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

3.3 KiB
Raw Blame History

Seed1 one labelled continuation

Claude 1794984996.938059 authorizes one fresh restart of folds 0/1 and first attempts of 2/3/4, with unchanged seed1/200 epochs, science recipe, max-epoch ×1.25 projection, per-fold first-epoch gate and actual idle dispatch. GPU2 remains reserved for Track A postfit. A further cap exceed stops and reports; no automatic extension or retry.

Claude 1794985731.591689 explicitly confirms the old 539.0720570000121 GPU-s is inside the final cumulative seed1 ceiling 167200, not an extra allowance:

  • Cumulative per-fold cap: 32000 GPU-s. New fold0 cap is 32000−269.53611274203286; fold1 is 32000−269.5359442579793; folds2/3/4 remain32000.
  • New-run training cap: 160000−539.0720570000121 =159460.927943 GPU-s. New-run total cap: 167200−539.0720570000121 =166660.927943, including unchanged7200 postfit reserve.
  • Runtime receipts and projections describe new-run cost. Add old539.072057 once to report cumulative seed1. Phase3 prior already equals original item1 72421.31599938683 + old seed1 =72960.38805638684; add only new-run cost/projection to it. Never add old seed1 again. Phase3 ceiling612000 remains unchanged.
  • Original STOP has no serialized coordinator CPU total; this limitation remains. It is not filled with estimated CPU.

Original root /home/frank14f/MeltMind-runs/phase3-seed1-2026-11-18, implementation ac45e2de, contract95d23536, budget-stop freeze a556b583 and published a45e375d0e are immutable. The old STOP, partial two epochs, original PASS/BLOCKED reviews and all failures remain. Both old attempted folds start fresh, without partial-checkpoint adoption. Scientific comparison SHA remains cba51291811c224966ccab82ff64a9b0690c83eaf2482316853ec9c656322493.

New root is exclusively /home/frank14f/MeltMind-runs/phase3-seed1-continuation-2026-11-18. Isolated worker/guard/queue copies differ only by their import/child-command names, preserving old pinned source. Common functions for scientific roles, projection, dispatch, real idle checks, GPU2 release priority and assignment validation are byte-identical. Common contract validation checks the new remaining constants against original STOP receipts and includes both entry generations in duplicate detection. GPU locks retain their original shared names. No unrelated counters, monitor logic or scientific fixes are included.

Six narrow CPU checks passed on first attempt: unchanged worker/guard/queue except entry names; unchanged projection/idle/release helpers; duplicate detection across both generations; full pending caps plus old charge exactly167200; settled new cost and phase3 old accounting once; fixed fold/total exceed STOP. Log: artifacts/phase3-seed1-continuation-acceptance-2026-11-18/attempt-01.log. No datasets, model weights or CUDA workloads were read/run by this preparation.

Status at preparation: no new run, review or launch. Require MeltAssist's actual new PASS bound to this implementation/contract/proof/comparison/full code map, then launch once with current host-load, RAM/disk, real idle and duplicate checks. The old PASS cannot authorize this binding. Foreign GPU jobs are avoided and never stopped; a current host-load snapshot does not establish the cause of the old slow epoch. Track A remains unchanged. No seed1 prediction/scoring or later item1b scope is implied.