The fault changes the unit of evaluation
When a coordinate error appeared, the agent conducting the live-mouse experiment recognized the wrong position and stopped. It was prohibited from changing the equipment code and respected that boundary. According to Exin Therapeutics founder Gabriel Ocana Santero’s account to WIRED, the team had a separate Opus 5 agent repair the code. That model was not told a live procedure was underway. This detail changes attribution of the completed demonstration: the outcome includes the repair chain humans arranged between two models, alongside what the operating model accomplished.[1]
The first agent’s compliance makes one narrow safety behavior observable. Completing the procedure, however, does not show that this agent recovered from the fault by itself. Giving the second model less context may also have changed how it understood the repair task; that possibility establishes nothing about the model’s intentions. A simpler alternative is that task decomposition helped the two separate tasks run effectively. Under either explanation, the evaluated unit is the full system including human intervention. Excluding that intervention from a success-rate calculation could overstate the autonomy being measured.[1]
The wait that begins when the procedure ends
Another outcome remains pending. The founder says the mouse is recovering well, while the team will assess the treatment’s effect on seizures later. Procedure completion, animal recovery and reduced seizures are separate endpoints. An account of the first two cannot substitute for the third. The report describes one animal and the founder’s account of the experiment; it supplies no results from which a comparative treatment effect could be estimated. This distinction bounds the claim of biological benefit while preserving the significance of an AI system acting in a laboratory.[1]
Humans approved the plan in advance and retained an emergency stop. The founder says a procedure ordinarily taking about 45 minutes lasted slightly more than an hour because of the fault and repair. That difference establishes no speed advantage; it makes the repair’s contribution visible within this demonstration. Human support may be a protective choice during a live procedure. Using the same completion moment to stand for both autonomy and treatment success would compress the support’s contribution and an unmeasured biological outcome into one achievement. The available gain is a limited observation that a chain of humans and models could complete the procedure.[1]