·
18 commits
to main
since this release
ChronosGuard-TS v0.3.0 freezes the preregistered ETTh1/ETTm1 rolling-origin forecasting evidence at commit ce535d5.
Verified scope:
- 2 datasets x 4 expanding folds, 7,928 pooled Test origins, OT-only, single seed.
- Compact PatchTST beat direct Ridge by MAE in 1/8 dataset/horizon comparisons; the other seven negative results are retained.
- Only ETTm1/96 reached nominal 90% empirical interval coverage; no coverage guarantee, confidence interval, or significance claim is made.
- The attached wheel is the audited v0.3 artifact: 124,433 bytes, SHA-256 9c1f1d4ddbef11acaeded19805afe01027544e09b573ef1d1842ee45c0516826.
Publication boundary:
- No raw ETT/NASA data, row-level forecasts, checkpoints, NPZ artifacts, or private telemetry are redistributed.
- This release does not include or authorize any v0.4 ETTh2/ETTm2 metric. The later v0.4 development attempt remains fail-closed after independent artifact verification rejected stale staging paths.
- Results are offline engineering evidence, not an official PatchTST reproduction, Transformer superiority claim, production SLA, or business-impact claim.
Known validation caveat:
- The immutable tag's historical GitHub Actions run is red for repository plumbing that was corrected on later
main: the then-current[dev]resolver omittedtorch, and a NASA audit test retained an obsolete SHA assertion. - The frozen v0.3 tag and wheel were independently revalidated after release: clean Python 3.11 install,
pip check, synthetic smoke, and the ETT frozen-artifact verifier all passed. The tag remains unmoved; currentmainpasses the full test and packaging matrix.