Files
MeltMind/docs/phase3-seed1-v2-entry-2026-11-18.md
T

5.0 KiB
Raw Blame History

Seed1 v2: total projection and fixed cumulative budget

The preceding original and labelled-continuation seed1 runs are terminal; neither is resumed or rewritten. Claude1794990750.013949 authorizes an independent v2 binding, clarified by1794990899.609759 and1794990914.918719. This entry changes the stopping rule and adds epoch-boundary host observations, preserving the same five original grouped folds, seed1, fresh initialization,200 epochs, learning-only normalization, optimizer and learning-rate schedule, last-epoch-only checkpoint, diagnostic validation, dynamic idle dispatch and GPU2 Track A postfit priority. Folds0/1 restart fresh,2/3/4 are first attempts. No partial checkpoint adoption, seed2, ensemble or further seed1 continuation is allowed.

Ledger and stopping logic

Old original539.0720570000121 + old continuation7943.241350897006 =8482.313407897018 GPU-s counted exactly once inside the unchanged cumulative seed1 cap167200. New v2 cap158717.68659210298 consists of training151517.68659210298 +postfit7200. phase3_prior80903.62940728385 already contains original item1 and both old seed1 attempts; add only new v2 cost to it. Phase3 ceiling612000 remains. Runtime counters are new-v2 only. The original stopped-coordinator CPU accounting gaps remain explicit.

For a measured active fold, projected cost remains spent +1.25×((200−completed_epoch)×maximum observed epoch time +remaining validation-count×maximum of original/observed validation timing)+120. This preserves the original timing formula. Completed-fold actual cost replaces its former projection. A projection is never allowed below its currently charged actual cost.

Never-started folds reserve the median of current started-fold projections. Before any epoch measurement, initial reserve is151517.68659210298/5 =30303.537318420596 per fold. For a started fold still awaiting its first measured epoch, use the conservative maximum of its actual charge, initial equal share, and measured-fold median, then include that provisional value in the started-fold median. This explicit intermediate-state rule avoids treating an active but unmeasured fold as zero work; it is included in the narrow review. A completed started fold contributes its actual cost to that median.

Total projection is the sum of started-fold projected/actual costs, pending-fold reservations, and7200 postfit. It must fit the new total158717.68659210298 and phase3 total612000. Per-fold projections are recorded but do not independently stop a fold. The40000 GPU-s per-fold ceiling is a safeguard on actual charged elapsed, not a projection allowance and not a five-way resource reservation. The existing guard's conservative setup/termination safety margins, sampled RSS/VRAM/CPU/disk protections and global actual-training boundary remain; no extra budget is created. A total overrun or actual runaway stops/reports without extension or another continuation.

Observations and scope

After each completed-epoch receipt, the coordinator records assigned GPU, currently visible GPU processes (UUID/PID/name/memory), own v2 worker identity and host load. The nvidia query has a5s timeout. Observation errors are retained as OBSERVATION_INCOMPLETE; this information never selects a device, changes the cost formula or attributes causality. Snapshot time is recorded; it is not continuous coverage of the preceding epoch. Snapshot elapsed time remains charged through the existing worker/guard elapsed ledger and is captured before constructing the new cost observation. No foreign process is stopped or modified. The prior epoch34 slowdown cause remains unproven, as Claude1794990899 accepted.

New root /home/frank14f/MeltMind-runs/phase3-seed1-v2-2026-11-18 is unique and refuses existing preparation or started markers. All three generations are included in duplicate detection and retain shared GPU lock names. Source copies isolate the new logic from both immutable failed runs; worker and guard differ only in import/child-entry names. Original scientific comparison SHA remainscba51291811c224966ccab82ff64a9b0690c83eaf2482316853ec9c656322493.

Twelve CPU synthetic checks passed first attempt: original science/guard/helper invariance, original projection formula, all generations detected, old ledger exact once, initial reserve, odd/even started medians, completed actual replacement, active-unmeasured reserve, total exceed despite safe actual, projected>40000 allowed when total fits, actual>40000 rejected even when total fits, and bounded observational snapshots/error preservation. No actual GPU query, data/model arrays or training was run by these tests. Log artifacts/phase3-seed1-v2-acceptance-2026-11-18/attempt-01.log.

Status: implementation/contract for new narrow MeltAssist logic-and-ledger review, not launched. Actual new review must bind implementation commit, contract SHA, proof SHA, unchanged comparison SHA and full code map. Old PASS cannot substitute. The present handoff is complete when the binding is posted for review; it does not claim a review verdict, launch, scientific result, prediction/scoring or downstream item1b gate.