Jev review templates
Review templates are screens and reason codes for humans who hold Jev-gated actions. If the UI says “Jev thought…”, you are lying — Jev did not generate a thought.
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.
Screen
Show: source text, question instructions, top-2 options with probabilities, confidence (or noul), suggested action, override control. Criteria are the explanation surface.
Reason codes
wrong_intent, missing_context, policy_exception, model_overconfident, other — feed feedback.
Do not generate a rationale for the reviewer
If you want a paragraph, call an LLM after and label it as generated assistance, never as “Jev’s reason.” Reviewers anchor on fluent lies.
HITL screen that does not invent a mind
Reviewers need enough to override, not a story. A production-minded layout:
- Source text — the filtered
stateyou actually sent, not the whole CRM dump. - Question pack —
instructionsandcriteriaas shipped (SHA visible). - Top-2 options + probabilities — or the Noul band and the Score level text.
- Suggested action from your enum, labeled as policy output, not as “Jev decided.”
- Override control with a required reason code.
- Timeout / error path — if the call 529’d, say so. Do not fake a Choice.
If the ticket is privileged (refund, ban, credit, access), default the suggested action to review or block when the door is late. Official SDKs retry 429/529 only; the UI still needs a human-readable “smart gate skipped” state.
Reason codes that feed a loop
Codes are cheap to store and expensive to skip. Minimum set that analysts can plot:
| Code | Use when |
|---|---|
wrong_intent |
The Choice set missed the real job. |
missing_context |
State was too thin or the wrong fields. |
policy_exception |
Mapping is right; this case is a documented exception. |
model_overconfident |
Probabilities look sharp; the human disagrees. |
other |
Always keep an escape — then tag the next week. |
Feed them into feedback. Do not invent a free-text “rationale from Jev” field that an LLM fills and the UI styles as the model’s inner voice. TypeSafe’s public loop is state in, typed answers out — no chain-of-thought product.
If you want a paragraph for the reviewer, call an LLM after the Jev hop and label it generated assistance. Reviewers anchor on fluent lies. Pair the screen with evidence templates so the packet is complete before anyone types.
Operate failure modes
- Rationale theater. A generated paragraph next to a Noul 0.51 looks like certainty. Hide the paragraph or mark it draft.
- Hidden criteria. If reviewers cannot see the question, they cannot say the question was wrong.
- Second-pass worship. Re-asking Jev after a human note and treating agreement as proof is circular. Log both answers; do not average them into “truth.”
- Queue starvation. Auto-rate up, remainder uglier. Sample autos weekly or a silent criteria typo becomes “the model got worse.”
When not to use this page’s advice
Do not treat these screens as a TypeSafe HITL product spec. Do not ship review UI for math, date reconstruction, or generation — those are jaggedness / out-of-scope, not “needs a nicer modal.” Do not reuse this template on SemIf/OpenJev and call the override “Jev review.” Replicas are not Jev and not endorsed.
See also
FAQ
Should reviewers see the prompt?
Yes. Criteria are the explanation surface.
Re-ask Jev after notes?
Optional second pass; don’t treat agreement as proof.
Is this a HITL product spec?
No. Field-guide notes only.
Can we show “Jev’s rationale” if we generate it locally?
No. Label it as generated assistance or omit it. Jev did not write a thought.
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.