repro-in-context-continual-learning / reports /published_verification.md
SabaPivot's picture
Upload reports/published_verification.md with huggingface_hub
39596f3 verified
|
Raw
History Blame Contribute Delete
1.85 kB

Published verification — paper #4632

Verified at 2026-07-18 00:01 KST.

The default published-phase challenge audit passes. The strict published-phase audit has no failures and reports only four remote checks as manual. Those checks were resolved independently as follows.

  1. The anonymously downloaded Trackio registry maps both latest and release-v3 to artifact version v2 for the bundle and sidecar.
  2. The anonymously downloaded CAS bundle is byte-identical to the local canonical archive: 6,303,662 bytes, SHA-256 52e5964e0b3399731b4126feb40826651c41eb96430cb549e1373910b08bc293.
  3. The anonymously downloaded sidecar is byte-identical: 123 bytes, SHA-256 3206985bb2bda863562309c1f5c3fa79cb6a0ed7b384be6986054d5a0fa8407f.
  4. All ten files in substantive_n500_v1/SHA256SUMS were downloaded without credentials and passed sha256sum -c.
  5. The public audit-provenance copy of the Job inspect record reports COMPLETED, l4x1, and 1,123 running seconds; the public full log confirms NVIDIA L4 and the saved substantive_n500_v1 payload. The HF Jobs API itself returns 401 anonymously, so these immutable records are retained in the public output bucket.
  6. The Space, both Buckets, and Collection resolve publicly with HTTP 200; the clean downloaded Space contains no local Trackio URI in any page and preserves the required pin order and Hub collection citation.