What Can Self-Investigating AI Establish? Joint Uncertainty over World Models, Access Models, and Interventions

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

What Can Self-Investigating AI Establish? Joint Uncertainty over World Models, Access Models, and Interventions This paper develops a methodological framework for evaluating self-investigating AI systems that do not merely revise hypotheses about a target environment, but may also need to revise assumptions about how evidence is generated. The central distinction is between a world model WW and an access model Π\Pi: anomalies in experimental results may arise from incorrect assumptions about the investigated system, but also from faulty assumptions about sensors, resets, calibration, intervention semantics, measurement invasiveness, or other properties of the experimental interface. The framework combines intervention-relative identifiability, explicit claim ceilings, access-model auditing, and a structured taxonomy of epistemic outcomes. A sequence of constructive single-system pilot cases illustrates parameter revision, model-class revision, access-model revision, unresolved causal locus, non-identifiability, and unavailable separating interventions. One pilot additionally motivates a practical distinction among design truth, implemented causal truth, and manifested experimental truth, showing how benchmark labels can diverge from the causal system actually executed. The paper does not propose a new theorem of identifiability, nor does it claim that AI can certify the completeness of its own representational space. Its contribution is a methodological synthesis for autonomous research agents: structural claims should be limited to what can actually be distinguished under the available hypothesis, access, and intervention classes. Successful self-revision can expand a represented epistemic space, but it does not by itself establish that the enlarged space is complete.