Open · Last updated 2026-09-21

Jev implementation checklist

This implementation checklist is independent and unofficial. Not a clone of TypeSafe’s build narrative. Keys come from their console or a marketplace, never from jev.pro. Canonical: docs.typesafe.ai.

This page is for builders crossing from curious to production-minded Jev use. jev.pro is an independent unofficial field guide — not affiliated with TypeSafe AI. Cover intents; do not clone rival IA. Confirm facts on docs.typesafe.ai.

Open-cluster pages are independent field-guide notes. Replicas and third-party interfaces mentioned anywhere on jev.pro are not Jev and not endorsed.

Client and channel

Pick one production door; record model id + auth in config (matrix).
First-party: official SDK or POST /v1/systemone with your own 429/529 backoff.
Do not mix boolean / noul or jev-latest / typesafe/jev-1.13 across doors.
Env: TYPESAFE_API_KEY (or the door’s name). Never commit tokens.
Pin jev-1.13.0 (or current versioned id) once thresholds exist; log response.model.

Questions module

Single file for instructions, criteria, and floors.
Dual-stack smoke test (Python + TS) on the same JSON body if both languages ship.
Filter state before POST; measure input tokens.

Errors and traces

Map 401 / 422 / 429 / 529 per errors and retries.
Marketplace 402 / customer_verification handled as callouts, not retries of the body.
Persist a decision trace on every production call (or sampled + all handoffs).

Eval

Gold JSONL + frozen questions (offline evaluation).
Score policy actions, not only argmax.
No invented 193× in dashboards — quote TypeSafe if you show vendor speed.

Order of operations (questions file → gold → HITL)

Implementation fails when people wire the SDK before they can name the judgment. Sequence:

Questions file — one module for instructions, criteria, and floors. Question ids are for you (official: not sent to the underlying model). Put the whole question in instructions.
Filter — allow-list state fields; measure usage.input_tokens. Unused fields still bill (vendor $0.042 / M input, confirm models page) and distract jev-1.13.
Door — one production door; record model id + auth (channels). First-party: official SDK or POST /v1/systemone with your own 429/529 backoff. Env: TYPESAFE_API_KEY (or the door’s name). Never commit tokens. We do not issue keys.
Pin — jev-1.13.0 (or current versioned id) once thresholds exist; log response.model. Do not mix boolean / noul or jev-latest / typesafe/jev-1.13 across doors.
Gold JSONL — freeze questions; score policy actions, not only argmax. Offline evaluation.
HITL — review queue before auto-act. Screens and reason codes: review templates. Persist a decision trace on every production call (or sampled + all handoffs).
Errors — map 401 / 422 / 429 / 529 per errors and retries. Marketplace 402 / verification = callouts, not retries of the body.

Dual-stack smoke (Python + TS) on the same JSON body only if both languages ship. Next: launch checklist. Intent still unset? Design checklist.

No invented 193× in dashboards — quote TypeSafe if you show vendor speed. Schema-safe ≠ factual correctness.

When not to use this page’s advice

Do not treat this as TypeSafe’s build narrative. Do not implement SemIf/OpenJev with these ids and call it Jev — not Jev and not endorsed. Do not skip HITL because the SDK example returned 200. Do not “finish implement” without a questions file and a gold set.

FAQ

Python only?

Official JS SDK exists too — JavaScript SDK.

Is OpenRouter a TypeSafe door?

If you use a marketplace, treat their error codes as callouts. Confirm their docs.

Next after implement?

Launch checklist, then operate (monitoring, review templates).

Can I ship without gold JSONL?

You can ship a demo. You cannot claim the floors are fitted.

Disclaimer

This is an independent unofficial site and is not affiliated with TypeSafe AI; official documentation is available at https://docs.typesafe.ai. Never treat jev.pro as TypeSafe official documentation. We do not sell, issue, or proxy API keys.