Vercel Labs’ form router uses Jev to assign lead, contact and issue submissions to registered teams. Its code makes the fallback policy explicit and keeps the first model’s statistics attached to its original answer.
Route with an application-owned threshold
- The server validates form fields and builds a Choice question from its own destination registry.
- It accepts Jev’s registered choice when TypeSafe confidence is valid and at least 0.95. The comparison uses the unrounded value.
- Low or missing confidence, an invalid result or an evaluation error triggers an independent GPT-6 Luna Fast decision using the same submission and criteria.
- The fallback sees neither Jev’s answer nor its statistics. Its valid choice becomes final, even if the two models disagree; there is no further review loop.
Confidence and the selected option’s probability are separate fields. A typed destination only meets the schema; it does not establish that the routing decision is correct. The 0.95 cutoff is the example’s policy, not a demonstrated accuracy guarantee.
Preserve failure and delivery boundaries
The code bounds both model calls and retries transient failures. If the fallback fails, the submission returns an error and sends no email. Successful routing produces an email preview; actual delivery requires opt-in and configured recipients. A provider’s acceptance does not confirm inbox delivery.
The repository says its tests use mocked model providers and email delivery. It has no built-in authentication, application rate limiter or durable submission history. Submitted fields go to the model gateway, and opted-in emails go to the email provider. Instructions to treat input as untrusted are not proof of prompt-injection resistance.
Verification: inspected the README, routing and submission implementations, and architecture notes. The project was not run, its model accuracy was not measured, and this was not a security audit.