Jev migration guide
Migration means moving decisions off a mega-prompt or a frozen classifier onto Jev questions — not moving off TypeSafe’s undocumented “v0 API” (there isn’t a second public evaluation route in the reference we used). First-party evaluation is POST /v1/systemone plus GET /v1/models.
This page is an independent field-guide note. jev.pro is not affiliated with TypeSafe AI. Confirm facts on docs.typesafe.ai. We do not sell keys. This is not a TypeSafe playbook.
Open-cluster pages are independent field-guide notes. Replicas and third-party interfaces mentioned anywhere on jev.pro are not Jev and not endorsed.
From prompt-only
Extract enums → Choice criteria. Extract yes/no → Noul. Keep drafts on the LLM. Dual-run and compare actions.
Choice is an unordered set (max 255). Add other / none if the old prompt could say “none of these.” Score is 2–10 ordered levels; write descriptions a person could apply. Do not treat the fractional expected score as dollars — keep arithmetic in code. Noul is P(yes); 0.5 means uncertain, not “medium.” Confidence is peakedness of Choice/Score; Noul has none. Structured instructions may be objects or arrays; backtick the field paths you actually send.
See vs prompt-only systems. Scope: when to use Jev. If the old prompt also generated the reply, split decide-then-draft; do not ask System One to write the email.
From a trained classifier
Keep it if volume is huge and labels frozen. Otherwise: labels-in-request. See vs classifiers. A cascade of atomic questions plus a confidence abort often beats a 200-way Choice blob. Compose questions in code; one POST, many questions.
Do not expect parity with a frozen model trained on years of labels. Measure disagreement on actions, not argmax-only accuracy.
Between doors
Rewrite model ids and boolean vs noul. Never lift-and-shift JSON. Door migrations are schema migrations. Boolean vs Noul has already bitten people — see the channel matrix. Pin versioned ids after you fit floors. Version habits: versioning guide.
Shadow mode
For two weeks, write Jev’s action next to the old prompt’s action. Do not act on Jev. Measure disagreement. Only then allow Jev to act on the high-agreement, low-stakes slice.
Big-bang auto-act is the failure mode. Keep the old prompt behind a flag for a sprint. Persist both actions on a decision trace so you can explain the cutover. Write the target action space (including abstain) before you compare.
Wire the client with the implementation checklist. Do not skip the launch checklist just because the old system already shipped. Architectures that silently swap aliases or fail-open: anti-patterns. Path overview: Jev playbook.
What this page does not claim
- No guaranteed parity.
- Not official TypeSafe migration documentation.
- Dual-run agreement is not a safety certification.
FAQ
Big-bang auto-act?
No. Shadow mode, then review-heavy, then narrow auto.
Downgrade path?
Keep the old prompt behind a flag for a sprint.
Is there a second public evaluate API to migrate from?
Not in the first-party reference we used. Evaluation is POST /v1/systemone.
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.
Open-cluster pages are independent field-guide notes. Replicas and third-party interfaces mentioned anywhere on jev.pro are not Jev and not endorsed. Hub: Open. Siblings: versioning guide, vs prompt only systems, implementation checklist. Canonical: https://docs.typesafe.ai.