Files
MeltMind/docs/track-a-h2-entry-2026-11-18.md
T

27 lines
5.5 KiB
Markdown

# Track A H2: five original v4 random splits
## Actual launch after independent PASS (2026-11-18 12:28 UTC+8)
Assist1794975952 actual receipt1fa923f8 was adopted byte-for-byte, frozen in review commitb7879eec and validated against implementation6701e4d6/contractbe4ac1f6. The sole service meltmind-track-a-h2-20261118.service is active/running with Restart=no, coordinator1262364. Job0 guard1262373/worker1262432 uses GPU2; job1 guard1262374/worker1262422 uses GPU3. Three idle polls each passed at5MiB/0%; workers actually reached epoch1 at51/311 and58/304 parts in the first snapshot. Initial normalization311/304 completed, source pins unchanged. Item1 worker1135393 remains onGPU1, job5 step56503/64000; GPU0 is reserved for its postfit. Pending jobs2/3/4; first-epoch cost gates pending at this snapshot. No postfit/test access. Actual root records include review-adoption.json, host-precheck.json, started.json and actual-start-audit.json; this supersedes the historical prepared-only status below.
Status: implemented, 20 CPU synthetic checks PASS, local implementation frozen; independent entry review pending. No new training, prediction or scoring launched. Input reuse audit 1eae27ad is not training acceptance.
Research question: compare the unchanged H2 recipe with the existing Greif-style A baseline under the same five original random 314/315 splits. This is protocol-level re-creation, not the unpublished author split identity. Family leakage remains possible in random-part splits. The metric remains per-part RMSE/max(full truth), min/median/max over each test set and mean over five splits; old A median0.0396 is context, not the new paired result. No topology/local-type hypothesis is closed by this comparison.
Claude3608/3932 require A reuse without refit, 615 frozen v4-native H1 graphs/features/raw IDs and preservation of14 old failures. Claude1794975093.432709 approves full actual train311/304/307/306/310, seed0,200epochs, original optimizer/schedule, last-epoch weights, no validation and original H1 v4 learning-only clocks. No63-part queue or cross-version ID mapping. Candidate test304/311/308/309/305 remain provisional until H2 predictions exist.
The independent entry is tools/track_a_h2_{prepare,common,worker,guard,queue}.py. Shared H2 scientific modules are unchanged and source-pinned. Every epoch uses all actual train parts and their full nodes/true graph edges. Feature normalization is computed per split on train only with the original FP64 population statistics; the network and optimizer retain the original FP32 recipe, separate FP64 Vf skip, no AMP/TF32. The five clocks are reused by hash with exact train membership and equal-part grouping; no clock targets reopened. Model seed0 is distinct from split seeds2026111601..1605.
Run root: /home/frank14f/MeltMind-runs/track-a-h2-2026-11-18. Final contract SHA256 be4ac1f6d5a5d6c0183aa50d2e046b22d7e5fa77c7714fef7eab6d616e68bee7. Local implementation commit6701e4d67f2f7ee6c70d22a3512f40c4ad9180f5; proof SHA f740a9c4654c67911d344e4b98c19adf3776179ba31dcaf859b5c44669420700. expected-review-binding.json specifies actual exact review identity. Initial unreviewed f0def4cd contract is retained in before-final-checks, superseded only before any launch.
GPU2/3 only, at most two workers; GPU1 remains item1 training and GPU0 reserved for its frozen postfit. CPU feeders8/10, guards28/29, coordinator30/31. Each dispatch takes a GPU lock and requires three idle polls. No other queue is stopped or modified. Total162000 GPU-s includes setup/waits/prediction:154800training plus7200postfit reserve; five job caps30960 each, setup60 charged once per job. Per-job CPU61920s/15GiB VRAM/16GiB RSS/8GiB output, aggregate40GiB RSS/48GiB output. These are engineering stop ceilings, not projected use.
Each job pauses after its own first epoch. Pending jobs reserve their entire caps, active elapsed time cannot disappear, terminal costs replace forecasts once. Cost uses maximum observed full-epoch duration with1.25future margin and120s final overhead. It is updated every epoch. No five-job barrier on two slots. Over-budget projection, worker/monitor failure, recorded or live VRAM breach, or global stop ends without retry. All setup and active waiting remain charged. Whole-job guards inspect underlying monitor_errors even if the supervisor reports PASS, and scan peaks in failed-worker records.
All five terminal PASS checkpoints and normalizations are Git-frozen before this training-only queue completes. It contains no postfit launch. A separate actual MeltAssist review is required before any test access; old review PASS cannot substitute. Prediction/scoring implementation is outside this training entry and must preserve the7200s reserve and total45h.
Validation:20 CPU-only synthetic checks passed (0.381684517CPU-s test body;0.383417746wall-s; imports excluded). Coverage includes exact roles, test-to-train rejection, no new holdout, test loader refusal before IO, clock and native-graph identity, two-slot dispatch, reserved item1 GPUs, pending/active/terminal accounting, slow first-epoch stop, global-stop gate, raw monitor-error rejection and failed peak enforcement. A mocked worker traverses all200epochs/400synthetic updates with no validation/test reads; it is not a real fit or GPU benchmark. First18-check pass is retained as prior test history; expanded20-check pass is final. Existing A/615graph audit was not rerun.
Next decision: independent review of these actual files, then authorized idle-GPU execution if PASS; first-epoch evidence decides cost feasibility. The original scientific task and item1 remain unchanged.