TypeSafe's Jev model returns judgments and probabilities that application code can turn into decisions. In OpenRouter'smarketplace tutorial, ordinary code first checks requirements such as photo counts, then asks Jev to interpret listing text. A separate policy determines whether to publish, reject, or request human review. Saving the model's answers lets developers test different policy thresholds without paying for another model call. Those thresholds need human-labeled examples: Jev's confidence describes how concentrated its answer probabilities are, which alone does not establish accuracy on a particular marketplace.
