Making AI Scientists Auditable from Evidence to Claim
The research introduces XCIENTIST, a research harness designed to make AI scientists auditable by linking evidence to claims. It solves claim drift by externalizing tacit skills like synthesis and validation through a three-layer architecture: a paper graph substrate, a constrained research harness for ideation and testing, and an inspectable UI. This framework enforces strict contracts during the ideation-validation-evolution loop to ensure all scientific steps are traceable and accountable.