CONFIRM: Scientific reporting governance for trustworthy agentic neuroimaging analysis

Researchers propose the CONFIRM framework, which separates "running the analysis" from "deciding which findings to report" in agentic neuroimaging analysis: it first fixes conclusions and pass criteria with a machine-readable preregistered analysis plan, then filters results through five deterministic gates—search history and multiplicity, confounding, statistical power, analysis stability, and replication—and offers three reporting tiers: Screen, Replicate, and Full. The authors show that LLM reviewers still support findings with site confounding even when the confound is presented as evidence. On NeuroClaimBench, which contains 489 executable neuroimaging claims and 150 methodological-violation controls, CONFIRM's false-support rate for controls is at most 3%, versus 23%–38% for baselines, while recovering at most 61% of literature-supported claims; when auditing an existing agent, it reduced false support for controls from 23% to 0% while retaining its literature-supported claims; its diagnostic feedback also produced more supported revisions than Self-Refine with fewer LLM calls.