a from-parts recomposition of the head order (`ctrl`)
+  scored 1541.3, delta +0.0 (1W/1L, ties elsewhere) — the composer is faithful,
+  so every arm delta is attributable to the swapped-in source alone.
+- Fold safety: supervised sources fit on the fold-train closed table run_bgkf
+  hands them; Cox-family censored pools built exactly per cv.py's hazard-branch
+  recipe (`mc.censored_pairs(FULL, max_origin)` restricted to fold-train devices).
+- Driver + sources: `/tmp/c23/run.py`; per-cell CSVs `/tmp/c23/cv_*.csv`.
+
+## Order sources tested (methods, not features)
+
+- **cox** — lifelines CoxPHFitter (penalizer 1.0 / 0.1) on FEATURES+vmin7_rank
+  (median-impute + standardize), proper censoring from the censored-pairs pool;
+  order = log partial hazard.
+- **lcw** — cost-aware LISTWISE ranker (torch, 5-seed bag, MLP 32/16, 300 ep):
+  softmax loss over per-origin groups, each positive's gain = expected late-day
+  cost `10·(42 − days_to_eol)` (floor 5). Meaningfully different from the dead
+  plain RankNet: listwise normalization over the whole origin group + linear
+  cost gains (0–420) instead of pairwise BCE with 1–2× weights.
+- **mgb** — GBM-Cox hybrid (two-stage boosting-for-survival): HGB regressor
+  (3-seed bag, depth 2, lr .06, 300 iter) on lifelines martingale residuals;
+  order = log-partial-hazard + HGB(residual). (Residuals realigned from
+  lifelines' duration-sorted output — verified.)
+- **\*g4** variants — same sources given the four gap-approach columns
+  (gap_ratio, rel_close, dwell_flat, gap_log), closing the features-vs-method
+  confound in the gap slot.
+
+## Results (paired vs head 1541.3; 48 origins; swaps target ~2.27)
+
+| arm | slot | SCORE | delta±SE | sig | W/L | wilcoxon | swaps |
+|---|---|---|---|---|---|---|---|
+| ctrl (parity) | head recomposed | 1541.3 | +0.0±0.0 | 0.4σ | 1W/1L | — | 2.27 |
+| cox1_gap | gap .20 | 1611.7 | +70.4±24.2 | 2.9σ | 13W/25L | .019 | 2.27 |
+| cox01_gap | gap .20 | 1614.0 | +72.7±24.9 | 2.9σ | 13W/19L | .065 | 2.27 |
+| coxg4_gap | gap .20 (+GAP4) | 1648.4 | +107.1±32.7 | 3.3σ | 12W/26L | .001 | 2.28 |
+| cox1_4th | 4th .10 | 1561.2 | +19.9±13.5 | 1.5σ | 11W/9L | .571 | 2.27 |
+| cox1_alone | alone | 1665.5 | +124.2±46.8 | 2.7σ | 13W/33L | .006 | 2.24 |
+| lcw_gap | gap .20 | 1618.5 | +77.2±28.2 | 2.7σ | 8W/26L | .003 | 2.27 |
+| lcwg4_gap | gap .20 (+GAP4) | 1619.3 | +78.0±29.6 | 2.6σ | 10W/24L | .011 | 2.27 |
+| lcw_4th | 4th .10 | 1578.2 | +36.9±21.5 | 1.7σ | 12W/14L | .340 | 2.27 |
+| lcw_alone | alone | 1710.4 | +169.1±61.5 | 2.7σ | 18W/29L | .012 | 2.25 |
+| mgb_gap | gap .20 | 1651.5 | +110.2±34.2 | 3.2σ | 7W/23L | .001 | 2.29 |
+| mgbg4_gap | gap .20 (+GAP4) | 1650.3 | +109.0±36.4 | 3.0σ | 5W/28L | .001 | 2.29 |
+| mgb_4th | 4th .10 | 1571.2 | +29.9±21.0 | 1.4σ | 7W/11L | .181 | 2.27 |
+| mgb_alone | alone | 1625.0 | +83.7±47.4 | 1.8σ | 20W/27L | .098 | 2.26 |
+
+(All deltas are LOSSES; positive = worse than head.)
+
+## Reading
+
+1. **The gap slot's signal is the objective, not the features.** With identical
+   GAP4 columns, Cox (+107), listwise-cost (+78) and martingale-GBM (+109) all
+   lose the slot by 2.6–3.3σ; the dies42 classification HGB is doing real,
+   method-specific work there.
+2. **Proper censoring buys nothing here.** Cox partial hazard alone (+124) is
+   far behind the champion composite; the closed-window dies42 label already
+   captures what the 42d planning problem needs. Martingale-residual boosting
+   does add order signal over pure Cox (alone: +84 vs +124) but never enough.
+3. **Cost-weighted listwise is the weakest source** (alone +169) despite being
+   aligned with the real cost objective — with ~1.2k positives across ~157
+   origin-groups the listwise gradient is too diffuse; the champion's bagged
+   logistic + FHT prior generalize better at this n.
+4. **Dilution monotonicity:** every source's damage scales with its weight
+   (alone > gap .20 > 4th .10), and no 4th-slot arm shows even a directional
+   win (best: cox1_4th +19.9±13.5, 11W/9L, wilcoxon .571 — noise-shaped loss).
+
+## Tier verdict
+
+Adoption bar (tiered, user-approved): new mechanisms need ≥2σ paired WINS with
+controls. **Nothing qualifies �� every arm is a loss; the best is a 1.4σ loss.**
+No follow-up recommended on this axis.
+
+DEAD (do not retest): Cox partial hazard as order source (any penalizer, any
+slot, ±GAP4), cost-weighted listwise/softmax ranker (per-origin groups, late-day
+gains, any slot, ±GAP4), GBM-on-martingale-residuals hybrid (any slot, ±GAP4).
+Together with the earlier kills (discrete-time hazard HGB, AFT-lite+IPCW, plain
+RankNet, imminence-weighted gap-HGB) this closes the survival/LTR order-source
+axis: **the champion 3-way order stands.**

agent_reports_bgkf/t4_transplant.md

@@ -0,0 +1,106 @@
+# T4 TRANSPLANT — rank-4 competitor mechanisms on the head config
+
+**Agent:** transplant lane, T4 count-flat campaign. **Date:** 2026-08-23.
+**Baseline:** head 60d65da (3-way order .50 logbag / .30 wiener / .20 gap-HGB,
+alpha 1.0, split_days 3, timing_weight 2.0) = **CV 1541.3** on bgkf.
+Reference per-cell CSV `/tmp/ref_head_1541.csv` (asserted 1541.31).
+**Parity:** the order3 replica reproduced the deployed head exactly
+(champ arm: 1541.3, delta +0.0, swaps 2.27) — every arm below is paired
+against a validated baseline.
+
+**Source:** rank-4 competitor's readable source (public 1407.63 vs our
+lineage 1575.31), cached `/tmp/hf_battery3/carl/` (`hierarchical_fpt.py`,
+`identity_ensemble.py`, `competition_planner.py`).
+**Code:** `/tmp/c27/` (precompute.py, aft_precompute.py, run.py, run_aft20.py);
+per-cell CSVs `/tmp/cv_bgkfcand_c27_*.csv`. No tracked file touched,
+monkeypatch only, no commits.
+
+All arms are ORDER-ONLY: blend value-preservation keeps the logbag p42
+multiset, so the count machinery is untouched (swaps 2.26–2.30 vs head 2.27).
+Fold-safety: all fits on fold-train only; pools within fold-train; the
+364-day seasonal lag reads LAST year (legal); frozen constants
+(beta 0.00549 V/degC) are not fits.
+
+## Results (paired vs head, 48 all-building scenario totals)
+
+| arm | mechanism | SCORE | delta±SE | sig | W/L | wilcoxon | swaps |
+|---|---|---|---|---|---|---|---|
+| champ | parity replica | 1541.3 | +0.0±0.0 | 0.4 | 1W/1L | – | 2.27 |
+| geo05 | geography tilt w=.05 | 1676.5 | **+135.2±38.5** | 3.5 | 7W/26L | .000 | 2.28 |
+| geo10 | geography tilt w=.10 | 1853.5 | **+312.2±60.6** | 5.2 | 6W/37L | .000 | 2.29 |
+| geo20 | geography tilt w=.20 | 1993.1 | **+451.8±65.3** | 6.9 | 5W/41L | .000 | 2.30 |
+| geomult | rank(p42pctl·(250+E_b)) | 1944.2 | **+402.9±68.7** | 5.9 | 2W/36L | .000 | 2.29 |
+| seas_hgb | seasonal feats in gap-HGB | 1585.9 | +44.6±26.5 | 1.7 | 9W/14L | .160 | 2.28 |
+| seas_p05 | seasonal gated post w=.05 | 1541.3 | +0.0±0.0 | 0.4 | 1W/1L | – | 2.27 |
+| seas_p10 | seasonal gated post w=.10 | 1541.3 | −0.0±0.0 | 1.0 | 2W/0L | – | 2.27 |
+| aft10 | FP-lifetime AFT 4th @.10 | 1540.2 | −1.1±0.8 | 1.4 | 6W/5L | .520 | 2.26 |
+| aft20 | FP-lifetime AFT 4th @.20 | 1559.7 | +18.4±15.6 | 1.2 | 13W/12L | .895 | 2.27 |
+| carlw30 | their EB wiener in .30 slot | 1669.7 | **+128.4±43.7** | 2.9 | 14W/27L | .008 | 2.27 |
+| carl4_15 | their FPT residual @.15 | 1540.0 | −1.3±8.0 | 0.2 | 6W/6L | .970 | 2.27 |
+| carl4_20 | their FPT residual @.20 | 1566.0 | +24.6±17.5 | 1.4 | 10W/9L | .595 | 2.27 |
+
+## Tier verdict: NOTHING ADOPTS
+
+No arm reaches the 1.5-sigma count-flat-tweak bar, let alone the 2-sigma +
+controls bar for new mechanisms. Best arms (aft10 −1.1±0.8, carl4_15
+−1.3±8.0) are washes; four arms are significant LOSSES.
+
+## Per-diagnostic notes
+
+**Geography-aware ordering tilt (their emergency-cost pricing).**
+Implemented planner-side via the `rank_score` column in `plan_tsroute`
+(top-n SET changes, n itself unchanged). E_b = both depot legs +
+building/room/battery service time + overtime share + daily penalty,
+exactly their `mean_dedicated_trip_hours` geometry. Monotonically
+catastrophic in w (+135 → +452), losses near-uniform across scenarios, so
+the peer's fallback (their 007 gate: apply only when fleet mean
+dedicated-trip cost ≤ 8h) is settled without a run: **all five fold-fleets
+pass the gate** (means 1.9–7.7h), so gated ≡ ungated ≡ rejected. Reading:
+under a fixed count, tilting marginal inclusions toward expensive-to-reach
+buildings evicts true positives; geography does not price OUR misses
+because the planner already routes and times whatever set it is given.
+Their pricing pays off only jointly with their bigger selection set.
+
+**Seasonal temperature signal (their one novel idea).** Full construction
+transplanted (health = V − 0.00549·(T−20), 90d drift slope clipped ≤0,
+42d forecast + same-device temp 364d ago re-added, re-smoothed, min,
+gate pmin ≤ 2.50V). Coverage: strict machinery valid on 54% of
+device-origins; the ≤2.50V gate fires on 9.2%. As a gated within-set
+permutation (their assign_multiset delivery) it is a **cost no-op at
+w=.05/.10**: the gated devices are already at the very top of the champion
+order, so re-permuting them changes no plan. As gap-HGB features
+(seas_pmin_rx, seas_dtemp) it is a 1.7-sigma LOSS (+44.6) — the signal the
+gate carries is already saturated by v_corr/tau_projected; what is left is
+noise the HGB overfits. Their novelty does not transfer because our stack
+already prices temperature (v_corr_*) and terminal proximity.
+
+**Their EB Wiener (12 weekly increments, leave-building-out pool,
+1.4826·MAD diffusion).** As the .30-slot replacement: 2.9-sigma LOSS
+(+128.4). Our cached per-device MLE (180d window, K=6 shrinkage toward
+population median) is simply a better drift estimator here than their
+12-week median-increment with EB shrinkage toward a fold-train pool.
+As their literal gated 15% rank residual on top of the champion: wash
+(−1.3±8.0); at 20%: negative. Their Wiener is a weaker sibling of a slot
+we already occupy with a stronger implementation.
+
+**AFT on first-passage lifetimes (left-truncated log-normal).** Faithful
+implementation confirmed: our fold fits give b0≈0.86–1.22, b1≈0.84–0.89,
+sigma≈0.22–0.24 vs their published train fit b0=0.874, b1=0.889,
+sigma=0.233; 79 event devices among reliable rows (theirs: 76). This is
+NOT our dead AFT-on-observed-EOL — different target, truncation handled.
+Still: −1.1±0.8 at their weight .10 (1.4 sigma, trivially small because the
+gated set again overlaps the already-selected top), and pushing the weight
+to .20 flips it negative (+18.4). Correct sign, no adoptable magnitude.
+
+## Is their 1407.63 method or operating point?
+
+**Operating point.** Every mechanism in their readable source, transplanted
+at matched swap count onto our order, is a wash or a loss (best −1.3,
+sub-0.5-sigma). Their board profile — 743 swaps / 3349 late-days vs our
+644 / 5009 — says they buy ~99 extra swaps and convert them at ~16.8
+late-days per marginal swap (ours convert at ~6.0 near our cutoff). That
+conversion rate is only reachable with more inclusions on a fleet whose
+tail we demonstrably cannot rank into the top-644; no count-flat ordering
+graft closes it. The appetite axis is CLOSED by user directive, so this is
+a diagnosis, not an action item: their gap is appetite + selection-set
+size, not transferable ordering signal.