An AI-safety partnership needs a public output charter

OpenAI and Poland's NASK National Research Institute announced on 23 October that they had signed an agreement to explore collaboration on AI safety, security and technical research. NASK's account describes the agreement as unique in the EU. Reuters independently reported the announcement. None of the opened items supplies a complete workplan, test protocol or publication schedule.

The useful next artifact is a public output charter. For each workstream, name the question, accountable organisations, intended evidence, expected artifact, review route and target date. Distinguish a public report, restricted technical finding, benchmark, mitigation, training asset and policy note; “research” is too broad to track.

Set disclosure rules before findings arrive. Some vulnerability details should remain restricted, but the charter can still report scope, model or system class, evaluator independence, dates, limitations and whether a finding was remediated or independently closed. Record withheld outputs and the reason rather than silently shrinking the programme.

Add a reproducibility route. Where code or data cannot be released, specify an authorised third party, synthetic test set or aggregate result that can challenge the conclusion. Public institutions should also state what domestic capability remains after the collaboration: methods, staff practice, infrastructure or reusable standards.

The counterargument is that early charters constrain exploratory research. Use ranges and revision notes rather than fixed hypotheses, but preserve changes. A charter can evolve without turning every later output into evidence that the original promise was met.

This article does not claim that the partnership lacks substance or that every result should be public. It asks the parties to make progress auditable without exposing sensitive systems. The immediate decision is to publish the first output charter before describing the agreement as a safety outcome.