Refuse, Decompose, Refresh: A Claim-Safe Protocol f... | AI Research

Refuse, Decompose, Refresh: A Claim-Safe Protocol for Closed-Loop AI Evaluation addresses a fundamental problem in AI testing: benchmarks can be perfectly reproducible while still supporting the wrong conclusions. In closed-loop AI systems, the policy being tested influences which states are visited and which components are observable. This means that if a system never visits a specific state, it is impossible to know if the component controlling that state is functioning correctly. This paper proposes a "claim-safe" protocol to ensure that evaluation results are based on valid evidence rather than misleading metrics.

A Contract-Based Approach

Rather than providing a single "PASS/FAIL" score, the authors argue that evaluations should function as an executable contract. This contract explicitly defines what evidence must exist before a score is calculated, which sampling units support the findings, and which variables represent the intervention being tested. The protocol relies on three core actions: The same ai evaluation question is explored in Beyond Aggregate Scores, which adds a research perspective.

• Refuse: If the system does not generate enough clean evidence or if the comparison between the reference and current data is not supported, the evaluator must abstain from scoring rather than producing a potentially misleading diagnostic.

• Decompose: Instead of a single summary label, the protocol reports execution integrity, operational false admission, and structural hypotheses as separate, independent fields.

• Refresh: When the system’s behavior shifts away from the reference distribution, the protocol treats this as a request to update the reference map rather than as direct evidence of a fault.

Measuring What Matters

A key finding of the study is that how we measure "exposure" to a fault matters more than the size of the fault itself. The researchers compared two ways of measuring the impact of a component fault: "cell fraction" (a geometric description of how much of the component was edited) and "affected clean traffic" (how much behavior actually flowed through the changed region). They found that traffic was a much better predictor of whether a fault could be detected. This suggests that evaluation metrics should be validated based on their ability to predict outcomes, rather than relying on intuitive but potentially flawed proxies. The ai agents story also surfaces in Google opens early access to AI..., adding another angle.

The Importance of Context

The paper highlights that "null" results—where no fault is found—must be interpreted relative to the reference distribution. In their experiments, they observed that clean, fault-free data triggered alarms simply because the demand regime had shifted. If these results were pooled together without context, it would look like a high rate of false positives. By keeping the "drift log" separate from the "fault detection" results, the protocol prevents researchers from misinterpreting a shift in system demand as a failure of the AI model.

Key Takeaways for Evaluation

The authors emphasize that abstention is a valid and necessary part of an evaluation result. When a system is not evaluated because it lacks the necessary observable support, that "not evaluated" status is a meaningful finding that should be reported alongside success or failure rates. By making the denominator of an evaluation transparent and refusing to force scores on unsupported cases, developers can build more trustworthy assessments that accurately reflect the capabilities and limitations of closed-loop AI systems. The same ai evaluation question is explored in When Should a World Model Move?..., which adds a research perspective. as detailed in the full paper on Arxiv