From text decisions to visual understanding: explore TypeSafe Jev’s structured judgments, experimental image inputs and Box localization. Follow Jev 1.14 updates, official news and real-world workflows.
“I was charged twice.
Can someone help?”
{
"choice": "billing",
"confidence": 0.98
}Example values for illustration. Not a live inference.
Published end-to-end response time
Three decision primitives + experimental Box
Deterministic output for direct code consumption
Source: TypeSafe announcement and docs. Not a site benchmark; results vary by workload. Explore the model →
Images can appear in state, question instructions or criteria. Use Choice, Score and Noul for visual judgments, and Box to locate one target.
PNG / JPEG, image asset references and normalized bounding boxes. Learn presence checks, image dimensions and request limits.
OPT-IN / DANGER ZONEUnderstand the experimental boundary →Images and Box are off by default and require an opt-in header per request; Box needs Jev 1.14. TypeSafe does not yet consider them production-ready. No native audio or video input.
Independent coverage based on official announcements and documentation. Articles include sources, dates and scope; case-study figures are not our benchmarks.
Multilingual improvements, experimental features and preview versus latest.
TypeSafe announces its Series ALed by a16z, with plans for developer infrastructure and enterprise features.
Jack & Jill: 88% lower screening costRead the reported cost, latency and quality results, including the limits of projections.
ALL NOTES / GUIDESRead all news and guides →Model introductions, integration notes and multimodal guidance in English and Chinese.
Explore how Jev AI fits task routing, information retrieval and agent checks. The model supplies judgments; your code controls what happens next.
Use Jev’s Choice judgments to classify intents and route incoming requests to the right team, tool or microservice with zero parsing boilerplate.
Published 70–500ms end-to-end response time. Inline decision-making fast enough for user-facing interactions.
Choice selects, Score evaluates and Noul judges conditions. Jev 1.14 adds experimental Box questions to return a bounding box around one image target.
"billing"0.880.94Set probability thresholds for automation. High-confidence decisions execute instantly in code; uncertain edge cases route to human review.
Jev AI explores TypeSafe’s Jev System One model: application state and focused questions go in; choices, scores and probabilities come out. It is built for decisions inside software automation.
Useful for writing, code generation and open-ended reasoning. Applications consume the generated content with their own processing and validation rules.
INPUT → GENERATED CONTENT → APPLICATIONUseful for selecting a path, scoring a dimension and judging a condition. Define the questions and answer space, then let code act on the returned decisions and probabilities.
STATE + QUESTIONS → TYPED DECISIONS → CODEWorking together: A larger model handles complex reasoning and generation, Jev supplies fast judgments, and code and tools own rules and execution. Each part has a focused role in the workflow.
Choice selects, Score evaluates, Noul judges a condition, and experimental Box locates an image target. Local examples illustrate the first three.
Explore Jev AI primitivesPick from a defined set
→ "billing"Evaluate against a rubric
→ 0.82Evaluate a truth statement
→ 0.94Image localization (experimental)
→ boxJev is TypeSafe’s System One model: text and experimental image inputs become structured judgments your code can use.
Choice selects an option, Score evaluates a rubric and Noul returns the probability of yes. Bounded outputs are easy to compose.
Ask several independent questions over the same state in parallel. They cannot see each other’s answers; code keeps execution order and deterministic rules.
TypeSafe calls its training approach RLCD: reinforcement learning for calibrated decisions. Use probabilities to guide action or review; typed output and high confidence do not guarantee a correct answer.
Alongside text, Jev now offers experimental PNG / JPEG inputs. Audio and video still need transcription or still-frame extraction; this site’s examples remain local illustrations. Jev AI model guide →
Browse browser agents, model routing, context compaction, MCP tools and independent implementations.
JEV AI / AGENT SKILLBuild Jev AI workflows with the official skill →Use the official TypeSafe skill to give Codex, Claude Code and other agents guidance on primitives, question design and uncertainty.
Jev AI is an independent guide and project resource for TypeSafe’s Jev model. Jev evaluates state and focused questions, returning choices, scores and probabilities. Experimental image inputs and Box questions add visual judgments and localization.
Jev 1.13 and 1.14 experimentally accept PNG and JPEG images; Box localization needs 1.14. Each request must enable TypeSafe-Danger-Zone-Images. Images and Box are not production-ready by TypeSafe’s standards. Audio and video are not native inputs, and this site does not run image inference.
The Jev model covered here focuses on structured judgments rather than free-form text. Use it to select candidates, score a dimension or judge a condition. Writing replies, articles or code still requires a generative model.
Use Choice to pick one candidate, Score to evaluate a described scale, and Noul for the probability that a condition holds. Independent questions can evaluate the same state together.
The current examples run locally for learning. They do not call the model, submit your data or demonstrate real accuracy. They illustrate the shape of each primitive’s output.
Structured output does not guarantee a correct answer. Choice and Score confidence describes distribution concentration, not accuracy. Validate on representative tasks and keep a review path.
Explore the local decision examples, find a browser tool, coding agent or search implementation in the project directory, then plan an integration with the getting-started guide. A site account does not grant model API access.
Find a practical starting point in the community directory