What makes human–AI research cumulative? A hypothesis on conditional decisions and the growth of research capability

AI systems can produce more answers without creating a more reliable capacity for further inquiry. We propose that part of the growth of human–AI research capability may depend on forming, selecting and revising conditional decisions: which evidence supports an action, where a method applies, and when a route should be retained, abandoned or reopened. The hypothesis distinguishes improvements in a model from the ability of a research system to use earlier work under new conditions. Post-training may develop the relevant decision skills; joint research may form and transmit usable rules; neither establishes the other. Reliable reuse could allow one investigation to support many subsequent investigations, but any productivity gain must be demonstrated after accounting for computation, verification, human effort and time. The same account allows compact rules, machine-generated alternatives, and cases where retaining a history is unnecessary or harmful. This perspective states the hypothesis, its observable consequences and its limits. It reports no new experiment establishing the proposed mechanism or a transition to explosive research growth.