Files
MeltMind/docs/phase2-h2-s1-postfit-2026-11-17.md
T

5.2 KiB
Raw Blame History

H2-S1 final-epoch prediction and score entry

Claude1794921374.599419 authorizes the original prediction/score entry and requires MeltAssist's exact semantic PASS before target access. This implementation prepares one attempt; preparation never launches it. Original training completed200epochs/64000steps and its immutable checkpoint is SHA35cc9e41a4197e186570089ac305c4dee100f3d37611e0ca17e01db441b93756, independently fixed in8bf1019e67de4b0feae80b932d6f29c5b39c6a75. Training success does not establish scientific utility.

Only fold0/seed0 final200 is evaluated:320learning/111validation/121test,37testcomponents,3125278fulltestnodes. Original learning normalization, clock c=1.026602777749673, full true graph, H2 FP32 model and aggregation, and FP64 F skip/seconds decoding are unchanged. The separate test-only input loader is AST-identical to original preprocessing after replacing the role guard and removing the learning-target branch. No target file is needed for prediction. It never assigns test cases a validation role.

All full-node H2 seconds are saved before extracting the original eligible query. A query predictions and A metrics come from the frozen E1b fold0 outputs, with original H1/A model identity, learning/test roles, component, cache, process setting and query IDs/counts/coordinates/wall mask checked. A is not fit or predicted again. Negative predictions are retained. Original query-weighted relative L2, last-location/Rg with raw-ID minimum tie rule, and Kendall tau-b are used. Original E1b component pooling and2000paired bootstrap/seed2026111408 remain: a component mean for an arm requires every constituent metric defined; paired differences use the intersection. No replacement of undefined values, part filtering, validation selection or newly chosen score.

The original S2 eligibility gate is strict OR: field H2−A CI upper<+.005, OR last-location/Rg H2−A CI upper<0. An undefined/nonfinite bound cannot pass its branch; equality does not pass. Eligibility is reported only. This entry launches neither S2 nor the subsequent data-sufficiency study. One fold/seed cannot isolate connectivity causality or establish across-split robustness.

The generic semantic validator binds the actual PASS to implementation commit, full code hash map, contract, implementation proof and comparison contract. Missing/FAIL/unrelated/stale/mutated or unfrozen reviews fail closed. Prediction also requires PASS. All121 predictions are hash-manifested and Git frozen before scoring; score revalidates PASS before parsing old A scores or opening cached test seconds. Target accesses are logged. Opaque source hashes may be checked in preparation; no test target arrays or old score values are decoded then.

Execution retains the original600GPU-second prediction reserve,3600CPU-second postfit cap and144000GPU-second total H2 cap. Prior pilot+train18766.64214332099GPU-s are charged once. Prediction includes60seconds setup plus idle polls and all inference/monitor time, on original GPU UUID GPU-e8853ae7-a302-6a17-bcfa-aa767e9ed20c. Three idle checks precede a single attempt; busy/failure stops with no retry. WorkerCPU0–1, coordinatorCPU2; original gate666520 and its CPU16–27 queue are untouched. Allocator expandable_segments:True; VRAM15GiB, treeRSS16GiB, output20GiB, one prediction process. The guard checks recorded allocated/reserved peaks during and after every exit, including failed workers; missing peak witnesses fail. Limits are sampled; any measured overshoot fails rather than being erased.

Postfit CPU ledger includes120seconds preparation reserve, coordinator CPU and the maximum of recorded worker charge versus actual child CPU (including Git children), counted once. Prediction reserves1200CPUseconds and score600CPUseconds, with120seconds tail space required at dispatch. The prediction guard's60CPUseconds local setup reserve is covered by the overall120 preparation reserve rather than added twice. Final resource receipts are retained; an unexpected prediction-guard exception conservatively charges elapsed GPU time. No auto-retry. No old source, original training artifact, or running gate queue is modified.

Synthetic acceptance:21CPU checks passed in1.056s (15new scope/identity/science/failure checks and6existing actual-file/Git semantic-validator checks). Includes a synthetic full graph with no target files, role rejection before I/O, FP64 enforcement, graph/query separation, checkpoint/role identity, original OR boundary, component grouping and failed-worker peak retention. The shared validator itself and freeze are not mocked; test-only filesystem constants point at temporary Git repositories. No real test prediction, target decode, or GPU workload occurs in acceptance. The exact log and hashes accompany the immutable contract. Actual semantic review and real prediction/scoring remain pending until separately evidenced.

Preparation attempt1 stopped before root creation because suffix matching mistook its own prepare script for an active prediction. Exact script basename matching fixes the engineering check; an added regression test rejects prepare/test names while retaining all true execution names. Original20-test PASS and final21-test PASS logs and the preparation failure are preserved. No scientific attempt occurred during either preparation.