Phase 61 findings: team stack on bgkf (no adoption, all count-channel); operating-point analysis (podium swaps 15-17.6 vs our 12.1); alpha-1.0 A/B commit prepared

agent_reports_bgkf/c3_findings.md

@@ -126,3 +126,44 @@ its multi-level supervision + CP-SAT planner is a structurally better
 pipeline. Recommended relays: our WienerFHT order blend, vmin7_rank, a
 building-held-out transfer check, and their own gen-10 resistance channel
 (state-dependent dV/dT) as the strongest open lead there.
+
+## Team solution ON our bgkf (user directive; phase 61)
+
+Full port: their panel rebuilt from our dataset (EOL reproduction 81/81
+exact), their model retrained (building-OOF, h=42 AUC 0.961), their CP-SAT
+planner. All arms on our exact protocol (5 snake folds, test fleets
+standalone, folds summed):
+
+| arm | SCORE | vs ft2 | swaps/scen | verdict |
+|---|---|---|---|---|
+| their FULL stack (rs 1.0) | 1916.6 | +228.3, p.017 | 13.2 | planner over-buys on ~90-device fleets |
+| their model -> our planner (raw) | — | +509.7, p<.001 | 7.6 | at-risk gate starves our count rule |
+| their ORDER, anchored on ft2 values | — | +412.1, p<.001 | 9.4 | their ordering loses decisively cross-fleet here |
+| 30% their order | — | +149.1, p.015 | 9.6 | monotone dose-response, all bad |
+| 20% in 3-way with wiener | — | +59.9 n.s. | 9.6 | |
+| **our ft2 model -> their CP-SAT** | **1592.0** | **-96.3 ± 69.2, wilcoxon .72** | **16.4** | mean-only, count-carried (25W/23L) |
+
+**No structural adoption**: nothing is sign-supported, and every gain is
+count-channel. INSTRUMENT CAVEAT both ways: their machinery targets
+~460-920-device fleets; our folds are ~90 devices. Public (them 1493.65 vs
+us 1661.45) remains the only cross-pipeline arbiter.
+
+## THE OPERATING-POINT FINDING (leaderboard decomposition)
+
+| | swaps/scen | late-days/scen | early-days/scen |
+|---|---|---|---|
+| J2W (1st, 1135) | 15.1 | 60 | 622 |
+| podium (2nd-3rd) | 15.5-16.3 | 58-70 | 787-830 |
+| team repo (1493) | 17.6 | 68 | 1086 |
+| OURS (1661) | 12.1 | 121 | 275 |
+
+Every team above us swaps 25-45% more and halves our late cost. Our
+CP-SAT mix lands at 16.4 swaps locally (-96 mean) — the podium's operating
+point, discovered independently by their solver on our predictions. The
+alpha-1.5 public failure (6ea03da: +8 swaps/scen on the OLD plain-logbag
+ordering, zero catches) predates the ft2 ordering. Whether ft2's ordering
+converts extra appetite publicly is THE open question, and only a public
+trigger can answer it. Prepared as a one-click experiment:
+**commit 296f8f2 = alpha 1.0 on the ft2 frontier** (single variable);
+head 7d7632c stays at the safe 0.8. Pushes are scoring-inert - 296f8f2
+only runs if the user triggers it.