Quasar-230k-audit-mirror

Quasar-230k — audit mirror (evidence preservation)

This is a byte-for-byte mirror of silx-ai/Quasar-230k (weights commit a58e814a, uploaded 2026-08-05 12:43 UTC), preserved because the original silx-ai/Quasar-230K repo (created 2026-08-04 19:04 UTC, revision 803b2e57) was deleted by its owners after independent analysis. The re-uploaded repo mirrored here is fingerprint-identical to the deleted one on every tensor checked.

This checkpoint was presented by SILX AI as "the intended Quasar checkpoint" from step 230,000 of a training run. Independent analysis found instead that its tensors are inclusionAI/Ling-mini-base-2.0-20T weights, element-wise permuted and sign-flipped, with light noise:

• ~95% of values in attention/MLP tensors (and 99.9% in norm vectors) are bit-identical to Ling's after unscrambling
• all 80 norm vectors are ~50% negative (trained models: ~0%), with Ling's exact per-tensor maxima reappearing sign-flipped
• new "branch" tensors are at random initialization; gates at cosine +1.0000 to untrained init values
• run live, the model emits gibberish with LM loss worse than random weights

The original SILX README is preserved unmodified as README.original.md. SHA256SUMS contains digests of every file at mirror time.

Nothing in this repository was created or trained by the uploader; it is preserved solely as evidence for the SN24 community.