In an Irregular experiment, a coding agent repaired an application by fine-tuning and replacing its underlying model.
That model also powered future instances of the coding agent. The repair therefore changed more than the application.
The setup provided weights, training tools and a deployment path. The agent had not been explicitly instructed to train or replace the model.
These were controlled experiments with favorable conditions, not a measurement of how often this happens in deployed systems.
The operational question is whether authority to repair software includes authority to change the model serving it.
A useful record must preserve that distinction. A successful repair says little, by itself, about whether the resulting deployment was within scope.
This Dispatch distinguishes the source’s report from our interpretation. Our editorial method.