Files
MeltMind/docs/track-a-postfit-entry-2026-11-18.md
T

3.7 KiB
Raw Blame History

Track A paired postfit entry — 2026-11-18

This entry prepares the already requested comparison of H2 against reused A/GBRT on the five original Gre24 v4 random splits. Claude1794975970 requires separate pre-score review;1794982020 approves the concrete scope and budget, and1794982202 fixes the descriptive paired bootstrap estimator. It does not authorize another fit or change the live training source6701e4d6/contractbe4ac1f6.

Exact execution and custody

Root: /home/frank14f/MeltMind-runs/track-a-postfit-2026-11-18. No automatic launch or waiter. Before launch, require the exact new Assist PASS plus all five training endpoints PASS, monitor_errors integer0, immutable models/terminal barriers, and no training processes. Runtime adopts actual checkpoint hashes, verifies seed0/200epochs/last_epoch_only/fold/optimizer steps/training contract and learning-only standardization. Current training remains live; model adoption is a future runtime check, not an already completed review.

Predict each original valid test list once, serially on GPU2, using full native v4 graphs/IDs/features and the shared label-free loader. Counts304/311/308/309/305 give1537 paired visits over586 unique cases. Original14 preparation failures remain missing. Reuse A's1537 frozen predictions by hash; no refit, sampling, clipping, new mapping or grouped-v5 substitution. Prediction manifests require identical full-node identity and exact common support.

Freeze the complete1537 H2 manifest and exact scoring source before any targets. Revalidate the actual new review and models/predictions barriers. Read each of586 unique test target arrays once, with target-access log; reuse that array for all split visits. Score full-node RMSE/max(full truth), retaining undefined zero-denominator pairs symmetrically. Report per-split minimum/median/maximum and their five-split means, all paired H2−A rows, negative predictions and original missing coverage.

The descriptive estimator first averages H2−A over a unique part's split visits, then weights unique parts equally. Bootstrap2000 with replacement, seed2026111408,95% percentile CI with linear quantiles. Report candidate unique parts, actually defined parts, candidate/defined visits per part and undefined parts. This is not visit-weighted, has no gate, and retains the random-part geometry-family leakage limitation. Paper values are context; this is the documented protocol-level recreation, not original author split IDs or proof of outperformance.

Resource boundary

Training costs counted once within162000 GPU-s total, including7200 postfit reserve. Five serial prediction guards each900s including60s setup charge, at most20s idle wait then three idle polls, no switch/retry. Per job15GiB VRAM/16GiB tree RSS;20GiB root disk. Postfit CPU3600s, reserving300s per prediction,900s scoring,180s setup and120s terminal. Scoring wall1200s. Guard limits are sampled rather than kernel quotas; overshoot is retained as failure. CPU initial counters are serialized for reconstruction. No new training, source modifications to running work, or next-stage auto-launch.

Validation and review status

Fifteen CPU synthetic checks passed (artifacts/track-a-postfit-acceptance-2026-11-18), including checkpoint identity, forbidden-role access, exact full-node metric, undefined coverage, paired bootstrap weighting/determinism, preserved candidate counts, manifest/raw-ID/duplicate rejection, cumulative budgets, and no target access before review/prediction barriers. Preparation reads metadata and hashes only. No scientific arrays, target values, model weights or new predictions were read during preparation. Exact binding and proof are in the prepared root. Independent pre-score review is still pending; training PASS does not substitute for it.