Files
MeltMind/docs/phase2-h2-s2-postfit-2026-11-17.md
T

5.2 KiB
Raw Blame History

H2 S2 pooled postfit: exact review preparation

Claude1794925470.101149 authorizes preparing this entry concurrently with the original four-fold training. This document describes implementation, not executed predictions or a received independent PASS. The preparation ends by delivering the exact binding to MeltAssist. It launches no waiter, prediction, scoring, fitting or new stage.

The unique prepared root is /home/frank14f/MeltMind-runs/phase2-h2-s2-postfit-2026-11-17. Original training remains /home/frank14f/MeltMind-runs/phase2-h2-s2-2026-11-17, implementation1f57bad66eb062d4f2269347532321929f873717, contract76e1976bed88fc390812238c3d80693446daea3840dc009ac7d8870ddc43a31d. Neither its code nor the gate outcome queue is changed.

Exact inputs and execution barriers

Reuse all121 original S1 fold0 prediction entries and per-part metrics by hash. No fold0 prediction, target-array read or per-part metric recomputation. All five folds retain the original H1/A learning/test roles, baseline model hashes, learning-only clocks, eligible query identities and multiplicities. B13-E1b predictions and per-part metrics are frozen references; its42 invalid-feature parts stay in the552-part roster.

Before any new prediction, require all four original training terminal entries PASS/exit0, exact settled ledger, original models and terminal Git freezes, absence of training processes, final epoch200 checkpoints and each fold's learning-only normalization. The future checkpoint hashes are adopted only from those verified original freezes, never guessed during preparation. The new adoption freeze binds all four checkpoints. Reject partial training, failed guards, altered checkpoints, wrong folds/steps, validation-selected checkpoints and doubled old costs.

The new expected-review-binding.json identifies the implementation commit, contract SHA256, implementation-proof SHA256, comparison-contract SHA256 and complete code SHA256 map. An actual MeltAssist PASS matching all five fields must be copied and frozen as implementation-review; the shared validator must pass. S1's PASS cannot satisfy this gate. Prediction and scoring each revalidate it; scoring does so before parsing old score values or reading test target arrays. All552 predictions are hash-frozen before scoring. Only431 folds1–4 targets are then read from the original caches; no new decoding.

Predictions and statistics

Four sequential full-graph predictions use the original fold GPUs, FP32 H2 model, original FP64 F skip, frozen final checkpoints and normalization. Query extraction follows the full-node prediction. Negative values remain unchanged; full-node and eligible unique-query negative counts are reported separately. No clipping or baseline refitting.

Pool552 parts /204 components with unchanged E1b pooled(): component means weighted equally, per-arm components require all constituent part metrics defined, paired comparison uses the intersection. Use2000 paired component bootstrap replicates, seed2026111408. CONFIRMED requires complete paired field support204/204 and a finite upper95%CI for H2−A strictly below0. Incomplete field support is PARK. H2−B13-E1b, last-fill location/Rg and Kendall tau are report-only; per-fold results also reported. Undefined metrics are retained, not zero-filled or silently dropped within components. This is recipe confirmation after observing fold0, not an untouched-five-fold preregistration or causal identification of connectivity.

Bounded accounting and failure preservation

Total H2 cap remains144000 GPU-seconds. Original S1 settled18858.21490871499 is counted once in the original fourfold training terminal ledger; that full settled cumulative amount is adopted once. New prediction allowance2400 GPU-seconds includes four600-second folds and all their idle/setup/failure time. Each fold has a conservative60-second setup charge, three idle polls, original GPU identity, and no retry. Predictions run sequentially;15GiB recorded/device VRAM,16GiB tree RSS and20GiB total postfit output caps apply. Failed workers retain peak records and cause a stop.

Postfit cap3600 CPU-seconds includes360 conservative setup (120 preparation plus four60 local allowances), worker/child CPU and coordinator cost. Before each fold reserve540 worker CPU-seconds for each remaining prediction,600 for scoring and120 for terminal work. Actual child CPU and explicitly recorded worker CPU are combined by maximum, not added twice. Score worker has600 CPU-seconds /780 wall-seconds /16GiB RSS. No resource or numerical failure triggers a retry. Failures, original S1 evidence and training pins remain intact. No data-sufficiency, extra training or other follow-on stage is launched.

Validation status

23 CPU synthetic checks passed: exact actual Git/file review validation, all-four terminal and ledger identities, checkpoint fold/final epoch rules, remaining resource reservations, failure peak/coverage checks, strict confirmation boundaries, component weighting and undefined-component semantics, exact fold0 metric/manifest reuse,552-entry uniqueness and target-read ordering. No real target values, new predictions or GPU workload were used. Independent semantic review remains pending until its actual matching receipt is received.