Decision Models
How can a model judge a question it has never been trained to name?
Start with the first report → See all reports ↓Overview
A support ticket can raise several judgments at once: which team should receive it, whether it is urgent and whether a refund is requested. A decision service can make these answers convenient to request, but a convenient interface leaves open how the model uses a new question, compares its options and produces meaningful probabilities.
The series starts with Jev as a concrete service and asks what its public evidence establishes. It broadens the comparison to classifiers, entailment models, constrained generators and direct scorers. The final article goes inside the training problem: how to connect supplied questions and choices to the evidence, test transfer to new rules and evaluate probabilities separately from answer accuracy.
Follow the decision from its interface into its training
Request a judgment
Separate the convenience of a service from model-level quality.
Compare the methods
Trace how each approach represents the supplied options.
Teach the relation
Learn to use new questions, rules and probability targets.
The comparison expands beyond Jev; the training question concerns request-conditioned judgment across methods.
Reading order
A recommended path through the enquiry. Each report explains its part; continuation links preserve the original branches.
- Separate capability from serviceIs Jev a better decision model—or a better decision service?
Introduces the decision interface and asks which operational benefits and quality claims survive fair comparisons.
Start here - Open the comparisonWhat exists outside Jev—and how close has it come?
Shows how the same bounded output can come from different methods, with different ways of incorporating a new option.
- Train for new questionsTraining a Language Model to Judge New Questions
Explains the supervision and transfer tests needed to make supplied meanings matter, then treats probability quality as a separate target.