Campaign 2 CLOSED: all 7 lanes final; frontier adopted at 3f9bf12; rk_4w shelved (torch, no sign edge); aux_ho raw = value channel (Phase 59)

EXPERIMENTS.md

@@ -3190,3 +3190,62 @@ all-24-building out-of-fold cost), SCORE = mean of the 48 totals.
 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).

agent_reports_bgkf/FRONTIER.md

@@ -1,10 +1,27 @@
-# FRONTIER BOARD (maintained by the coordinator - agents READ before each new candidate)
+# FRONTIER BOARD - CAMPAIGN 2 CLOSED (2026-08-23 morning)
 
-Updated: 2026-08-22 ~21:40. Baseline: champion logbag/base = 1760.2 (swaps 9.5).
+All 7 lanes closed. **ft_vmin7r_wnr20 ADOPTED and DEPLOYED at commit
+3f9bf12** (submit.py Ft2RUL, BATTERYSWAP_FT2 default ON; deployment parity
+vs the CV frontier bit-exact, max cell diff 0; soft degradation to legacy
+champion on any failure; BATTERYSWAP_FT2=0 = the measured-1661.45 config).
+Baseline: champion logbag/base = 1760.2 (swaps 9.5).
 
-## Standings (paired on 48 origin-sums vs 1760.2)
+## Final standings (paired on 48 origin-sums vs 1760.2)
 - **ft_vmin7r_wnr20  -71.8 +/- 23.5 (3.05 sigma, wilcoxon p=.037, swaps 9.58)
-  = NEW FRONTIER, the campaign's one fully-clean survivor.**
+  = THE FRONTIER, adopted. The campaign's one fully-clean survivor that
+  also clears deployment review.** Worst-case reference-shift bound: vmin7
+  recomputed WITHIN each ~5-building test fleet still -60.8 +/- 26.3
+  (2.31 sigma); scored fleets are full ~922-device populations (closer to
+  the shared-reference case).
+- rk_4w -111.5 +/- 27.7 (4.02 sigma, Wilcoxon .001, 29W/13L) = strongest
+  anchored number, NOT adopted: no sign-test edge over the frontier
+  (head-to-head -39.7 +/- 20.3, 15W/15L, Wilcoxon .131) and RankNet is a
+  torch dependency with cross-version determinism risk. SHELVED
+  (artifacts/rankers.py + run3w.py).
+- wd raw vm7+aux_ho+wiener.20 -126.4 +/- 43.8 = REPORT-ONLY: rk_4wctl
+  settled the adoption form - aux_ho as ORDER adds +6.1 n.s. on the
+  frontier, so the raw excess is VALUE-CHANNEL (the calibration channel
+  the public refuted in phase 56). Statistical tie with rk_4w (+14.9).
   Composition: FEATURES + vmin7_rank (fleet pct-rank of 7d-min voltage,
   already in the table) with the wiener .20 value-preserving order blend.
   Count-clean by every control (corr +0.048; count-flat scenarios carry the
@@ -53,6 +70,12 @@ Also dead: z-scores (weaker vrel), building relatives, dynamics ranks.
   hgbd1l25-on-vrel leader row is void - vrel base quarantined.)
 
 ## Closed lanes
+- LEARNING-TO-RANK (rk, complete 08-23, no adoption): see final standings
+  - rk_4w/3w552025h are real, seed-replicated, fully anchored, but no
+  sign-supported edge over the frontier + torch dependency. Within-building
+  pairs are the WORST pair scheme (+310 at full weight); within-origin
+  days-weighted pairs the right one. aux_ho order-slot test settled the
+  wd adoption form (order adds nothing; raw = value channel).
 - MODEL FAMILIES (md, complete 08-23, no survivor): keep champion
   logbag(C=1, bags=15) - locally optimal (C=0.3 -9.7 n.s.; MORE BAGS HURT
   +21/+23 at 30/60; scalers/penalties/imputers noise). Monotone HGB is

agent_reports_bgkf/md_findings.md

@@ -1,11 +1,17 @@
-# md — Model families & hyperparameters (building-GroupKFold)
+# md — Model families & hyperparameters (building-GroupKFold) — FINAL
 
-STATUS: resumed after outage (deadline-day morning). All pre-crash reads below are
-valid — harness is deterministic, deltas are vs the same baseline recipe (1760.2,
-swaps 9.50). Remaining queue: d2 l2-bowl completion (l2=2, l2=10), value-anchored
-twin of bowl winner, 3-way order source on vm7 frontier base.
+STATUS: DONE (deadline-day morning). Baseline logbag = 1760.2, swaps 9.50
+(regenerated file verified identical before any post-outage pairing; harness
+deterministic, so pre-outage reads are retained). Coordinator queue completed:
+HGB bowl finished, count check on winner done, 3-way composition on frontier done.
 
-## Results (base variant, paired vs /tmp/cv_bgkf.csv, 48 origin-sums)
+## Headline
+Regularized MONOTONE HGB is the only real model-family signal in this lane, and it
+is REAL but REDUNDANT: its ordering core survives value-anchoring (-44.7 ± 38.6)
+yet adds nothing (+1.0 ± 24.9) on top of the vm7+wiener.20 frontier stack.
+**No survivor to promote from the md lane.**
+
+## Main table (base variant, paired vs /tmp/cv_bgkf.csv, 48 origin-sums)
 
 | label | SCORE | delta | SE | W/L | swaps |
 |---|---|---|---|---|---|
@@ -19,49 +25,66 @@ twin of bowl winner, 3-way order source on vm7 frontier base.
 | md_bl_hgbd1_w3 (mean blend w=.3) | 1702.0 | -58.1 | 38.4 | 27/18 | 9.52 |
 | md_hgb_d2s (mono d2, lr.05x400) | 1704.6 | -55.5 | 72.2 | 29/19 | 9.50 |
 | md_hgb_d2l25 (mono d2, l2=5) | 1710.3 | -49.9 | 66.1 | 27/21 | 9.52 |
+| md_rk_hgbd2_w5 (ANCHORED TWIN of d2, w=.5) | 1715.5 | -44.7 | 38.6 | 24/19 | 9.60 |
 | md_hgbmono_l2 (mono, depth3, l2=1) | 1715.1 | -45.1 | 70.2 | 27/21 | 9.35 |
 | md_hgb_leaf50 (mono, leaf50, l2=1) | 1716.1 | -44.1 | 71.4 | 25/23 | 9.44 |
-| md_blcm_hgbd1 (blend, count-matched) | 1719.7 | -40.5 | 31.2 | 23/22 | 9.65 |
-| md_hgb_l2s (mono, l2=1, lr.05x400) | 1726.2 | -34.0 | 69.9 | 23/25 | 9.46 |
-| md_rk_hgbd1_w5 (rank-only twin, w=.5) | 1744.4 | -15.8 | 33.0 | 16/25 | 9.71 |
+| md_blcm_hgbd1 (blend d1, count-matched) | 1719.7 | -40.5 | 31.2 | 23/22 | 9.65 |
+| md_hgb_d2l2 (mono d2, l2=2) | 1721.2 | -39.0 | 75.3 | 24/24 | 9.46 |
+| md_hgb_l2s (mono d3, l2=1, lr.05x400) | 1726.2 | -34.0 | 69.9 | 23/25 | 9.46 |
+| md_hgb_d2l10 (mono d2, l2=10) | 1732.8 | -27.4 | 62.8 | 26/22 | 9.62 |
+| md_rk_hgbd1_w5 (anchored twin of d1, w=.5) | 1744.4 | -15.8 | 33.0 | 16/25 | 9.71 |
 | md_robust (logbag + RobustScaler) | 1747.8 | -12.4 | 13.8 | 24/8 | 9.46 |
 | md_impmean (logbag, mean imputer) | 1747.7 | -12.4 | 11.5 | 16/5 | 9.50 |
 | md_logbag_c03 (C=0.3) | 1750.5 | -9.7 | 8.5 | 23/9 | 9.48 |
 | md_enet (saga elasticnet l1r=.5) | 1756.5 | -3.7 | 3.4 | 15/5 | 9.48 |
 | logbag (baseline) | 1760.2 | 0.0 | - | - | 9.50 |
 | md_l1 (saga L1, C=1) | 1766.7 | +6.6 | 10.4 | 13/6 | 9.50 |
-| md_hgbmono_slow (mono, lr.05x400) | 1768.6 | +8.5 | 73.6 | 23/25 | 9.38 |
+| md_hgbmono_slow (mono d3, lr.05x400) | 1768.6 | +8.5 | 73.6 | 23/25 | 9.38 |
 | md_logbag_c3 (C=3) | 1768.8 | +8.6 | 10.2 | 7/9 | 9.50 |
 | md_logbag_b30 (bags=30) | 1781.2 | +21.0 | 13.7 | 11/12 | 9.50 |
 | md_logbag_b60 (bags=60) | 1783.2 | +23.0 | 13.7 | 13/16 | 9.52 |
-| md_hgbmono (mono, gbm defaults) | 1788.6 | +28.4 | 71.0 | 22/26 | 9.54 |
+| md_hgbmono (mono d3, gbm defaults) | 1788.6 | +28.4 | 71.0 | 22/26 | 9.54 |
 | md_gbm (unconstrained control) | 1809.5 | +49.3 | 85.7 | 24/24 | 9.94 |
 
-## vrel-family runs (base QUARANTINED by coordinator — historical record only)
+## Frontier composition (variant vm7 = BASE + vmin7_rank; ref reproduced exactly)
 
-| label | SCORE | delta | SE | W/L | swaps | vs vrel-logbag |
+| label | SCORE | delta | SE | W/L | swaps | vs vm7+wnr20 ref |
 |---|---|---|---|---|---|---|
-| md_hgb_d1_vrel | 1640.7 | -119.5 | 63.5 | 26/22 | 9.15 | -52.7 ± 53.3 |
-| md_bl_hgbd1l25_vrel | 1642.4 | -117.8 | 46.1 | 30/18 | 9.17 | -51.0 ± 34.2 |
-| md_bl_hgbd1_vrel | 1667.2 | -93.0 | 45.5 | 30/18 | 9.19 | -26.2 ± 34.3 |
-| md_blcm_hgbd1_vrel (count-matched) | 1667.6 | -92.6 | 46.8 | 31/17 | 9.29 | ~-25.8 |
-| md_logbag_vrel (reference) | 1693.4 | -66.8 | 41.4 | 32/15 | 9.29 | 0 |
+| md_vm7_wnr20 (frontier ref, logbag+wiener.20) | 1688.4 | -71.8 | 23.5 | 24/16 | 9.58 | 0 |
+| md_ord3_d2_vm7 (.55 logbag/.25 hgb_d2/.20 wiener) | 1689.3 | -70.8 | 30.2 | 21/21 | 9.65 | +1.0 ± 24.9 |
+| md_ord3_d2w15_vm7 (.65/.15/.20) | 1695.2 | -65.0 | 29.8 | 21/19 | 9.67 | +6.8 ± 17.4 |
+| md_ord3_d1l25_vm7 (.55/.25/.20, stumps) | 1705.3 | -54.9 | 31.7 | 22/21 | 9.62 | +17.0 ± 27.3 |
+
+## vrel-family runs (base QUARANTINED by coordinator — record only, not claimed)
+md_hgb_d1_vrel -119.5 ± 63.5 | md_bl_hgbd1l25_vrel -117.8 ± 46.1 (vs vrel-logbag:
+-51.0 ± 34.2) | md_bl_hgbd1_vrel -93.0 ± 45.5 | md_blcm_hgbd1_vrel -92.6 ± 46.8 |
+md_logbag_vrel reference -66.8 ± 41.4 (reproduced frontier read exactly).
 
 ## One-liners
-- gbm control reproduces prior read exactly (+49.3). Monotone constraints alone: +49 -> +28.
-- Constrained + regularized is a clean monotone trend down: depth3 l2=1 -45, l2=5 -68;
-  depth2 l2=1 -86.4 (best solo); stumps l2=1 -79.0 ± 54 with PERFECTLY flat swaps 9.52.
-- Depth2 bowl so far: l2=1 -86.4, l2=5 -49.9 — l2=2 and l2=10 pending to bracket.
-- Mean blend logbag+stumps halves the SE: -75.7 ± 39.0 (1.94 sigma), swaps 9.65.
-- QUARANTINE checks (base): rank-only twin keeps -15.8 ± 33.0 of -75.7 (ordering is a
-  minority share); count-matched twin keeps -40.5 ± 31.2 with swaps unchanged at 9.65
-  (win is NOT sum(p42)-count inflation — the value channel is per-device magnitudes
-  feeding ec-margin, not expected count).
-- Linear family: flat. C=0.3 mildly good (-9.7 ± 8.5, W/L 23/9), more bags actively
-  hurt (+21/+23 at 30/60 — keep bags=15), saga L1/enet/imputers/scalers all within noise.
-- vrel compositions were my biggest raw numbers (-118/-120) but the base is quarantined;
-  kept above for the record, not claimed.
+- gbm control reproduces prior (+49.3). Monotone constraints alone: +49 -> +28.
+- Constrained+regularized trend: depth3 l2=1 -45 / l2=5 -68; depth2 l2=1 -86.4 (bowl
+  bottom); stumps -79.0 flat swaps. d2 bowl is NOT smooth (l2=2 -39, l2=5 -50,
+  l2=10 -27): l2=1 is the bottom, the rest is per-config noise (SE 63-75). No solo
+  config passes 2*SE.
+- COUNT CHECKS on the winner: d2's value-anchored twin (champion p42 values, half-
+  weight d2 ordering, count machinery untouched) keeps -44.7 ± 38.6 at swaps 9.60 —
+  a genuine ordering core, not a count exploit (raw d2 swaps 9.25 drift explained
+  as value-channel, not sum(p42): count-matched blend also kept -40.5 ± 31.2).
+- FRONTIER TEST (the decisive one): as third order source on vm7+wiener.20 the HGB
+  increment vanishes — +1.0 ± 24.9 (w .25), +6.8 ± 17.4 (w .15), stumps +17.0.
+  HGB's cross-fleet ordering signal is subsumed by wiener + vmin7_rank.
+- Linear family is FLAT: C=0.3 -9.7 ± 8.5 (W/L 23/9) is the only mild positive;
+  MORE BAGS HURT (+21.0/+23.0 at 30/60 — keep bags=15); saga L1/enet, RobustScaler,
+  mean-imputer all within noise. Champion logbag(C=1, bags=15) is locally optimal.
+- Not run post-outage (deprioritized per coordinator, low prior): xtrees, mlp,
+  torch-MLP, Cox p42=1-S(42), class_weight (count-coupled by construction).
+
+## Conclusion / recommendation
+NO SURVIVOR from the md lane. Keep champion logbag(C=1, bags=15) as the p42 model;
+do not add bags; do not swap scalers/imputers/penalties. Monotone depth<=2 HGB is a
+validated, quarantine-clean, decorrelated ORDER source in isolation (-44.7 anchored)
+— worth keeping on the shelf ONLY as a fallback third source if the wiener order
+source is ever quarantined; on the current vm7+aux_ho+wiener stack it adds nothing.
 
-Conclusion (interim): regularized MONOTONE HGB (depth<=2) is the one real model-family
-signal on this instrument; solo never passes 2*SE (SE ~55-73), blends reach 1.9 sigma.
-Next: bowl completion, anchored twin of bowl winner, then 3-way order source on vm7.
+Artifacts: /tmp/cv_bgkfcand_md_*.csv (post-outage runs), driver /tmp/md2/{run,cmp}.py,
+models /tmp/md2/mdmodels.py (HGBMono monotone map, MeanBlend/CM, OrderBlend3).

agent_reports_bgkf/rk_findings.md

@@ -1,6 +1,6 @@
 # rk: learning-to-rank (Campaign 2, building-GroupKFold)
 
-STATUS: running — headline (within-building pairwise logistic) mapped across weights; variants next.
+STATUS: CLOSED (coordinator adjudication 2026-08-23 — see final section at bottom). No adoption; rk_4w shelved (torch dependency, no sign-test edge over the shipped frontier).
 
 Baseline: champion logbag/tsroute, SCORE 1760.2, swaps 9.5/scenario (48 paired all-building scenario sums).
 
@@ -38,3 +38,90 @@ Conclusion: pending — within-origin contrast, days-weighting, vrel/rank-input
 - None past 2*SE yet; variance is RankNet seed noise -> trying bigger seed bag next.
 
 Conclusion (interim): promising non-survivor; best rk_rnod50 at 1.8 SE.
+
+---
+# DEADLINE-DAY RESUME (post-outage rebuild)
+
+STATUS: running. /tmp and venv lost overnight; driver REBUILT from recipes above
+(/tmp/rk2/rankers.py, run.py, run3w.py) and validated EXACTLY against recorded
+reads on the restored environment (venv -> blind-repo symlink, sklearn 1.9.0 /
+torch 2.13.0; regenerated baseline verified 1760.1688):
+- rk_rnod50s7 re-run: SCORE 1700.09, delta -60.08 +/- 28.24 (2.13 SE), 26W/10L,
+  swaps 9.52 — identical to the decimal. Count check clean: swing max +/-1,
+  changed 5/48, mean +0.02.
+- rk_3wrepro re-run: 1688.36, -71.81 +/- 23.50 — frontier base reproduced exactly.
+
+IMPORTANT CORRECTION to the pre-crash "FINAL" section above: the vrel base was
+QUARANTINED after that section was written — all rk_vrnod* rows are reads on a
+contaminated base and are WITHDRAWN from recommendation. The champion-base
+survivor rk_rnod50s7 and its replication stand.
+
+Pre-crash reads recovered from session (CSVs lost, re-running now): 3-way on
+frontier base (vm7 values; order = A*logbag(vm7) + 0.20*wiener + C*ranknet):
+- C=0.15 (A=0.65): -89.3 +/- 25.6, 25W/15L, swaps 9.62
+- C=0.25 (A=0.55): -90.8 +/- 25.8, 27W/16L, swaps 9.56
+Queue: confirm C=0.25; ranker with hot_90/oob_90+vmin7_rank inputs; aux_ho
+4-way slot test; count checks; final table.
+
+## Resume results (confirmed on restored env; all paired vs regenerated baseline 1760.1688)
+| label | SCORE | delta | SE | W/L | swaps |
+|---|---|---|---|---|---|
+| rk_3w552025h | 1657.9 | -102.2 | 27.7 | 31W/11L | 9.62 |
+| rk_3w552025 | 1669.3 | -90.8 | 25.8 | 27W/16L | 9.56 |
+| rk_3wrepro (frontier base) | 1688.4 | -71.8 | 23.5 | 24W/16L | 9.58 |
+| rk_rnod50s7 (champion base) | 1700.1 | -60.1 | 28.2 | 26W/10L | 9.52 |
+
+- rk_3w552025h = NEW LANE BEST (3.69 sigma vs champion): vm7-logbag values, order
+  = 0.55*logbag(vm7) + 0.20*wiener + 0.25*RankNet(fvm7h), where fvm7h = champion
+  13 + vmin7_rank + hot_90 + oob_90 as RankNet inputs (wd's aux-ho pair helps my
+  ranker too). Marginal ON frontier base: -30.4 +/- 19.5 (17W/14L).
+- rk_3w552025 (no hot/oob inputs) confirmed pre-crash read exactly: -90.8;
+  marginal on frontier base -19.0 +/- 14.6.
+- COUNT AUDIT (phase-56 style) on rk_3w552025h: 41/48 scenarios count-identical
+  to champion; count-flat scenarios ALONE score -108.1 — win is pure ordering,
+  count shifts (7 scenarios, -2..+3) slightly DILUTE it. Clean.
+- Win distribution: warmup -132, panel -113, nearend +1 — NOT warmup-concentrated.
+- Running: seed-bag replication (seeds 100-106) of rk_3w552025h; aux_ho 4-way
+  slot test (control 0.50lb/0.20wnr/0.30auxho vs treatment 0.35/0.20/0.25auxho
+  + 0.20ranknet, vm7 values anchored).
+
+---
+# LANE CLOSED (coordinator, 2026-08-23 ~09:10)
+
+The agent was lost twice (detached-chain stall, then a machine-sleep API
+kill) but its full run queue completed on disk; the coordinator finished
+the adjudication from the CSVs. All pairings on the 48 per-origin
+all-building sums vs baseline 1760.2.
+
+## Final table
+| label | composition (vm7 values; order weights) | delta | SE | W/L | notes |
+|---|---|---|---|---|---|
+| rk_4w | .35 lb + .20 wnr + .20 rknet + .25 auxho | -111.5 | 27.7 | 29/13 | 4.02 sigma, Wilcoxon .001 |
+| rk_3w552025h | .55 lb + .20 wnr + .25 rknet(fvm7h) | -102.2 | 27.7 | 31/11 | count audit clean |
+| rk_3w552025hb | same, seeds 100-106 | -93.5 | 27.9 | 26/15 | seed-block replication |
+| rk_3w552025 | rknet without hot/oob inputs | -90.8 | 25.8 | 26/16 | |
+| rk_3wrepro | .80 lb + .20 wnr (= ft frontier) | -71.8 | 23.5 | 23/16 | exact reproduction |
+| rk_4wctl | .50 lb + .20 wnr + .30 auxho (NO rknet) | -65.7 | 22.7 | 24/16 | |
+| rk_rnod50s7 | champion base + rknet w=.5 days | -60.1 | 28.2 | 25/10 | |
+
+## Head-to-heads (the decisive ones)
+- rk_4w vs ft frontier: -39.7 +/- 20.3 (1.96 sigma) but W/L 15/15,
+  Wilcoxon .131 — mean-supported ONLY, no sign-test support.
+- rk_4w vs 4wctl (the RankNet increment): -45.8 +/- 20.9 (2.19 sigma),
+  Wilcoxon .101 — few scenarios carry it.
+- rk_4wctl vs ft frontier: +6.1 +/- 18.8 — **aux_ho as an ORDER source
+  adds NOTHING beyond wiener+vmin7_rank** (settles the wd adoption-form
+  question: the raw composition's excess was value-channel).
+- rk_4w vs wd raw vm7+auxho+wnr20: +14.9 +/- 38.0 — statistical tie.
+
+## Verdict
+**No adoption from this lane.** rk_4w is the campaign's strongest fully-
+anchored vs-baseline number (all compositions injected at blend weight 1.0
+= value-preserving by construction; swaps identical to the vm7 primary's
+up to tie-handling), and RankNet's contribution replicates across disjoint
+seed blocks — but its increment OVER the shipped frontier has no sign-test
+support (15W/15L/18T), and RankNet is a torch dependency (MLP 32-16, Adam,
+300 epochs x 7 seeds) the competition Docker is not known to carry, with
+cross-version determinism risk. SHELF: rankers.py + run3w.py preserved in
+agent_reports_bgkf/artifacts/; recipe = ThreeWay in run3w.py
+(A=.35 lb / B=.20 wiener / C=.20 RankNet(fvm7h) / D=.25 aux_ho).