SIRDA

How should disputed technical evidence become a decision?

Evaluators can agree on the test result and disagree about its significance. The process must expose uncertainty and specify how evidence becomes an authorized response.

SIRDA editorial hypothesis

Technical findings can state tested configurations, uncertainty, and competing interpretations while a separate decision process applies a published burden of proof. The response can then be limited by severity, reversibility, and the strength of the evidence. This hypothesis fails if a disputed metric silently determines policy or if uncertainty blocks every bounded response to a credible imminent danger.

0 objections ยท View challenges

Challenges to the editorial hypothesis

No published objections yet.

The argument

A capability result, an estimate of harm, and a judgment about acceptable risk answer different questions. A complete record separates them. It shows what was observed, what was inferred, and which assumptions connect the two. AISI's evaluation research provides one concrete reason to state the computational conditions of a test.

An authorized panel then applies the rules governing decisions under uncertainty. Those rules need to specify who carries the burden of evidence, when an additional test is warranted, and how temporary restrictions differ from continuing prohibitions. Dissent belongs in the record, including explanations of what additional evidence would change an evaluator's view.

Scientific evaluation should inform policy. It should not silently become policy. An evaluator can contribute technical expertise while the constitutional procedure remains responsible for deciding which consequences that finding authorizes.

How to test it

Give independent teams the same evidence and ask them to distinguish observations, inferences, and value judgments. Repeat the exercise with a false positive, an incomplete confidential record, and a time-sensitive threat. Examine both erroneous restriction and unjustified inaction.

Sources

More compute, more capability: Why AI agent evaluations need to account for test-time compute

Experiments showing that test-time resource limits affect measured agent capability.

Related Ecosystem

Related questions

Related failure conditions

Discussed in

Contributions

No published contributions yet.