TypeSafe Jev: The Decision Model That Answers With Choices, Not Paragraphs

Most AI systems in production are asked to do something deceptively simple: read a record and decide what happens next. Should this ticket go to billing or to infrastructure? Is this lead worth a sales call? Does this request violate policy? A general-purpose language model can produce a fluent answer to any of those questions, but fluency is not the same as a usable output. TypeSafe Jev was built around that gap — it is a hosted decision model that returns typed answers instead of generated prose, so the result can be routed, scored, or blocked by ordinary application code without a parsing layer in between.

What TypeSafe Jev Does: Choices, Scores, and Noul Probabilities

Jev takes two inputs: application state, such as a support ticket, a JSON record, or a form submission, and a set of typed questions about that state. What comes back is not a sentence. It is a decision expressed in one of three output types, each designed to map directly onto a branch in a program.

The first type is a choice — a single selected option from a set of up to 255 candidates. Because the model returns probabilities alongside the selection, callers also receive a confidence signal rather than a bare label. That matters in production: a 0.97 confidence routing decision can be automated, while a 0.41 can be queued for human review. The second type is a score, expressed as a level on an ordered scale of 2 to 10. Scores are the natural fit for urgency, risk, or lead quality, where the useful information is relative position rather than a category. The third type is a yes/no probability, which TypeSafe refers to as the “noul” type — a calibrated likelihood that a statement holds, useful for policy checks and guardrails.

TypeSafe describes Jev as a “System One” model, a framing that positions it as fast, instinctive pattern recognition rather than extended reasoning. That distinction shapes where it belongs in an architecture. Jev is not trying to write explanations or hold a conversation; it is trying to make a bounded judgment about structured input quickly and consistently, then hand that judgment back to software that already knows what to do with it. The current release is Jev 1.13.

The operational profile is deliberately narrow. Input is text and JSON only — there is no image support. The context window is 64k tokens, and throughput is documented at 80 requests per second per key. Official SDKs exist for Python and JavaScript, which keeps integration close to the languages most application and data teams already use for service code. TypeSafe also states that customer requests are not used for training, a detail that tends to matter during procurement reviews in regulated or privacy-sensitive organizations.

Where Jev Fits: Routing, Scoring, and Guardrails Before an Agent Acts

Because the outputs are typed, Jev’s most natural home is the layer that sits between raw application state and an action — including actions taken by an AI agent. A support platform can pass a ticket body and metadata into Jev and ask for a routing choice across its queue taxonomy, then act on the returned option and confidence. The same ticket can carry a parallel urgency score on the 2-to-10 scale, giving an on-call team a ranked view without anyone reading the text first.

Intent classification follows the same shape. Instead of prompting a generative model and hoping the label matches an enum, the candidate intents are declared up front, and the answer is guaranteed to be one of them. Risk scoring and lead quality fit the score type for the same reason: the output is a level, so thresholds and tiers can be defined before deployment rather than discovered in output samples.

Policy checks and guardrails are where the typed approach earns its keep. A yes/no probability on a statement such as whether a request touches a restricted operation gives an agent framework a cheap gate to consult before it executes a tool call. If the probability crosses a configured threshold, the action is held. This is the pattern that makes Jev useful as a pre-action check rather than a replacement for the reasoning model itself — the agent still does the work, but the decision about whether that work is allowed is made by a component whose output never needs to be interpreted.

The constraints shape the design too. With text and JSON input only, Jev is not a fit for screenshot triage, video frame analysis, or document-layout tasks. The 64k token context is generous for tickets and records but not for very large concatenated corpora. And at 80 requests per second per key, high-volume pipelines need to think about key distribution and batching strategy before launch. None of these are flaws so much as boundaries that determine whether Jev is the right component for a given job.

Pricing, Benchmarks, and How to Compare Jev With Alternatives

TypeSafe’s documented pricing, as of October 2026, is $0.042 per 1M input tokens with output free. Because the output is a typed decision rather than generated text, output-side cost disappears entirely, which makes forecasting unusually simple: estimate input tokens, multiply, and the decision layer’s cost is known. There is no listed free tier, so evaluation typically starts on a paid key.

Cost is only one axis. The alternative most often measured against Jev is DecisionNode from Bynn Intelligence, Inc., which also returns typed answers with probabilities but extends the input side to images and video frames, adds number, point, and box answer types, and offers batch discounts, WebSocket streaming sessions, and a free tier. Preliminary benchmarks published in October 2026 put the two close on general text performance, with DecisionNode MoE at 0.816 against Jev’s 0.803 across a mean of 15 text suites. The split widens in both directions elsewhere: Jev scored 0.600 on prompt injection against 0.764 for DecisionNode, while Jev led on knowledge questions with an MMLU-Pro score of 0.820 against 0.698.

Those numbers argue for a specific evaluation method rather than a winner. A team whose workload is knowledge-heavy, text-only, and price-sensitive will read the MMLU-Pro gap as significant. A team deploying an agent that acts on user-controlled input will read the prompt injection gap differently, because a guardrail that can be talked around is not a guardrail. The reliable path is to build a small labelled set from your own traffic, run both models against it with identical questions and answer types, and compare accuracy at the confidence thresholds you would actually ship. A detailed breakdown of Jev’s capabilities, costs, and alternatives — maintained independently of TypeSafe AI — is available at typesafe jev, including guidance on structuring that kind of evaluation before committing a production workflow to any single decision model.