Onboarding a team to Jev System One
Team onboarding for Jev is a people-and-repo problem: one production door, one questions module, and a shared story about what the model cannot do. TypeSafe’s product is the evaluation API. Your team still owns control flow. Jev (TypeSafe System One) fills typed slots; it does not replace your IAM, your review queue, or your incident process.
This page is for leads who already heard of Jev and need to bring engineers, product, and reviewers onto the same contract. jev.pro is an independent unofficial field guide — not affiliated with TypeSafe AI. We do not sell, issue, or proxy API keys. Confirm primitives and limits on docs.typesafe.ai. For a reading order that separates official docs from this guide, use accessing the unofficial Jev documentation guide. Public pages here are English only.
Steps
- Give engineers What is Jev plus official primitives — not a rival handbook clone. Cover ≠ clone. Pair that with the learning path.
- Stand up a workspace: vault secret, smoke Noul, pinned versioned id. Keys come from console.typesafe.ai after access. Official HTTP uses
Authorization: Bearer; SDKs readTYPESAFE_API_KEY. - Split roles: who may change criteria, who may change floors, who may override in the review UI. See permissions and team roles. Reviewers do not get the production key.
- Teach the three primitives and the line “Noul has no confidence.” Pair with jaggedness notes (literal wording, weak numeric precision) so nobody treats schema-safe output as a fact oracle.
- Ship a review template that shows questions and probabilities — not fake rationales. Analysts act on policy, not argmax. Do not invent a 193× KPI.
- Pick one production door. If that door is Vercel AI Gateway, auth is a Gateway credential or OIDC and the model id is
typesafe-ai/jev, not a TypeSafe first-party alias.
Pitfalls
- One shared production key in a group chat. Rotate; scope per environment.
- Letting agents emit one HTTP call per question. Install the official agent skill pattern so questions batch on one
POST /v1/systemone. - Onboarding by promising “no hallucinations.” Schema-safe ≠ factually correct. Out-of-schema is a different failure than a wrong-but-typed answer.
- Teaching vendor claims as SLAs. TypeSafe currently lists $0.042/M input (output free), 250k tok/s, and 1200 RPM. Label them vendor claims; they can change.
- Treating a 401 during onboarding as “the team is not provisioned.” 401 is a bad or missing service key. Do not retry 401.
Shared artifacts the team should freeze
Freeze a questions module, a pin (jev-1.13.0 or current versioned id), reason codes, and a fail-closed default for irreversible acts. Product can draft criteria; engineering owns the closed option set and the floors. Optional: invite TypeSafe to your Slack — we are not their support desk.
“Pro” on this site means production-minded craft (boundaries, calibration, failure modes), not a paid TypeSafe tier.
See also
FAQ
How long until we are “Pro”?
Pro on this site means production-minded craft — not a paid TypeSafe tier.
Should product write the criteria?
Product can draft; engineering owns the closed set and the floors.
Do we invite TypeSafe to the Slack?
Optional. We are not their support desk.
What this page does not claim
This page does not claim jev.pro can provision seats, issue keys, or join your Slack as TypeSafe. It does not promise hallucination-free answers. It does not freeze vendor price or RPM.
Disclaimer
This is an independent unofficial site and is not affiliated with TypeSafe AI; official documentation is available at https://docs.typesafe.ai.