Specialized decisions,
frontier reasoning.

The most capable AI system is not always the one with the largest model at every step. In healthcare, where speed, predictability, and restraint matter alongside intelligence, architecture matters just as much as model capability.

Why Jev is getting attention

TypeSafe AI introduced Jev in September 2026 as its first public “System One” model: a model designed to return typed probabilistic decisions rather than generate open-ended strings. The proposition is unusually concrete. Define the available choices, send unstructured state, receive a decision plus a probability distribution and confidence that ordinary software can use.

It is easy to see why that launch generated enthusiasm. Model builders have spent years making language models broader; TypeSafe is asking what happens when a model gives up prose generation and specializes in the smaller decisions that make software move. The company makes ambitious claims about speed and economics. Those claims deserve independent workload-specific testing—which is why we ran our own benchmark rather than simply repeat the launch numbers.

Background: TypeSafe’s Jev announcement and current model documentation. Our run pinned jev-1.13.0; aliases can move when new releases ship.

A different question than “which model is smartest?”

Modern EHR intelligence has at least two kinds of work. Some tasks require broad synthesis: reading a longitudinal record, weighing evidence, or explaining a recommendation. Frontier generative models are remarkable at that work. Other tasks are narrow decisions: identify the kind of request, choose the right destination, and know when confidence is too low to proceed.

We tested whether a model built specifically for structured decisions could improve that second layer. The comparison pinned jev-1.13.0, TypeSafe's decision model, against the general-purpose generative model already serving as our production routing baseline. This was not a clinical-judgment evaluation, and neither model was allowed to execute an action.

The study

We assembled 100 hand-written request patterns spanning ten broad kinds of work that an AI-native clinical system may need to distinguish. Each pattern was expressed in 20 surface forms, creating 2,000 cases. Half were reserved for held-out evaluation. The dataset was balanced by design, used no patient data, and measured only whether the system selected the expected destination.

Jev reached 96.1% exact-route accuracy on the held-out set. The general-purpose baseline reached 84.6%. Because the surface variants share semantic roots, we also resampled by the underlying request pattern; the direction of the advantage remained positive under that more conservative analysis.

The useful statistic is not only the 11.5-point accuracy lift. The baseline made 154 incorrect routing decisions; Jev made 39. That is 115 fewer errors, or a 74.7% relative reduction in this held-out sample. Relative error reduction is sensitive to the benchmark's balanced composition, so we report it alongside—not instead of—the raw counts.

Speed and cost are clinical design constraints

A routing decision sits at the beginning of an interaction. Every extra fraction of a second compounds the feeling that an assistant is “thinking” before it can begin useful work. In our observed run, median latency was 168 milliseconds for Jev and 593 milliseconds for the general-purpose baseline. Measured token cost was about $0.055 versus $0.328 per 1,000 decisions.

Those numbers are workload-specific, not universal price claims. But the pattern matters: assigning a narrow decision to a narrow model can reserve frontier reasoning for the places where its breadth creates value.

Confidence is part of the architecture

A healthcare AI system should not turn uncertainty into confidence theater. Jev returns a confidence score with its choice. At a conservative 0.90 threshold, the model accepted 80% of held-out cases and was correct on all accepted cases in this study. The remaining 20% can fall back to a more capable model, request clarification, or wait for human review.

That does not make the complete system perfect. A fallback is only as good as the model receiving it. The lesson is to expose uncertainty and design the escalation path deliberately, rather than hide it inside a single opaque generation.

What happens when the models work together

We also measured a simple cascade: Jev decides first, and requests below the 0.90 confidence gate move to the general-purpose model. On the held-out set, 20% of requests fell back. The combined path reached 97.5% exact-route accuracy, with a 175 millisecond median and an estimated cost of $0.120 per 1,000 decisions.

The cascade's 95th-percentile latency rose to 835 milliseconds because the slowest requests paid for two sequential model calls. That tradeoff is instructive: confidence gating improved overall accuracy and average economics, but it did not improve every latency statistic. Production architecture is a set of negotiated constraints, not a single leaderboard number.

The future EHR will not be powered by one model. It will be an orchestrated system that knows when to decide quickly, when to reason deeply, and when to stop.

What this means for an AI-native EHR

We are building Wavera around a simple principle: use frontier intelligence where medicine needs context and synthesis, and use specialized models where software needs a fast, typed decision. Keep deterministic controls around consequential actions. Make uncertainty visible. Measure the whole system, not just a model in isolation.

This study is an early engineering result, not a clinical validation claim. The next step is careful shadow evaluation on representative, de-identified traffic and continued review of the boundaries between informational and consequential requests.