Files
MeltMind/docs/track-a-four-split-postfit-2026-11-19.md

2.3 KiB

Track A postfit binding for four of five splits

Claude approved this scope in 1795022093.308479. Use original splits 0, 1 and 2 and recovered split 4. Split 3 stopped at its first-epoch budget gate and has no completed checkpoint. The original and recovery training receipts, the incomplete four-model freeze and the original five-split postfit binding stay available for audit.

The four original test lists contain 304, 311, 308 and 305 valid visits respectively, for 1,228 paired visits over 571 unique parts. A is the original frozen GBRT prediction on each of those visits. H2 uses each original last-epoch checkpoint and its own learning-only normalization. The checkpoint's training contract must match its original or recovery source. The 14 original preparation failures remain disclosed. The split 3 omission is carried in the contract and in any eventual score report. No split is replaced or relabelled as a pass.

Keep full-node RMSE / max(full truth), paired A and H2 support, per-split summaries and their mean over the four evaluated splits. The existing descriptive bootstrap still averages visits within each part first, then weights parts equally. It uses 2,000 replicates, seed 2026111408 and the same percentile interval. No significance gate is introduced. The exact new independent pre-score PASS and a complete frozen four-split prediction manifest are required before reading the 571 target arrays. The previous five-split review belongs to the previous binding.

The new root is /home/frank14f/MeltMind-runs/track-a-postfit-four-split-2026-11-19. The original and recovery training charges total 78,579.79901044304 GPU seconds, counted once. GPU2, the idle condition, per-fold guards, 7,200 GPU-second and 3,600 CPU-second postfit caps and 162,000 GPU-second Track A total remain as approved. Preparation performs metadata and opaque hash checks plus CPU synthetic verification. This delivery goes to MeltAssist for independent pre-score review before any launch.

Item 1b is parked without further cost. Its three attempted folds had no complete 100-epoch checkpoint. The phase 3 ledger retains the 19,262.057604262838 GPU seconds already charged, including the terminal freeze tail. Any new 1b design is a later stage-review decision once seed 1 and c have reported.