Jev
Jev is TypeSafe AI's first public model, which they describe as a "System One" model for automation: reached as typesafe-ai/jev through the Vercel AI Gateway's decisions endpoint, it does one narrow thing, answering typed questions about an application-state string. A question is a choice from an enumerated set, a predicate (a calibrated yes or no), or a score against ordered criteria, and every answer comes back with a probability and the full distribution rather than prose we would have to parse. Input runs about four cents per million tokens with output free, so a decision is effectively free compared to a frontier call.
We use it in Replicator at decision points that are genuinely boxed and always have a fallback path: assessing what an inbound attachment is for, picking which organization a document belongs to, and judging reuse-versus-build in the PDF widget pipeline. The rule we deploy under is deliberately narrow. Jev is for enumerable choices behind a fallback, never destructive operations, and never a replacement for a guardrail we established with a measurement.
That last clause came from running the experiment. Barkup Bench Study AU put Jev against fable-5.1 on 102 frozen PDF regions, both engines fresh on byte-identical inputs. Action agreement came in at 0.676 against a pre-registered 0.85 gate, so the swap was declined and fable stayed. The useful part was the texture: 31 of 33 disagreements ran one direction, Jev reusing at 45.1% where fable sat at 16.7%, with its extra reuses carrying median probability 0.71. That is a systematically looser policy rather than noise, and it was 49 times faster and roughly 500 times cheaper while being the wrong judge for that stage.
How We Use It
- Enumerable decision points where a probability matters more than an explanation
- Calibrated yes or no predicates over application state, behind a fallback path
- Cheap, fast classification and routing that would be wasteful as a frontier model call
Related reading: Study AU: The match-stage engine A/B, where Jev proved to be a looser policy rather than a drop-in judge
AI / LLM
Why this matters: Controversial, misunderstood, and overhyped though they may be, AI and LLMs are transforming industries and society. Used carefully and responsibly, they are still a valuable tool for innovation and progress. We design and deploy LLM-powered features with a focus on safety, performance, and maintainability, leveraging managed model providers and an SDK-first approach for rapid iteration. Preferred defaults: Vercel AI SDK + AI Gateway for multi-model orchestration; Anthropic and OpenAI as primary model providers.