25 Luglio 2026 - ultimo aggiornamento 2 mesi fa
POLICY DECISION REQUESTED
An AI system should not be described as reliable merely because it completes a task once or avoids a prohibited disclosure. For stateful work, reliability also requires evidence that the authorized state needed to complete the task survives the sequence.
Apply AI Alliance should convene a six-month technical working group to draft and pilot a Stateful AI Evaluation Protocol. The protocol should produce a core state-contract schema, minimum reporting requirements, three sector pilots, a public-procurement model clause and a formal contribution to European AI standardization.
A Policy Standard for Evaluating Stateful and Agentic AI in Europe
Year of publication
2026
- Effettua il login per postare commenti
Commenti
Matthew, this is a rigorous and much-needed contribution.
You've identified a critical gap: current evaluation frameworks test for correctness and boundary compliance, but not for state integrity across complete workflows. The R2 finding that a system can maintain perfect boundary compliance while losing required permitted state in 25% of executions is precisely the kind of failure that traditional benchmarks miss—and that real-world deployments cannot afford to ignore.
What I would add is that state integrity evaluation is not complete without an auditable evidence chain that ties runtime behavior to director liability. A system can pass a sequence-based test today and drift tomorrow. The policy question is not just "how do we test state integrity at a point in time?" but "how do we continuously verify that state integrity is maintained, and how do we prove it to regulators, procurers, and affected parties?"
This is where evaluation meets Article 14 (human oversight) and Article 72 (post-market monitoring). The evidence package you describe—preregistration, execution records, hashes, closure records—is precisely the kind of immutable audit trail that makes stateful integrity claims verifiable, not just assertable.
The SAEP proposal is a strong foundation. I would encourage the working group, if formed, to consider evidence-chain requirements as a parallel track to the measurement protocol itself. A state contract and sequence-based cases tell us what to measure. An evidence chain tells us how to trust the measurement.
Would be keen to explore how the community thinks about integrating auditability into the SAEP framework.