Useful framing — one distinction worth adding for the enterprise audience this series is aimed at: HF's Safe labels answer 'did the platform scan this file?' They don't answer the question an org's own release process has to answer: 'did these exact bytes pass OUR checks?'
Those drift apart in practice. A model gets converted, quantized, or re-uploaded under the same name; the label survives, the bytes change, and the eval evidence stays attached to the name rather than the artifact.
The shape we've landed on is binding three things together at admission time: the artifact's SHA-256 digest, the evaluation contract it was checked against (pinned — same suite, same thresholds), and a signed receipt tying them together that the load side can verify offline before anything runs.
MFENX publishes a complete admission record anyone can verify in five minutes — and the same free check rejects the model if one byte changes: https://mfenx.com/verify/
Happy to gate a public model mentioned in this series as a worked example, if that's useful.