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

3.9 KiB

Phase3 item1 v2 — separate complete update-budget schedules

Claude1794948643.973609 amends e24fff7c's common-trajectory design; Claude1794948710.366709 fixes the total item cap at120000 GPU-s within phase3's612000 GPU-s. The earlier c8e2ee4c/a84e5e82 entry is preserved, never started, superseded. This is a labelled design revision, not unchanged execution of e24.

Original fold0 pools remain nested:34/65/132 learning components,111/190/320 parts; validation111/test121 unchanged. Same-recipe25/50 pools get fresh seed0,200epochs with the S1 epoch schedule. Same-recipe100 endpoint reuses S1 final checkpoint and existing external predictions/metrics by exact identity; no refit. Each pool reuses its own previously frozen learning-only clock on exactly the same list, recomputes learning-only standardization, and retains the same zero-residual initialization/full-node FP32 graph model/FP64 skip/node-weighted objective/AdamW/clip1. Geometry-category coverage is unavailable because no frozen category ontology exists; component/source-exposure and node/edge coverage are reported without inventing categories.

Step-matched50%-U,50%-2U,100%-U,100%-2U are four separate fresh seed0 fits. U=32000,2U=64000. Each has3200 warmup steps from1e-4 to1e-3, then cosine to1e-5 at its own endpoint. Formula: for s<=3200,1e-4+(s-1)9e-4/3199; afterward1e-5+(.001-.00001)(1+cos(pi*(s-3200)/(S-3200)))/2 with S32000 or64000. Only the terminal checkpoint is adopted per fit. No common-trajectory U checkpoint reuse. g_n now measures the joint change in update budget and complete LR schedule, not pure additional updates along the same trajectory.

There are six new fits/checkpoints. IDs0/1 are epoch25/epoch50; IDs2/3 are50%-U/2U; IDs4/5 are100%-U/2U. FixedGPU assignments repeat0/1 only after prior jobs release them; at most4 simultaneous workers. Each traverses its fixed list with SeedSequence([0,traversal]); partial final traversals are retained. Node/edge exposure and repetitions are reported. Validation diagnostics every10epochs for same-recipe or3200steps for update arms never select weights. The online loss is not a frozen-model training-domain error. New test predictions and target scoring are absent from this executable entry.

The comparison contract fixes same-recipe and fixed-budget component-equal field/last/tau, separate cF-only baselines for all three clocks, training-own-domain and common validation/test evaluation, and jointly paired2000 bootstrap replicates(seed2026111408). Family slope divides bylog2(132/65), delta=.001. Update effect includes schedule changes. Wide/undefined intervals are not equivalence; one fold/seed excludes training/pool sampling uncertainty. S1 test reuse is disclosed. No3U, newfold/seed, fresh-family admission or CAD-count extrapolation.

Budget: training112800 plus postfit reserve7200 =120000 GPU-s, phase3 total612000 unchanged. Per-jobGPU andCPU caps11000/17000/14000/28400/14000/28400;15GiB VRAM,16GiB RSS,20GiB output each, aggregate64GiB RSS/80GiB. Setup/import/idle/waits/failure time charged. Every fit pauses after its first complete epoch until a cell-specific projection passes; the total includes settled cells once, current measured projections and unstarted-cell conservative projections plus postfit reserve. Refresh at each5-percent checkpoint. Stop on individual/aggregate excess, monitor/identity failure, no retries or cap changes. No dependency on all six being simultaneously active.

After six terminal passes, hash and Git-freeze six checkpoints, then stop TRAIN_COMPLETE_POSTFIT_PENDING. A new exact Assist PASS on this implementation/contract/comparison/full source map is mandatory before training launch. Postfit is subsequently implemented and bound to these checkpoints, with another exact pre-score PASS; old S1/S2 reviews never transfer. Existing user continuation authorization suffices; these are engineering identity reviews, not new user-go loops.