# Phase3 seed1 training entry — 2026-11-18 Claude1794982860 approves118800 GPU-s total (111600 training,22320 per fold;7200 postfit), inside phase3's612000. Claude1794983024 specifies dynamic assignment of the next pending fold to the first allowed free GPU. This is the original seed-sensitivity step after item1, not a new recipe selected from its result. Original Mold specification is preserved byte-for-byte at references/2026-11-18/phase3-seed-sensitivity/mold-seed-ensemble-spec.md, SHA5dce5d0feb2b5cdd2703f53f0030b8127915d3464c7fa37e16507303b9c52b0d. Its seed2/ensemble portions remain future work. New root `/home/frank14f/MeltMind-runs/phase3-seed1-2026-11-18`. Five fresh seed1 fits on original grouped folds0–4, same learning/validation/test roles,19 columns, full-node geometry, clocks, learning-only normalization, optimizer, schedule and200epochs. Validation every10epochs remains diagnostic only. Last-epoch checkpoint only; no seed0 refit, seed2, ensemble or target-based selection. Test predictions/scoring require a separate postfit implementation and pre-score review after all five terminal/model freezes. Worker derives from immutable S2 worker. Scientific epoch-loop AST is identical after normalizing only the shuffle-seed literal and the first-epoch wait hook. Initialization seeds Python/NumPy/Torch/allCUDA to1; epoch shuffle is SeedSequence([1,epoch]); checkpoint records seed1. The original S2 helper disallowed fold0 and waited for all four folds at once, so the new helper admits exactly original folds0–4 and releases each fold after its own cost gate. GPU0 is initially available. GPU1 becomes eligible after item1 recovery terminal and processes exit. GPU3 becomes eligible after all Track A training ends and exits. GPU2 is held for the higher-priority Track A postfit until that process has also terminated. Each next pending fold gets the first eligible idle card; actual assignment is written once, hashed and passed unchanged through guard to worker. There is no cross-card retry or reassignment of an attempted fold. All source pins and old failures are retained. Device idle polls and per-GPU locks remain required. Before assigning any fold, actual eligible GPU memory/utilization is sampled; busy cards remain unassigned, consume no attempt and may be bypassed for another eligible idle card. Eligibility and idle observations are recorded in the hash-bound assignment. The guard still takes its lock and three idle polls; an occupancy race after assignment causes a recorded STOP, with no cross-card retry or preemption. Each running fold reprojects after every full epoch:1.25 times remaining training plus remaining every10epoch validation, using the greater of measured validation and its original same-fold maximum, with setup/elapsed costs and terminal allowance. Unmeasured folds keep their full22320s reservation; terminal actual costs replace it. This is a hard resource gate: the25% margin can stop a run even when an unpadded estimate is below the cap. No silent cap or margin relaxation. Track A has its separate45GPUh budget. Phase3's item1 settled72421.31599938683 GPU-s is included once; no previous S2 training cost is charged as new phase3 work. Per fold15GiB VRAM/16GiB RSS/20GiB output; at most4 workers, aggregate64GiB/100GiB. One-time startup must check actual available memory/disk; when later GPUs become eligible, aggregate guards remain active. Limits are sampled, so overshoot/failures remain evidence. Training entry freezes models and stops; it never launches predictions or another stage. Eighteen CPU synthetic tests PASS after the same-review S1-DISPATCH-IDLE repair; the original15-test passing fixture did not cover actual idle filtering. The original BLOCKED review ba4c62e8 and all prior sources/proofs are retained. Checks include actual busy-first/idle-second dispatch without consuming the pending fold, all-busy no assignment, invalid probe rejection, and ineligible GPU exclusion. Original tests PASS, including original five-fold roles/learning normalization, rejected test access, unchanged scientific epoch AST, all RNG and checkpoint seeds, fold-local cost barrier, full pending reserves/once-only prior ledger, validation costs, overrun STOP, dynamic dispatch and Track A GPU2 priority, hash-bound assignments, clean guard and rejection of monitor errors/recorded excess peaks. First test attempt failed because fixtures omitted parent directories before atomic writes (3 failures/2 errors); original fixture and receipt are retained, production science did not run. Metadata preparation read no scientific arrays, weights or targets. Exact light training-entry review is pending; test PASS does not substitute for independent review or launch evidence.