Files
MeltMind/docs/phase3-data-sufficiency-entry-2026-11-18.md

4.5 KiB

Phase3 item1: data-family size versus optimizer updates

Authority: Claude1794946961.453149 triggers plan304f5771; mathematical specification e24fff7c5cb91dff7acf1532f502daec60e4b7fec21b2d4d4dcde1e793be7fb4. No other phase3 item launches here. S2 remains terminal; gate-v2 failed, so fresh-family admission remains closed.

This entry creates four fresh fold0/seed0 H2 fits. The original nested grouped learning lists are25/50/100 percent, with common validation111 and test121. Actual component/part/node/edge counts and source exposure coverage are frozen in contract.json. No new geometric category label is inferred; a canonical geometric-category ontology is not available in these input contracts, and that coverage is explicitly unavailable. Components are the independence unit, not nodes or time slices.

Two same-recipe cells (25/50 percent) use the unchanged200-epoch learning-rate function and last-epoch selection. The100-percent endpoint reuses the original S1 checkpoint and test predictions/metrics with exact identity checks: same roles, seed, zero-residual initialization,19 columns, full-node graph, per-node training loss, clock, normalization formula, schedule,200epochs and selection rule. Training and validation evaluation for that endpoint may require frozen-checkpoint inference in the separately bound postfit; no refit is allowed.

The other two cells (50/100 percent) each run fresh to64000 updates, saving U=32000 and2U=64000 on one trajectory. For1<=s<=3200, lr=1e-4+(s-1)9e-4/3199; for3200<s<=64000, lr=1e-5+(.001-.00001)(1+cos(pi*(s-3200)/60800))/2. Both pools share exactly this step schedule; U is not separately decayed. The50-percent pool has190 parts, so its final traversal is partial:336 complete traversals plus160 parts. The original SeedSequence([0, traversal]) permutation is used; no resampling of pool membership. Per-part repetition/node/edge exposure counts are recorded. Equal updates are not equal compute. No3U extension.

All cells retain the original H2 architecture, zero output initialization, fresh seed0, FP32 (no AMP/TF32), FP64 cF reconstruction, full nodes/true edges, node-weighted loss Nparts*Ni/Ntotal, AdamW and clip1. Each pool has its own original learning-only clock and recomputed learning-only feature normalization; those clocks were originally computed on exactly these lists and are reused by hash. Diagnostic validation targets are permitted; they never choose checkpoints or stopping. Same-recipe diagnostics every10 epochs; step trajectories every3200 updates. The online training loss is explicitly not a frozen-model training-domain metric.

The complete comparison contract separately fixes the later common-domain training/validation and one external evaluation, cF-only baselines, field/last/tau, component-paired2000-replicate joint contrasts, actual132/65 component slope denominator, delta=.001, classification including undefined/wide intervals. S1 external fold was already used for stage decisions; this is diagnostic reuse. A single fold/seed interval omits training seed and pool sampling uncertainty. No extrapolation to required CAD counts or blanket claim that data limitation is ruled out.

This executable entry stops after all four terminal guards and six new checkpoints are hashed and Git-frozen: TRAIN_COMPLETE_POSTFIT_PENDING. It has no test prediction or target-scoring function. A new exact Assist PASS binding the current implementation, contract, comparison contract and complete source map is required before this training launch. The later postfit implementation must independently bind the resulting checkpoints and get an exact pre-score PASS before prediction/custody. Old S1/S2 reviews never transfer. This staged engineering implementation does not add a new user authorization gate.

Budget: item1 total86400 GPU-s (24GPUh), within the new phase3 612000 GPU-s (170GPUh). Training allowance79200, reserved postfit7200. Four job caps10000/16000/26600/26600 GPU-s, including setup/import/idle/waits/failures; CPU limits equal each job cap. Per job15GiB VRAM/16GiB RSS/20GiB output, four GPUs at most; aggregate64GiB RSS/80GiB. First full epoch of every cell pauses until an aggregate1.25-safety projection passes all individual caps and item total; subsequent5-percent progress checkpoints refresh it. Source pins stay unchanged; no retries, threshold changes or automatic extensions. All terminated work remains charged. The queue freezes terminal checkpoints and stops; postfit costs are subsequently added to this same item ledger. Metadata preparation and CPU synthetic checks are separately recorded, with no GPU workloads.