Brain Researcher: AI-assisted research infrastructure workspace for neuroimaging analyses

AI agents can now write code, run analyses, and propose hypotheses, but a generated output is not a scientific claim by default. A claim becomes trustworthy only when a researcher can see the evidence behind it, the alternatives it was weighed against, the checks it survived and the limits beyond which it should not be interpreted. Here we present Brain Researcher, a workspace-centric infrastructure for auditable AI-assisted science that turns researcher judgment into executable commitments: allowed alternatives, validation rules, provenance records and claim boundaries that constrain what the system may do and what each result may assert. We use neuroimaging as a demanding testbed, where heterogeneous datasets, multi-stage pipelines and flexible analytic choices make automation valuable and interpretation fragile. Across seven foundation models, Brain Researcher lifted tool-selection capability from 51% to 63% at baseline to about 93% and raised the share of claims backed by evidence that could be located and judged supportive nearly sixfold. Applied to four live questions from collaborating scientists, the system returned bounded claim records, with some accepted and others qualified, blocked or rejected, instead of unqualified findings. In two self-evolving campaigns, fixed human-written gates rejected attractive hypotheses, deferred unsupported claims and licensed revised successors. AI-assisted science can become both more capable and more trustworthy when human judgment is part of the executable infrastructure: the researcher's authority moves upstream, not out of the process.