Numbat Leak Looks Like Generated Noise

This is a useful reminder that internally consistent model files are not evidence of provenance or training.

• –Reproducible tensors strongly suggest deterministic generation instead of learned weights.
• –The claimed 14T-parameter architecture, 772 shards, and NVFP4 formatting create spectacle but not authenticity.
• –Developers should treat unknown Hugging Face checkpoints as untrusted code and avoid `trust_remote_code=True` without inspection.
• –The incident highlights the need for signed checkpoints, reproducible training records, and independent evaluation before calling a model leaked or real.