Jev API
Jev is TypeSafe AI's System One model family for typed, probability-based decisions: classification with Choice, rubric evaluation with Score, and yes/no judgments with Noul. Compare Jev 1.13 and 1.14.0, including version-specific experimental capabilities.
Jev 1.14.0 API
typesafe/jev-1.14.0
Jev 1.14.0 improves typed decisions and multilingual support, adding experimental Box outputs alongside opt-in image and longer-context features. Platform availability and experimental feature support must be checked separately.
Jev API Background
Overview
Jev is designed for applications that need a decision they can act on directly. A request supplies a state and a set of typed questions; the response contains structured answers rather than a generated explanation. The state can describe a customer message, document, product, incident or other application record. The application defines the available outcomes and decides how the results affect its workflow. Jev complements a text-generating model when a system needs both written responses and explicit decisions.
Development History
This family page brings together the platform's jev-1.13 and jev-1.14.0 entries. TypeSafe released Jev 1.14.0 on October 14, 2026, describing incremental improvements across tasks and a particular focus on multilingual quality. Its original API uses the version IDs jev-1.13.0 and jev-1.14.0; model identifiers on this platform follow the selected version card. An upstream release does not imply that every provider already exposes it.
Key Innovations
- Typed decision primitives let software consume selected labels, numerical scores and probabilities without extracting them from prose.
- Independent questions can share one state. Applications can collect several judgments and combine them using explicit business logic.
- Experimental image inputs are available upstream for both versions, while the experimental Box question type is specific to Jev 1.14.0.
Jev API Technical Specifications
Architecture
TypeSafe calls Jev a System One model: a model built around narrow judgments supplied as typed questions. Each question is evaluated independently against the same state. It is intended for well-scoped decisions; broad tasks requiring multiple reasoning steps should be decomposed into separate judgments and combined in application code. Public documentation does not provide a parameter count here, so none is assumed.
Capabilities
- Choice selects from a developer-defined set of named options, returning the selected option, the probability distribution and a confidence value. It fits categories without an inherent order, such as support departments or product classes.
- Score evaluates an ordered rubric. It returns a numerical score that can lie between rubric levels, probabilities over those levels and confidence. Write descriptions that distinguish the meaning of each level.
- Noul evaluates a yes/no statement and returns a probability between zero and one. Use it for individual predicates rather than treating a probability as an unconditional Boolean result.
- Choice and Score confidence summarizes concentration of the output distribution; it is distinct from a guarantee of correctness. Thresholds and escalation policies belong to the application.
- The upstream default context limits are 64k tokens for the whole request and 32k for the state plus any single question. Questions must fit both limits. These are upstream reference limits; check the chosen platform entry for its exposed configuration.
- Upstream experimental long context requires TypeSafe-Danger-Zone-Long-Context: true and raises both context limits to 262k. Image input requires TypeSafe-Danger-Zone-Images: true. Header passthrough and experimental support must be verified with the actual provider.
Limitations
- Jev is not a general prose-generation interface. Use a separate generative model when the product needs a written answer, open-ended composition or a detailed explanation.
- Image inputs, Box outputs and extended context are opt-in experimental upstream features, not an assurance of production readiness or availability through this platform.
- Longer context can reduce accuracy and increase latency. Supply relevant evidence and avoid making a single question carry several unrelated judgments.
- Do not infer support for audio or video from experimental image support. Convert unsupported media to appropriate text or structured evidence before evaluation.
Jev API Performance
Strengths
- Explicit output types suit workflows where software must branch, rank or route using the result.
- Choice can compare known categories; Score can retain an ordered evaluation instead of forcing every judgment into a binary decision.
- Confidence-aware handling can distinguish automatic decisions from cases that need additional evidence or human review.
- The shared-state, multiple-question pattern makes it possible to evaluate separate aspects of one record while keeping the final policy visible in code.
Real-world Effectiveness
Use a representative evaluation set before adopting a version or changing a threshold. Include ambiguous records, missing information, minority categories and the languages used by your customers. Measure task accuracy, wrong automatic actions, review volume, latency and cost under the actual provider. TypeSafe's release announcement describes broad improvements in 1.14.0, but does not establish a universal improvement percentage for every workload. Retain a pinned older version until the replacement passes your own acceptance criteria.
Jev API When to Use
Scenarios
- Support triage: choose the owning team from a fixed list and evaluate urgency separately with an ordered rubric.
- Catalog and document classification: map records to known labels while preserving uncertain cases for review.
- Content evaluation: assess individual policy conditions or quality dimensions, then combine them using explicit application rules.
- Agent routing: choose between known tools or handlers, with ordinary permission checks applied before an action executes.
- Retrieval pipelines: evaluate candidate relevance or verify a narrowly stated claim against supplied evidence.
Best Practices
- Write one clearly scoped judgment per question. Put the evidence in state and define mutually distinguishable Choice options or descriptive Score levels.
- Use confidence and probabilities as inputs to a tested decision policy. Include a review or abstention path where uncertain decisions should not execute automatically.
- Pin a specific model version during evaluation and deployment. Re-test thresholds when moving from 1.13 to 1.14.0; do not assume identical output distributions.
- Verify the selected version's platform status and price before integration. Official upstream pricing and endpoints are not a substitute for this platform's version-specific API documentation.
- Keep image and extended-context experiments separate from the initial production text workflow until the chosen provider and your evaluations support them.
- Reference documentation: https://docs.typesafe.ai/introduction ; https://docs.typesafe.ai/models ; https://docs.typesafe.ai/primitives/choice ; https://docs.typesafe.ai/primitives/score ; https://docs.typesafe.ai/confidence ; https://docs.typesafe.ai/images ; https://typesafe.ai/blog/jev-1-14 .