A biomedical agent can keep the right verdict while citing different evidence: BioAgent-DF restores tool values at their original coordinates to trace the swap
Synopsis
BioAgent-DF swaps reference and candidate values saved at the same scientific computation site back at their original coordinates, then lets a biomedical agent continue on healthy hardware, to audit whether a numerical fault changed what it reads and cites. Across eight SciFact batches, Qwen3.5-4B shows repeatable citation changes in three capture–claim combinations and Gemma 3 4B IT shows repeatable read and citation changes in four, covering two claims; full restoration recovers the reference evidence sets while verdicts stay unchanged. In hERG cardiac-safety screening, silent numerical faults changed testing decisions while the fixed-budget shortlist kept the same number of compounds labeled as blockers.
Interpretation
A biomedical agent can return the same verdict while reading and citing different evidence, so auditing only the final answer misses changes along the evidence path. The prior condition treated the agent's final verdict as the auditable result, although two executions on one claim can select different documents or cite different sentences. Across eight SciFact batches, Qwen3.5-4B shows repeatable citation changes in three capture–claim combinations and Gemma 3 4B IT shows repeatable read and citation changes in four, and these seven combinations cover only two claims.
Putting the saved reference values back at their original coordinates returns the agent to the reference execution's read and citation path. Candidate and reference values are saved at the same scientific computation site with matching shape, coordinate indices and representation, and the difference is recorded as a residual. Full restoration recovers the reference read and citation sets in every confirmed combination while verdicts remain unchanged; under the explicit-example prompt it matches the reference in all nine Qwen and fifteen Gemma groups.
In hERG cardiac-safety screening, numerical faults change which compounds are assigned to testing, and aggregate measures do not reveal it. The fixed-budget shortlist keeps the same number of labeled blockers before and after the fault, yet the compounds assigned to patch-clamp testing or to advancement change. On 2,009 internally held-out compounds, six of eight Case 2 sources change the triage vector; after the agent had read faulty predictions, replacing the prediction file recovered none of the changed decisions, whereas restoring the execution state saved before that read and continuing with recomputed scores recovered all 39.
Perspective
The demonstrated scope is two task families, SciFact claim verification and hERG compound screening; numerical faults come from finite outputs captured on two classes of faulty GPUs, and agents replay those values on healthy hardware. The hERG recovery experiment uses local commit records and has not been executed in a physical automated laboratory.
The repeatable evidence-path changes cover only two claims and two 4B models, and come from Case 2 captures selected for upstream retrieval changes; without that selection, candidate values cause no more discordance than a healthy repeat. The R1+M margin threshold in hERG is taken from the largest residual in retrospective sources, making it an empirical recovery rule; certified deployment needs valid per-score error bounds or a fallback to trusted full recomputation.