Auditing Executed Work in Computational Biology Agents

Scientific agents increasingly execute multi-step analyses, using intermediate results to guide subsequent actions. Recent scientific-agent systems incorporate reflection or validation steps to assess these results [Gottweis et al., 2026, Mehandru et al., 2025]. We study the conditions under which LLM reviewers can reliably distinguish correct from incorrect scientific work, varying the evidence available to the reviewer, conversational framing, and judge model. We test this on real computational biology trajectories from BixBench: a Claude Sonnet agent executes analyses in the benchmark data environment, and three judges review the 116 answers carrying deterministic ground truth. Showing the executed notebook raises acceptance of wrong answers from 0.30 to 0.90 for Sonnet and from 0.06 to 1.00 for GPT-5 mini (0.86 and 0.97 when the prose is also shown), while acceptance of correct answers rises by a similar margin, leaving balanced accuracy near chance. The verdicts alone suggest that the notebook does not help, but a score combining each judgment’s verdict and stated confidence retains discrimination. Its AUROC reaches 0.73–0.77 for Sonnet and GPT-5.6 Terra when the notebook is present, compared with 0.48–0.51 when only the answer is shown. Most strikingly, GPT-5 mini accepts all 116 answers when shown the notebook without the agent’s prose, yet its confidence ranks correct above incorrect answers 66% of the time. In addition, we find no evidence that the producing model reviews its own work more leniently. The code is released at https://github.com/thunderingluck/bio-auditor.