Verifiable Model Record In preparation

An open standard

A signed record of what an AI model is, and who stands behind it.

Anyone can check a Verifiable Model Record (VMR) offline, against the issuer keys they choose to trust. No network, no account, no vendor in the middle.

What a record states

A record carries hashes and signed claims, never the model's weights or its training data.

The model
Its files, each identified by a cryptographic hash.
Its training input
Committed by hash, so it can be matched later without being revealed, or stated as not disclosed.
Its lineage
The models it was built from, and how: fine-tuned, adapted, merged, quantized or distilled.
Its environment
The hardware and software it was produced with, by hash.
Its issuer
Who signed the record, and when.

Checked by anyone, offline

A verifier checks a record's structure, its signature, and whether the issuer's key is one the verifier has chosen to trust. It needs no network connection and no account, and it names the check a record fails.

Graded against the rules that apply

Policy packs are machine-readable rule sets. A pack grades a verified record rule by rule against the laws and standards it cites, and each rule answers compliant, non-compliant or indeterminate, with the clause it encodes.

A "compliant" answer means the record passes that pack's rule. It is not a legal finding, and reference packs are not legal advice.

For any AI model

VMR is designed for any AI model, from any vendor, whose weights the issuer holds. Only a party that holds a model's weights can issue a truthful record for it.

Status

  • The specification, version 0.1, is in preparation. Nothing is published yet.
  • When it is published, the specification and its schemas will be released under Apache 2.0 and its test vectors under CC0, so any implementation can check itself against them.
  • Free, open-source reference tools for making and checking records (Apache 2.0) will be released with it, together with reference policy packs for the EU AI Act, the NIST AI RMF, ISO/IEC 42001, C2PA AI Disclosure and IETF RATS.