Files
MeltMind/docs/phase3-item1-postfit-2026-11-18.md
T

5.8 KiB
Raw Blame History

Phase3 item1: one-review postfit binding

Claude1794950613.195499 supersedes the earlier two-review wording: MeltAssist reviews this code and the unchanged comparison contract once, while training continues. PASS binds the implementation, contract, implementation-proof, comparison and complete source map. Future checkpoint hashes are not invented and the training PASS cannot substitute. This preparation launches no predictor, scorer or waiter and reads no target arrays.

Entry: tools/phase3_item1_postfit.py /home/frank14f/MeltMind-runs/phase3-item1-postfit-2026-11-18, using the original reproduction venv and project working directory. The root's expected-review-binding.json supplies exact identities after preparation. Missing review or incomplete training refuses before a started marker. There is one attempt, no retry or automatic next phase.

Runtime custody

All six original v2 fits must be PASS with zero exit codes, unique job/model identities, frozen models and terminal receipts, no global stop, and no queue/guard/worker processes. Their guard hashes, worker-result hashes, final checkpoint/normalization hashes, roles and steps are reconciled. Original training GPU costs are summed once and must match the settled ledger. The actual six checkpoints and the historical S1 endpoint are then adopted and frozen. Tensor metadata checks enforce fold0, seed0, pool, complete-schedule endpoint, steps and original training contract before prediction.

Seven endpoint jobs run serially on original GPU0 after three idle checks. Each uses only its own learning set plus the common111 validation and121 test geometries, original full-graph FP32 model and FP64 clock decoding. No AMP/TF32, target branch, normalization fit, clipping or model fit is introduced. The historical S1 arm reuses all121 original test prediction hashes and stored metrics; only its learning/validation predictions are new. Three original learning-only clock fields are also emitted.

Expected manifest:3265 model entries (3144 new,121 reused) and1317 clock entries =4582 total, covering552 unique parts. Complete arm/pool/domain membership, component identity, node counts, file hashes and exact S1 provenance are checked. The prediction manifest and reviewed code are frozen before any target access. The scorer then revalidates the same exact PASS, actual model freeze and prediction freeze; no second review stage is required.

Only then is each cached target array read once:320 learning,111 validation,121 test. Target-access records precede each read; they are application evidence, not syscall-level tracing. Existing S1 test metrics are reused unchanged. All new metrics retain the original fixed PPS query, multiplicity once, minimum raw-ID tie convention, query argmax last-location/Rg and original tau pairing. Full-node predictions and negative full/query counts are retained without clipping. There is no fresh VTU/label decoding.

Comparisons and interpretation

Comparison contract is byte-identical to original SHA332f2515911c99aadf2486086ddff82bd195d889faa8391296443e7aeacdfee9. Nested learning pools have34/65/132 components. Slopes use their actual log2 ratios. Common validation/test report s_U,s_2U,g50,g100,s_2U−s_U, epoch slopes, clock slopes and endpoint-minus-clock contrasts. All contrasts share2000 evaluation-component bootstrap draws, seed2026111408. Each component requires every constituent metric defined; missing support and excluded parts/components remain explicit. Own-learning summaries do not subtract metrics across different training domains.

Field classification uses s_2U and g100: strict CI low>0.001 means benefit, an entire CI within[−0.001,+0.001] means local equivalence; wide/crossing intervals or incomplete support remain unidentified. The four complete-support combinations retain data benefit, update-and-schedule benefit, both, or local plateau. The same observed fold is descriptive, not independent generalization confirmation. Component bootstrap does not measure training seed or pool uncertainty; update effects include complete LR schedule changes. No3U/extrapolation, validation selection, topology closure or next-stage inference is added.

Resource limits

The original item cap is120000 GPU-seconds; training occupies at most112800 and postfit at most7200. Historical S1 GPU cost is not charged again. Each of seven serial endpoint jobs is limited to900 charged GPU-seconds including60 setup and idle polls, and300 CPU-seconds (240 worker plus60 guard reserve). Setup/inter-job wall time is also included conservatively. No automatic retry or alternate GPU selection follows failure. Each enforces15GiB allocated/reserved/device VRAM,16GiB RSS and20GiB output; failed-worker peak records remain checked.

Postfit CPU cap3600 includes180 setup reserve, coordinator CPU and reaped child CPU,900 scoring reserve and120 terminal reserve. Remaining-job reserves are checked before dispatch. Scoring has900 CPU-seconds/1200 wall-seconds/16GiB RSS. Original training costs are kept separate and counted once; final costs include terminal-freeze work. Resource failure preserves receipts and stops.

Acceptance and scope

Initial18 tests passed; retained attempts2/3 have27 PASS each (attempt3 includes the runtime original-training source-pin check), adding six isolated actual Git/file review-validator checks and three custody/terminal tests. Checks cover strict/inclusive CI boundaries, missing whole components, common bootstrap cancellation, actual count denominator, role rejection before I/O, label-free prediction structure, checkpoint selection, pending training, budget reservations, retained failed VRAM peaks, exact review/source mutations, and rejection before target access. These are CPU synthetic checks, not real prediction/timing validation. Exact resource feasibility remains enforced at runtime. Live training source pins are checked read-only; no training source is modified.