Rival-Directed Verification for AI-Assisted Drug Development

Large language models and agentic systems are changing how science is practiced. Researchers can now generate plausible hypotheses, run the analyses that support them, and report candidate findings faster than those findings can be evaluated. Drug development shows the imbalance clearly: AI systems help propose molecules, nominate targets, and assemble mechanistic accounts, while the assays, animal studies, and clinical trials that would test them remain limited by biology, cost, and regulation. Deciding which findings warrant that capacity has become the binding constraint. The standard for that decision is not new, and it does not depend on where a finding came from. Accounts of how knowledge grows differ in their details, but they converge on a demand that is rarely met in practice: a claim earns credibility only through exposure to possible failure. We argue that meeting this demand systematically is now both necessary and, for the first time, achievable. A passed verification test should therefore contribute support only if it was capable of detecting the rival explanations it is intended to address, a capability that emerging AI-science verifiers do not record. We operationalize this in CounterScreen, a framework that commits a finding to explicit falsifiers and named rivals before any test is run. For each rival, we propose rival-conditioned detection power (RDP), which estimates how often a committed test would fire in worlds where that rival, rather than the finding, produced the evidence. In a synthetic pilot, we show that tests that would be reported in identical terms, including two that both adjust for the same confounder, differ almost completely in whether they could have caught the rival. We propose an evaluation of whether RDP predicts held-out flaws and improves the selection of findings for scarce experiments.