Vercel — How to classify, route, and score with Jev and AI SDK (2026-09-19)

Ben Sabic, Vercel Knowledge Base, 19 September 2026. Arrived by Telegram the same day. The second source in this spoke on typesafe-ai's Jev, four days after the launch post (system-one-models-and-jev, 2026-09-15) — and it changes that page's status in three places.

T2, not T3. It is first-party for the integration (Vercel's SDK, Vercel's gateway) and second-party for the model: Vercel resells Jev and repeats TypeSafe's price, so it is not an independent check on any vendor claim. What lifts it above the launch post is that it documents a shipping API surface with limits and error behaviour, and that in two places it states the limitation the vendor's own post did not.

What changed since the launch post

That first row answers one of the three open items left on system-one-models-and-jev: availability after early access. The price held across the transition, which is worth recording because a free output tier reads like a launch position and has now survived one.

The API, in the shape that matters to this spoke

Three question types, each returning a typed answer with a probability, evaluated independently against one shared state — so adding a question does not change the others' answers, and latency barely moves:

• choice — one option from a named map; returns choice plus a full probabilities distribution.
• score — graded against an ordered rubric of 2–10 levels; returns a probability-weighted mean (the guide's worked example scores 2.86 across four levels) plus the distribution.
• boolean — returns probability that a statement is true.

The typing claim is concrete rather than rhetorical: a choice question's answer is inferred as a union of the caller's own option keys, so choice === 'tech' against options billing | technical | account | other fails at compile time.

A second statistic nobody had reported. result.providerMetadata.typesafe.confidence returns a per-question number from 0 (probability spread evenly) to 1 (all on one option), summarising how concentrated the distribution is. It is not the selected option's probability, and it is not returned for boolean answers. The guide's routing example gates on both — confidence ≥0.6 and selected probability ≥0.7 — and sends everything below either floor to a human.

The two sentences that qualify the vendor's pitch

system-one-models-and-jev records TypeSafe's claim that Jev "can't hallucinate." This spoke read that as true about a narrow object: the model cannot emit prose, so it cannot fabricate text. The reseller's guide says the rest of it out loud:

"Your schema constrains those answers, but doesn't guarantee they're correct."

And on the calibration promise, which the launch post left as an assertion with no curve behind it:

"Calibration describes how predicted probabilities match observed outcomes across many examples; it doesn't guarantee an individual answer is correct."

Neither is a retraction — both are compatible with everything TypeSafe published — but they are the distinction the launch framing blurred, made by the company selling access. A typed decision model trades one failure mode (invented text) for another (a confidently wrong typed value), and this is the first source in the corpus to say so.

The guide also treats the probability distribution as optional across providers — other evaluation models "may omit them" — which quietly makes experimental_evaluate a category interface rather than a Jev driver. Whether a second System One model appears to fill it is the thing to watch.

Engineering detail worth keeping

The SDK ships Experimental_EvaluationMockModelV4 in ai/test, so the threshold logic — the part that is application code, not model behaviour — is unit-testable with no network call. That is a small thing that says something about the category: when a model returns a typed value instead of prose, the interesting bugs move out of the prompt and into ordinary branching, where ordinary tests reach them.

The guide is careful about one more seam: its requestsRefund boolean "records intent, not approval." Classifying a request is not deciding it.

Related

system-one-models-and-jev · typesafe-ai · ai-gateway · llm-api-pricing · inception-labs · vercel · synthesis

Cross-spoke note

The AI SDK itself is not paged, on purpose. Vercel's AI SDK is a framework for building LLM applications, which is ../agentic-tooling-wiki's subject, not this spoke's — this page keeps only what bears on model access: the model ID, the gateway, the limits, the price and the answer shape. ../cloud-wiki owns vercel as a hosting provider. If the SDK earns a Thing page, it belongs in agentic-tooling-wiki with a link back here, not a second copy of this.