Access · Last updated 2026-09-21

Promoting Jev decisions between environments

Environment promotion for Jev means moving a decision contract — questions, criteria, thresholds, and a versioned model id — from sandbox to production the same way you promote code. TypeSafe does not document a “promote workflow” HTTP verb. Jev (TypeSafe System One) evaluates the body you send; it does not know which git SHA you intended.

This page is for builders who already heard of Jev and are ready to leave the sandbox. jev.pro is an independent unofficial field guide — not affiliated with TypeSafe AI. We do not sell, issue, or proxy API keys. Confirm aliases on docs.typesafe.ai. Keys come from console.typesafe.ai after access — never from this site.

Steps

1. Fit floors in a sandbox against labeled fixtures. Pin jev-1.13.0 (or the current versioned id), not only jev-latest. See sandbox access.
2. Promote the questions module and the pin in the same change. Log response.model in every environment. If the resolved id drifts from the pin, stop the rollout.
3. Use separate secrets per environment. First-party HTTP still sends Authorization: Bearer; SDKs still read TYPESAFE_API_KEY. A leaked staging key should not be production. A 401 after promote is a vault miss — do not retry 401.
4. Replay a held-out week after promote. Score policy actions, not only argmax. That replay is your stored POST, not a hidden replay API.
5. Ship a fail-closed default for irreversible acts before you need it. If Jev is down or 529, production must already know what to do.
6. If production uses Vercel AI Gateway, keep promoting Gateway auth (credential or OIDC) and the model id typesafe-ai/jev. Do not paste a first-party pin into that door unchanged.

Pitfalls

• Promoting jev-latest as if it were a SHA. Aliases move when TypeSafe ships.
• Changing criteria in prod without a diff. Reviewers cannot audit a prompt you edited live in the Playground.
• Copying Noul thresholds onto Choice confidence during the promote. Different primitives, different numbers.
• Inventing a TypeSafe staging hostname. Public docs we used name api.typesafe.ai. Isolate environments in your config.
• Inventing per-environment RPM. Vendor defaults are published globally and can change. TypeSafe currently claims 1200 RPM and 250k tok/s; do not invent a staging quota.

What you promote and what you leave behind

Promote: questions, criteria, floors, the versioned model id, and the fail-closed policy. Leave behind: playground-only wording, jev-latest as the production pin, and any staging secret. Re-check: vendor list price ($0.042/M input, output free — vendor claim) so finance is not surprised by unused state tokens after you add fields.

Workspace wiring that made the sandbox honest should already live in git — workspace setup. Auth concepts stay the same after promote; only the secret path changes — authentication concepts.

See also

• Access hub
• Sandbox access
• Workspace setup
• First API call

FAQ

Is there a TypeSafe staging URL?

Public docs we used name api.typesafe.ai. Isolate environments in your config, not by inventing hosts.

Can I promote by replaying production state in sandbox?

Yes if you stored a replayable trace. That is your POST, not a hidden replay API.

Do rate limits differ per env?

Vendor defaults are published globally and can change. Do not invent per-env quotas.

What this page does not claim

This page does not claim TypeSafe ships a promote verb, a staging host, or per-environment quotas. It does not issue keys. It does not treat jev-latest as frozen. 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.