4.1 KiB
H2-S1 final200: field error improves on the fixed fold; last-location improvement is not established
H2-S1 postfit is COMPLETE on121test parts/37components/3125278fullnodes. The original strict OR eligibility gate passes through its field branch. This is a single fixed fold0/seed0 last-epoch200 result, not an across-fold confirmation or an isolation of true-connectivity effects. No S2 or data-sufficiency job has started here.
| Metric | H2 | Frozen A | H2−A | 95% paired component CI |
|---|---|---|---|---|
| Query relative L2 | 0.0674683605950 | 0.0729235099381 | −0.00545514934311 | [−0.0103481542816, −0.000975495359645] |
| Last-location distance/Rg | 0.584434418427 | 0.474630454095 | +0.109803964332 | [−0.134109874211, +0.392611260187] |
| Kendall tau-b | 0.923444363225 | 0.921122049510 | +0.00232231371431 | [−0.00234933387428, +0.00765524734102] |
Original component pooling,2000paired bootstrap and seed2026111408 were unchanged. Every metric retained37components in each arm, with zero undefined bootstrap replicates. Field error fell by0.545515percentage points. Its CI upper bound is below both the original+.005 threshold and zero. Last-location's point estimate is worse and its interval crosses zero; it does not support a last-fill improvement. Tau's interval also crosses zero. Do not substitute field success for the last-location claim.
Inference used the same final200 checkpoint35cc9e41, original learning normalization and clock, full true graph, FP32 H2 and FP64 skip. Full predictions were saved and hash-frozen before query scoring. A predictions/metrics were reused from the same frozen baseline/model/queries; no A fit or new A prediction occurred. Negative predictions were preserved:7786full-node and145query predictions, with no clipping.121cached test targets were read once for the one score; new decodes0, new fits0, retries0. Validation was not used for selection.
Assist's actual semantic PASS SHA075140841e0cffdbef7dfeaaee47f33e0da7206eee9b4a27b0c5cb9b1e365694 was fixed in31a04cb7c341af0a0c256db644d0ed0d0baaa55f and the actual validator passed before launch or target access. It binds implementation6c208f2838bd14d5eb9c9d45123002bb410a3697, contract45c8bcb1ddd2c8d73aa1dca65c0606a9fb4a515fd59c86786cf00f2622d791ec, comparisonf3cbe6dd and all49source hashes. Twenty-one CPU synthetic checks passed; the earlier preparation self-match error and its20-test log remain retained.
Predictions were fixed inb4f488cc6ec25b9935545aa7663ee5dc8ebfd05c. Score/result SHA82b3a77ea40ff63c694ec7ac170662816116c970bad2a13eee8faba84219662e and terminal71e8d5a1f809e1c00548843f147386c13df0fea0 are immutable. A separate terminal audit verified all121prediction hashes, exact target roster/accesses, PASS-before-run and predictions-freeze-before-score ordering, source pins and no live H2 processes. The one systemd service hasMainPID0, SubStateexited, ExecMainStatus0; RemainAfterExit keeps its completed unit loaded and does not imply active computation.
Prediction guard PASS:91.572765394GPU-s including60setup/idle/monitor allowance, within600. Final postfitCPU including terminal freeze159.316316656s within3600, including120preparation reserve. Pilot+train18766.642143321GPU-s were counted once, making final H2 total18858.214908715/144000s. Recorded allocated/reserved peaks229690368/262144000B, below15GiB; both guards reported no errors. The prediction guard's separate cumulative field samples elapsed time again and is0.000016260077s above prior+its charged field; the final ledger consistently uses prior+the recorded91.572765394charge. Both raw values are preserved; no cap is affected.
Artifacts: /home/frank14f/MeltMind-runs/phase2-h2-s1-postfit-2026-11-17. No rescore/retrain/retry is needed. The original gate-outcome queue was not modified. Per Claude1794921374 the next GPU direction is data-sufficiency after S1 scoring; its separate contract and execution remain outside this completed entry. Claude leads stage interpretation before any next-stage launch. Research goal remains a useful cross-part molding surrogate; these results do not settle topology/type partitioning or connectivity as separate hypotheses.