TypeSafe AI Releases Probabilistic Decision Model 'Jev' on AI Gateway

TypeSafe AI has released Jev, a probabilistic decision model for software, via AI Gateway. Jev takes state inputs and directly returns typed Choice, Score, and Boolean answers along with probabilities.

Unlike existing LLMs that generate text token by token for applications to parse and validate, Jev evaluates all declared questions in parallel. This eliminates unnecessary text generation and streamlines processes by automating clear cases and routing uncertain cases for review.

Performance and Use Cases

TypeSafe reported that Jev was up to 193.6x faster and 444.6x cheaper than existing LLMs in workflow evaluations. Key use cases include:

• Selecting the next tool or sub-agent in agent loops
• Deciding whether to continue, retry, ask the user, or stop
• Scoring urgency or risk before execution
• Validating model outputs and applying guardrails

Developer Integration

AI SDK 7 exposes Jev via the experimental evaluate API. Supported in AI SDK 7.0.105 and above, it is called by specifying model, state, and questions. Specifying typesafe-ai/jev as the model yields results where question IDs and Choice keys are preserved.

Results include separate confidence scores for Choice and Score via result.providerMetadata.typesafe.confidence. Additionally, Zero Data Retention and No Training options can be enabled per request, and evaluation calls are logged in logs and custom reporting and reflected in budgets.