A smaller, sharper interface
An application often needs a decision, not a conversation. Which workflow should handle this request? Does a condition hold? Which of these options best fits the current state? A paragraph adds another interpretation step between the model and the program. A well-defined output interface makes that boundary easier to inspect.
Options in. Probabilities out.
Manchego takes a state, a question, and a closed set of options. Its decision-model interface returns a probability for each option in one forward pass. The published contract includes choice, yes/no conditions, and ordered scores. This is a different job from open-ended text generation: the software defines the available actions, and the model scores them.
Accuracy is only part of the story
A useful evaluation asks two different questions: how often is the selected option right, and how trustworthy is the confidence attached to it? A probability is not a promise. Thresholds, abstention, and human review need to be designed around the application and evaluated on representative held-out examples.
Read the boundaries, too
The Manchego model card includes benchmark context, calibration caveats, and limitations. Those details belong next to the result, not in the small print. Our working principle for decision systems is straightforward: define the options, expose the uncertainty, and test the entire decision path.
Source material
This editorial draft draws on the public release documentation below. It does not report a new experiment.
Manchego model cardPrepared for this website on 25 Sep 2026.