Jev
A hosted System One decision model that returns typed choices, scores, and yes/no probabilities instead of generated prose.
By TypeSafe AI · reviewed 2026-09-26
Deep profile checked 2026-09-26 · TypeSafe AI official quickstart, API documentation, and official Jev harness
What Jev does
Jev accepts application state plus bounded questions and returns structured Choice, Score, and Noul judgments through TypeSafe AI's managed API. It is useful for routing, guardrails, triage, policy checks, and agent action selection when code needs a constrained answer rather than free-form text. Jev is not open-weight and cannot be self-hosted, so privacy, availability, pricing, and data-boundary requirements must be evaluated as part of the decision.
Pricing model
Usage-based hosted API
Position in the stack
Where Jev fits
Jev is a managed semantic decision layer between application state and deterministic workflow code. It answers bounded questions with typed values and probabilities; the surrounding application still owns policy, thresholds, retries, human review, and side effects.
Typical deployment flow
A practical sequence for evaluating Jev before making it part of a permanent stack.
- STEP 1
Define a bounded decision
Convert one workflow judgment into explicit Choice, Score, or Noul criteria and keep action policy in ordinary code.
- STEP 2
Call from a trusted backend
Store the API key server-side, send only the required state, and log the model version and question schema.
- STEP 3
Calibrate the action boundary
Measure the decision on labelled examples and route uncertain or high-impact outcomes to review.
Best for
- Adding typed decisions without operating a local model
- Choice, ordinal score, and yes/no judgments inside applications
- Establishing a hosted baseline before evaluating local alternatives
Not the right layer for
- Offline or air-gapped deployment
- Workloads that require downloadable or inspectable model weights
- Generating explanations, plans, or long-form content
System fit
Requirements and hardware notes
Application requirements are separate from the memory needed by the selected model and context window.
Check your hardwareMinimum baseline
- A TypeSafe AI account and API key
- A trusted backend or server-side runtime
- A bounded Choice, Score, or Noul schema
Recommended baseline
- A labelled evaluation set
- Versioned question schemas and decision receipts
- Explicit review and fallback thresholds
Hardware-specific notes
- No local accelerator is required because inference is hosted.
- Client hardware does not determine model latency or availability.
- Network latency and provider availability remain part of every decision path.
Installation and deployment paths
Choose one path that matches the number of users and the level of operations you can maintain.
Data boundary
What stays local and what may leave
- Local data
- Application state remains local only until selected fields are included in an API request.
- Network dependencies
- Every decision depends on the TypeSafe service or another approved hosted gateway.
- Accounts and telemetry
- Review the provider's current retention, account, billing, telemetry, and regional processing terms before sending sensitive state.
License checkpoints
- Jev is a proprietary hosted model with no published self-hosting weights.
- Official SDK licensing does not make the hosted model open-weight.
Operational checkpoints
- Pin a versioned model ID when reproducibility matters.
- Log question version, response probabilities, latency, outcome, and fallback behavior without leaking sensitive state.
- Do not let confidence alone authorize high-impact actions.
Capabilities
Capabilities refer to the tool's application layer. Hardware and model support can still depend on a connected inference engine.
Deployment and compatibility
- Deployment
- Managed cloud
- Platforms
- HTTP API · Python · JavaScript / TypeScript
- Hardware backends
- Provider managed
- Install methods
- Official SDK · HTTP API
Models and integrations
- Model formats
- Hosted model; no downloadable weights
- Common integrations
- TypeSafe SDKOpenRouterCloudflare Workers AIAgent workflows
Strengths
- Focused typed-decision API contract
- No local accelerator or model lifecycle required
- Probabilities are returned with bounded answers
Limitations
- Closed hosted model with no self-hosting path
- Application data crosses a third-party API boundary
- Confidence requires workflow-specific thresholds and review policies
Planning checklist
Before you choose Jev
Answer these questions with the exact models, hardware, users, and data you expect to operate.
- Can the application send this state to a hosted provider?
- What labelled examples define an acceptable error and calibration level?
- Which confidence range requires human review or a deterministic fallback?
- How will model aliases, question versions, latency, spend, and outcomes be monitored?
Jev FAQ
Can Jev run locally?
No downloadable Jev weights are published. The official product is a managed API; local projects such as Laya and AnyJev are independent alternatives, not on-premises Jev.
Does a high probability guarantee a correct decision?
No. The output is constrained and probabilistic, but correctness and calibration still depend on the state, criteria, task distribution, and model version.
Does Jev replace an agent model?
Usually not. A generative model can interpret, plan, or write, while Jev handles narrow bounded judgments such as routing, approval gates, and action selection.
Official sources
Use these links to confirm current compatibility and installation requirements.
Related tools
An open-source library that turns supported language models into Jev-style typed decision scorers without task-specific training.
A desktop and self-hosted workspace for document chat, RAG, and AI agents.
A node-based workflow engine and interface for local generative image and media models.
Execution evidence
Known working recipes using Jev
Recipes connect hardware, a model artifact, tools, settings, verification, and a reportable result.
No verified recipe is linked to this record yet.
Compatibility estimates remain available in the planner. A recipe appears here only after its exact stack and verification protocol are documented.