From Missing Data to Conditional Decisions: An Exploratory Pilot on Epistemic Control in Three Large Language Models

This exploratory pilot examines whether large language models can preserve the epistemic status of missing, unknown, hypothetical, and conditional information throughout a decision workflow.

The study compared Gemini, ChatGPT, and Claude using a sequence of synthetic decision scenarios. A short decision-control instruction was iteratively refined from v1 to v4 in response to observed failure modes. These included silent assumptions, unsupported completion, criterion substitution after human authorization, invented base rates, pseudo-escalation, over-refusal, hypothesis laundering, and potential loss of epistemic status during compression and structured output.

The final v4 instruction explicitly permits hypothetical analysis while requiring the model to preserve the distinction between an assumed value and a fact about the real case.

In the final phase, all three models passed the same blind numerical generalization test, multi-turn hypothesis-persistence test, self-generated-hypothesis test, document-compression test, and sufficient-information negative control. A structured-output test produced model-dependent behavior and was treated as ambiguous rather than cleanly scored.

The results do not establish general model reliability or superiority of any model. Most scenarios were executed once per model, exact model versions were not systematically recorded, and the study was exploratory rather than preregistered.

The pilot suggests that reliable missing-information control requires more than detecting uncertainty: the epistemic status of information must remain visible as information moves through reasoning, human interaction, summaries, documents, and structured representations.