TypeSafe launches Jev for fast, structured AI decisions
Jev takes context and questions with predefined answer formats, returning choices, ratings, and probabilities that an application can use directly. It is designed for bounded decisions inside software rather than conversation or free-form writing.
Summary
TypeSafe introduced Jev on September 15 as an AI model specialised in decisions that software can use directly. Rather than generating free-form text, it receives information about a situation and questions with defined answer formats. It returns a choice among options, a rating on a scale, or the probability that the answer to a question is “yes”.
The product is aimed at applications that need to make many small judgments: routing a request, classifying a document, assessing urgency, or deciding when a person should review a case. TypeSafe calls this category “System One”, referring to fast, bounded decisions.
In practice
Consider a customer message saying they were charged twice. An application sends Jev the message and account details. It might ask for the main request type, choosing among “refund”, “information”, and “other”; whether the customer explicitly asked for their money back; and how urgent the case is on a scale defined by the team.
Jev answers all three questions about the same context without drafting a reply to the customer. For a choice, it returns the selected option, a probability for each option, and a value summarising its confidence. For a yes-or-no question, it returns a number between zero and one. The application’s code combines those answers with its own rules to route the case, seek confirmation, or send it to a person.
According to TypeSafe, the model was trained to produce calibrated probabilities: across many cases, answers assigned higher probabilities should more often prove correct. The company says it uses its own architecture and a training method it calls “reinforcement learning for calibrated decisions”. Questions about the same context are evaluated in parallel, without generating text one word at a time.
What we still don't know
The format prevents Jev from returning an option outside those defined by the application. That removes format errors, but does not guarantee that its chosen option is right. Missing context, poorly written criteria, or a task requiring extended reasoning can still lead to bad decisions. TypeSafe’s documentation recommends splitting complex problems into smaller questions and testing confidence thresholds on data from the intended use case.
TypeSafe says Jev can be much faster and cheaper than language models on these tasks. Its published comparisons are company-run: in the workflow tests, reference answers are based on the average decisions of two advanced models, rather than independently established ground truth for every case. TypeSafe also acknowledges that some examples and test choices may favour Jev. Its performance remains to be assessed in external applications with their own rules and consequences.
Why it matters
- Jev illustrates a way to use AI inside a program: the model makes bounded judgments while code retains control of the actions.
- Probabilities and confidence can help determine which cases proceed automatically and which need review, provided their reliability is measured on the actual task.
- The difference between a valid answer and a correct decision remains essential: constraining the output removes one kind of error, but does not replace evaluation or oversight.