Brain Researcher Raises Agent Tool-Selection Accuracy From 23.3% to 93.6% by Enforcing Analytic Rules

arXiv 2608.19902·medium signalXBlueskyLinkedInCopy link

AI agents can run scientific analyses but reproduce human failure modes including selective analysis, premature declarations of success, and optimizing imperfect criteria, so an output only becomes a defensible claim after alternatives are weighed. Brain Researcher is an agentic harness that operates inside a neuroimaging researcher's own computational environment under explicit rules for admissible analyses, required checks and claim scope. Across seven models it raised first-choice tool-selection accuracy by 70.2 percentage points (23.3% without it to 93.6% with it) and verifiable grounding from 4.6% to 22.0%, with multiverse analyses exposing analytic-choice sensitivity and review classifying claims as accepted, qualified, revised, blocked, rejected or deferred.