diff --git a/docs/claude/research-tree.md b/docs/claude/research-tree.md index 317ce3402..d5887f71e 100644 --- a/docs/claude/research-tree.md +++ b/docs/claude/research-tree.md @@ -34,6 +34,7 @@ Rules: | P3 | O1 with size-matched K = 8 partitions | — | — | PRUNED (eligibility 23 %; lesson: control cut area as a covariate, do not match it) | Log (newest first): +- 2026-11-20: **k4 on the 170 new-family holdout (second use, fixed rule): INCONCLUSIVE** (binding 02fe8d0d, permit ZENODO170-K4-SECOND-USE-1120; result 4f6d388e, ledger ad5f730b; 170 reads). Primary tail-support hit rates are high for all arms (ceiling): *full-node H2* argmax 93.98 % → k4 95.09 %, +1.11 pp [−.09, +2.61]; mean distance .0484 → .0447 Rg. *Common PPS:* A 94.47 %, H2-argmax 93.60 %, H2-k4 94.71 %; k4 − A +.23 [−2.78, +3.08]; k4 − argmax +1.11 [−1.96, +3.83]. Secondary single-peak hits: A 17.97 %, H2-argmax 12.14 %, H2-k4 13.97 % (descriptive). **Reading:** the k4 gain confirmed on Greif (folds 1-4 clean subset) does not reach significance on the new-family set, where the tail-region metric is near ceiling; same direction, smaller effect. GINO-Greif resume started 17:03 UTC (GPU0). - 2026-11-20: **k4 confirmation on Greif folds 1-4** (bindings c1080ca1 + amendment 652d6d8b, MeltAssist PASS; second-use ledger, 431 reads; result 74eee994). Primary metric = distance to true last-1 % volume support, hit ≤ .10 Rg; family-paired 95 % CI. Primary report on the 320 parts / 132 families never used to select k: *H2 full-node* argmax 85.05 % → k4 88.51 %, +3.46 pp [+.64, +6.20] → **k4 gain CONFIRMED**; mean distance .107 → .076 Rg. *Common PPS nodes, three arms:* A 88.56 % (.074 Rg), H2-argmax 88.04 % (.085), H2-k4 89.09 % (.062); k4 − A +.53 [−3.02, +4.55] → inconclusive; k4 − H2-argmax +1.05 [−1.69, +3.75] → inconclusive. Secondary (distance to the single true global peak): A 21.1 %, H2-argmax 15.1 %, H2-k4 18.8 % hits; A remains better at pinpointing the exact peak. **Reading:** k4 is a confirmed improvement for H2's endpoint on full nodes; against A on the shared query nodes H2-k4 is level (no evidence of a gap either way on the tail-support metric), whereas the earlier 'last fill worse' finding was on the global-peak metric, where A still leads. Zenodo 170 second-use run pending relay. - 2026-11-20: **Last-fill readout: k4 mesh smoothing before argmax improves H2's endpoint (fold 0, pre-registered).** Spec MoldSol 226c2b1e (frozen with my amendment), binding b1955c29 (MeltAssist PASS after two fixes), CPU only, no retraining. Validation (fold 0, 111 parts / 35 families, pooled seeds, hit = official d ≤ .10 Rg): argmax 85.89 %; k1 88.91 % (+3.02, LB +.27); k2 89.58 % (+3.68, LB +.74); **k4 90.17 % (+4.28, LB +.94) → frozen winner**; centroid 58.18 % (−27.71). Test (fold 0, 121 / 37, one read each): seed 0 hits 82.04 → 84.42 % (+2.38), mean distance .1505 → .1270 Rg; seed 1 83.03 → 92.98 % (+9.95), .0582 → .0374 Rg; combined paired gain +6.16 pp, 95 % CI [+.39, +13.60]. **Reading:** supports the spurious-maximum hypothesis from the 34-case classification: reading the endpoint from a mesh-smoothed field, not the raw argmax, recovers part of H2's last-fill loss. Field predictions unchanged. Limits: one fold, wide CI; not yet compared against A on the same metric; not yet applied to other folds or the 170 holdout. - 2026-11-20: **NEW-FAMILY HOLDOUT RESULT: H2's field gain TRANSFERS** (Zenodo 17831586, 170 cases / 98 ABC families, zero roster/near-duplicate overlap; frozen 5-fold ensembles of A, H2 seed 0, H2 seed 1, no retraining; pre-score PASS a670a4d4/f314ca15 + split PASS 06c35e1a; predictions frozen 202a01a1 before any target read; each Fill_time decoded once; result a236d696, target ledger 0f3cd58d). Family-weighted field relative L2, H2 − A, 98-family paired bootstrap (2,000, seed 2026111408): seed 0 −1.322 pp, CI [−1.721, −0.959]; seed 1 −1.364 pp, CI [−1.769, −0.971] → both FIELD_GAIN_TRANSFERS (pre-registered). Last fill: hit counts equal across the three; distance up for seed 0, slightly down for seed 1 (descriptive, no gate). Reading: on genuinely new part families from a different public release, the learned mesh model's field advantage over the tree model holds and is larger in absolute terms than on the grouped Greif protocol (−0.50 pp); last fill is still not improved. Limits: same solver/material/process as training; single gate per part; one release. Compute: 1a 37,748 CPU-s, 1b ~196 GPU-s, scoring 83 CPU-s.