agent-evidence-vectors v0.37.0

Checking an open-weight model release against AI Act Article 53(2)

The proposed CRA and AI Act table in orcwg/cra-hub#364 has an open-source meeting point. For a provider of an open-weight general-purpose AI model, Article 53(2) lifts Article 53(1)(a) and (b) but not (c) and (d), and it does not reach models with systemic risks. Whether a release qualifies depends on what was released: a free and open-source licence, public weights, public architecture and usage information, and a model that does not have systemic risks. A claim about public bytes can be tested against a record of those bytes.

This corpus tests such a record today: vectors-open-weight-art53. It is a check of what a release record carries. It is not legal advice, and it reaches no conclusion about any real model.

What the Article says

EU AI Act, Article 53(2): "The obligations set out in paragraph 1, points (a) and (b), shall not apply to providers of AI models that are released under a free and open-source licence that allows for the access, usage, modification, and distribution of the model, and whose parameters, including the weights, the information on the model architecture, and the information on model usage, are made publicly available. This exception shall not apply to general-purpose AI models with systemic risks."

Points (a) and (b) are the technical documentation for the AI Office and the information for downstream providers. Points (c) and (d), the copyright policy and the public training-content summary, are not exempted. Two other texts bound the exemption: Article 51(2) presumes high impact capabilities above 10^25 floating point operations of training compute, and Recital 103 withholds the exceptions from AI components provided against a price or otherwise monetised, including through personal data used for reasons other than exclusively improving security, compatibility or interoperability, with an exception for transactions between microenterprises.

What it checks

A provider (or a reviewer, or an integrator) writes a small JSON record of public locators for a release. The checker reads six conditions and answers exempt, not-exempt or not-established:

Condition What the record must carry
licence an SPDX identifier on the pinned open list, none of the restricted ones (non-commercial, no-derivatives)
weights access: public, an https locator and a SHA-256 over the weights
architecture an https locator
usage an https locator
monetisation an explicit empty list of monetisation conditions
systemic risk stated training compute at or below 10^25, no Commission designation, or a recorded Commission decision that the model is not systemic

Silence never passes. A missing field is not-established, a visible failure is not-exempt, and a failure outranks a gap. The answer also reports whether the record carries the copyright policy and the training summary, because the exemption leaves them in force.

Run it

pip install agent-evidence-vectors
agent-evidence-vectors --corpus vectors-open-weight-art53
python3 -m agent_evidence_vectors.openweight release.json --require-exempt

The first command runs the packaged cases through the reference reader. The second judges your own record and exits 1 unless it is exempt, so it works as a CI step in the repository that publishes the model. A record looks like corpora/vectors-open-weight-art53/cases/exempt-open-release/release.json. To judge another checker: agent-evidence-vectors --corpus vectors-open-weight-art53 --verifier './verifier'.

What it does not check

Recording activity

The table's recording meeting point has its own guide: Checking an AI Act log after the fact.