EVIDRA: Evidence-Guided Intervention Selection for Research Agents

Scientific discovery requires research agents to decide how to revise their process in response to experimental evidence. We present EVIDRA, a framework for evidence-guided intervention across prompts, workflows, context, memory, tools, and sampling. Each intervention links a testable hypothesis to an edit target, a recoverable baseline, and expected observations. The controller prioritizes interventions by expected improvement, information value, cost, and risk, then combines scoped execution, rollback checks, and independent assessment to build reusable research evidence. On four graph-optimization tasks, EVIDRA achieves an equal-task mean gap-closure score of 0.8736, compared with 0.8682 for the strongest baseline, and leads on Cluster Editing and Treedepth. Three mathematical studies yield an exact counterexample to Lonely Runner translation monotonicity, an improved Erdős–Straus sieve-weight certificate, and a tighter certified upper bound for a four-variable iid probability problem.