Skip to content

Latest commit

 

History

History
46 lines (38 loc) · 3.71 KB

ARTIFACT_RETENTION.md

File metadata and controls

46 lines (38 loc) · 3.71 KB

Artifact retention

On October 27 and 28, 2026, the maintainer requested minimal local retention and stopped bulk cloud archiving. Generated training corpora, obsolete checkpoints, scratch builds, downloaded work files and the dedicated private Drive archive were intentionally retired. Older transport receipts describe reconstruction at the time of those runs. Their Drive IDs no longer provide a current restore path.

The checkout retains source, configuration, tracked compact results, the required runtime, frozen inference executable and network, and the whole original parallel sample, now closed with 1,110 attempt records and 555 pair records. Its initial 262 records and 131 pairs were preserved unchanged.

The original campaign floating-checkpoint path is unavailable, but the selected C76 checkpoint still exists in the tracked results/precision-v4-inputs/selected.pt. A current byte check confirmed all 9,518,976 bytes match SHA256 c76d8e9ef0f2255531bc9085f5d0330a7a4152cf9b21a65ff8aee740ab8b2005 and the original selection identity. This was an existing retained copy, with no acquisition, retraining or selection change. Its availability does not restore the retired bulk corpora or obsolete checkpoints.

Historical domain results remain historical results. This retention decision does not make incomplete gates pass, provide public raw-data reproduction, or permit retraining and model selection from old outcomes. Complete raw historical reconstruction from this checkout is unavailable.

The original parent deadline expired on October 28 at 01:18:17 UTC. At that point, no fifth parallel owner was admitted and 938 attempts were unrecorded. That expired scope was not renewed. All original failures and cumulative limits remain unchanged.

Later on October 28, the maintainer explicitly authorized continued work and restored the follow-up. A new finite continuation records that authority separately from the expired parent. The new storage boundary has source-bound synthetic integration evidence. The separately admitted fifth owner then reached the original cumulative active limit and closed with 1,110 of 1,200 original attempts recorded. Five comparisons were independently checked; the last comparison remains incomplete. No further parallel start or active time is admitted. The follow-up remains active for other authorized work. Historical raw-artifact retirement and the frozen actor remain unchanged. Full acceptance and release remain blocked.

The 2,000 original v4 external PGNs remain local. Their headers, outcomes and color pairs were checked in the final external report binding. Per-move game JSON, pair journals and archived actor configurations are unavailable in that retained campaign tree. The old full independent replay remains an accepted historical result; this retained PGN audit does not reproduce that replay.

The final data/learning lineage check independently bound the original full audit to the final deployment and checked current availability. All240original scheduled checkpoint paths, six full split-array paths and three original audit catalogs are unavailable. The selected C76 checkpoint remains at its retained alternate path, alongside an8192-row validation slice with original parent hashes. Missing catalogs prevent an item-by-item inventory of retired per-game corpus leaves. The actual final G07 adapter accepts the completed historical/source/lineage relationships while explicitly preserving unavailable full raw reproduction.