6.6 KiB
H2-S2 final results and executor handback — 2026-11-18
H2 is CONFIRMED on field error under the frozen pooled rule; last-filled-location error is worse than A. This is five-fold recipe confirmation after observing fold0, not an untouched-five-fold preregistration or a test isolating true connectivity. The original score and decision are unchanged.
Stored five-fold result
552 parts / 204 equally weighted evaluation components; 2000 paired component bootstrap replicates, seed2026111408. Field and last/Rg lower is better; Kendall tau-b higher is better.
| Reference | Metric | H2 mean | Reference mean | H2 − reference | 95% CI |
|---|---|---|---|---|---|
| A | relative_l2 | 0.065791907 | 0.070830541 | -0.005038634 | [-0.007536893, -0.002830534] |
| A | last_location | 0.456616842 | 0.376713808 | +0.079903034 | [+0.006940890, +0.155577284] |
| A | kendall_tau_b | 0.930128834 | 0.926649720 | +0.003479115 | [+0.001039504, +0.006226751] |
| B13_E1b | relative_l2 | 0.065791907 | 0.069702214 | -0.003910307 | [-0.006306713, -0.001644771] |
| B13_E1b | last_location | 0.456616842 | 0.387589688 | +0.069027155 | [+0.002026449, +0.134446230] |
| B13_E1b | kendall_tau_b | 0.930128834 | 0.928112244 | +0.002016590 | [-0.000203388, +0.004374493] |
Only pooled H2−A field CI upper<0 with complete552/204 support is the confirmation gate. B13/last/tau are reported outcomes. Negative predictions remain unclipped: 40,852 full-node / 665 unique-query nodes. There are42 G-invalid parts; the complete roster is retained.
Folds1–4 only: descriptive sensitivity table
These431 parts / 167 components exclude the already observed fold0. This additional aggregation uses only stored per-part metric scalars. It does not create a new confirmatory gate or rerun prediction/target scoring. The component-paired percentile intervals describe evaluation-family uncertainty, not initialization or learning-pool uncertainty.
| Scope | Parts / components | Metric | H2 | A | B13-E1b | H2 − A [95% CI] |
|---|---|---|---|---|---|---|
| folds1–4 | 431 / 167 | relative_l2 | 0.065420477 | 0.070366829 | 0.069507228 | -0.004946352 [-0.007654256, -0.002482391] |
| folds1–4 | 431 / 167 | last_location | 0.428297978 | 0.355019700 | 0.362882224 | +0.073278277 [+0.003853634, +0.144415239] |
| folds1–4 | 431 / 167 | kendall_tau_b | 0.931609825 | 0.927874413 | 0.929205798 | +0.003735412 [+0.001028048, +0.006708727] |
| fold1 | 91 / 41 | relative_l2 | 0.065184002 | 0.068237407 | 0.068877591 | -0.003053405 [-0.008083001, +0.001877264] |
| fold1 | 91 / 41 | last_location | 0.422701937 | 0.369172459 | 0.346130733 | +0.053529479 [-0.057908574, +0.164094887] |
| fold1 | 91 / 41 | kendall_tau_b | 0.932850707 | 0.931022035 | 0.930128723 | +0.001828672 [-0.002471989, +0.006219912] |
| fold2 | 99 / 41 | relative_l2 | 0.066387316 | 0.073047736 | 0.070409577 | -0.006660420 [-0.010371051, -0.003071975] |
| fold2 | 99 / 41 | last_location | 0.493185845 | 0.377786625 | 0.392994129 | +0.115399220 [-0.017912771, +0.287285496] |
| fold2 | 99 / 41 | kendall_tau_b | 0.928807593 | 0.924209251 | 0.927113294 | +0.004598342 [+0.000720233, +0.008583173] |
| fold3 | 119 / 42 | relative_l2 | 0.064426254 | 0.067619224 | 0.066692168 | -0.003192970 [-0.008443514, +0.002218274] |
| fold3 | 119 / 42 | last_location | 0.414406747 | 0.328601580 | 0.365257217 | +0.085805167 [-0.108830846, +0.265556234] |
| fold3 | 119 / 42 | kendall_tau_b | 0.934189224 | 0.931064356 | 0.932806995 | +0.003124868 [-0.002796985, +0.008750355] |
| fold4 | 122 / 43 | relative_l2 | 0.065695185 | 0.072524702 | 0.071996795 | -0.006829517 [-0.014237604, -0.000585898] |
| fold4 | 122 / 43 | last_location | 0.385332089 | 0.345620957 | 0.347823464 | +0.039711132 [-0.062223681, +0.143465643] |
| fold4 | 122 / 43 | kendall_tau_b | 0.930579141 | 0.925252123 | 0.926803528 | +0.005327017 [-0.001767658, +0.014622626] |
Full comparisons against both references are in artifacts/phase2-h2-s2-terminal-report-2026-11-18/folds1to4-descriptive.json.
Execution, audit and accounting
TRON started the already reviewed postfit once under Claude1794946447.294149 because MeltAstra was unreachable. Coordinator985167; implementation25e79995, contract0bd03f42, actual pre-score PASS9c163702. This exception did not allow source edits, retry, publication by TRON, or launching an unbuilt entry. MeltAstra resumes ownership here. No second S2 launch occurred.
Four new prediction guards plus one score guard PASS. Fold0 has121 exact original predictions/metrics reused; folds1–4 add431 predictions and431 existing target accesses. No new fit or decode. Models freezef033b454 → predictions freeze86480b4d → score-stage PASS validation → target access → terminal4b2a710d. Original result SHA256 c43a6ab2460affdcafd3d51522e7b5e2d60a684b2b1a79a90e06042b0e19bc45; complete 2a55287167daed3ff541210f0fb3f32b8971b70976329cf29ca81828cbf06685.
MeltAssist independent existing-evidence audit PASS, SHA256 d4999e4a99e82d91b793cffd70b673a977de022878e28d8a9955ca1a4acf59da, is included unchanged. It checked63 source pins/five binding fields,8 freezes/176 Git blobs,4 checkpoints,552 prediction hashes and121 exact reused rows. MeltAstra rechecked all662 audit-recorded current hashes and552 predictions. This is evidence verification, not independent scientific recomputation.
Audit limits: ordering is supported jointly by Git ancestry, fixed entry and receipts; target log is overwritten, not syscall-level tracing. CPU formula and worker lower bound agree with492.447055337s, but starting resource counters were not retained, preventing independent exact coordinator/reaped-child decomposition. Overwritten progress witnesses cannot all be rehashed historically; retained peak maxima/final results were checked. Current process absence cannot independently establish historical absence.
GPU ledger: oldS1 18858.214908715 + newfourfoldtraining79602.784961751 + prediction355.327867327 = 98816.327737793GPU-s /144000. Final postfit CPU492.447055337 /3600s. Assist audit0.841653768CPU-s is separate; this report/aggregation cost is separately recorded in report-cost.json and does not rewrite the original ledger.
Stage decision
Claude1794946961 triggers the frozen304f5771 phase3 plan. Next is item1 data-sufficiency e24fff7c,24GPUh within the new170GPUh plan, first-epoch projection and a new exact MeltAssist review. This report does not launch it. Later seed1/last-fill/context/ensemble remain future stages in the prescribed order. Gate-v2 outcome equivalence failed, so fresh-family admission remains closed. Scientific goal remains useful cross-part filling prediction; neither topology/local-type organisation nor connectivity is globally settled by this comparison.