Genoria AI releases Operon-S1, splitting judgment out of the agent loop: flipping option order moves about 6% of answers, down from about 37%

In multi-step research tasks, the key judgments a biomedical AI agent makes are now checked separately instead of being folded into execution. The same gene-expression data can yield a different gene list under a different filtering threshold, and code that runs successfully does not mean the standard used meets the study's requirements. On a 49-item scorable bioinformatics test, the original agent answered 12 items wrong, and for 5 of them the correct value already appeared in the run output but the agent chose an answer under a different convention.

Judgments are turned into questions with explicit options, answered by a replaceable judgment model, and then fed back as reminders and review requests. Execution stays with ProtoPilot, which plans, writes code, runs analyses and assembles answers; the judgment system does not take over tools or code. Judgment happens at 13 decision-point types at task start, during the run and before submission, covering analysis type, cause of an error and sufficiency of evidence, and records keep the judgment made, whether the agent adopted it, and the final task outcome separately.

Judgment capability can be trained on its own, but a new version must clear a verification threshold before it becomes the starting point for the next round. Training labels come from task construction, verification programs or code execution, and a new version is compared against the previous version and rule-based methods, with a check that the model's confidence matches its actual performance. On 750 multiple-choice questions not used in that training round, flipping only the option order changed about 6% of answers, down from about 37%, and one error-classification judgment ran on CPU with a median time of about 0.6 seconds.

The demonstrated scope is computational: 7 test groups and 1,417 tasks spanning sequence reasoning, data analysis and protocol troubleshooting, plus one probe design and quality-control analysis; in the probe case, historical capture-sequencing data predating Operon-S1 serves only as an experimental reference, and the protocol-troubleshooting test analyzes protocols and failure descriptions rather than operating instruments.

The study has not shown that it can raise efficiency and hold scores at the same time across most tests: the untrained initial LAYA version cut SeqQA cost by 5.8% and execution steps on the exploratory protocolqa2 test by 10.3% with no detectable score change in either group, yet the same version lost about 1.3 percentage points on seqqa2; the stability gain also does not represent accuracy on the whole research task.