Jev feedback guide
Feedback means reviewer overrides flow into gold JSONL and sometimes into criteria edits. It does not mean TypeSafe updates weights for your tenant. Official models pages we used describe no customer LoRA.
This page is an independent field-guide note, not a TypeSafe playbook. jev.pro is not affiliated with TypeSafe AI. Confirm facts on docs.typesafe.ai. We do not sell 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.
Loop
HITL decision → store override + reason code → weekly error analysis → PR on instructions/criteria or on your downstream model — not a fine-tune ticket to this site.
Pair the override with the decision trace: question SHA, pin, probabilities, action. Without the map you cannot see whether the miss was a flat Choice or a Noul sitting at 0.48. Noul has no confidence field; 0.5 means uncertain, not “medium.” Confidence is peakedness of Choice/Score only.
Criteria are descriptions a person could apply. When the weekly dump says “reviewers keep flipping billing vs collections,” the fix is usually a criterion sentence, not “more AI.” Structured instructions should keep backtick paths stable so gold still matches production state.
Reason codes or it did not happen
Overrides without codes cannot be mined. Require a code. Weekly, dump codes × question id. The fix is usually a criterion sentence, not “more AI.”
Do not send reviewer essays back into state unfiltered — you will train your gold, and you may also inject instructions into the next call. First-party evaluation is still POST /v1/systemone; a poisoned state is a poisoned question.
Playground thumbs-down is a vibe, not a production loop. Official AutoResearch cookbook (vendor): Jev features into a classical model you train. That is your feedback sink, not a TypeSafe weight update.
Poison
Do not silently treat every override as truth (reviewers err). Dual review on disagreements. A rubber-stamp “looks fine” is not gold. Human handoff is a success mode — glossary — not a model failure to erase.
Do not use feedback to invent a fourth primitive or to ask Jev for dollars. Score stays 2–10 ordered levels; keep arithmetic in code; do not treat score as dollars. Choice stays an unordered set (max 255; keep other). Compose atomic questions in code; do not merge three misses into one mega-Choice.
Longer production-feedback explainer: production feedback. Harness: evaluation guide. Scope: when to use Jev. Implementation and launch still apply when you change the pack: implementation checklist, launch checklist. Treating unofficial UIs as a feedback console is an anti-pattern.
What this page does not claim
- No claim the hosted model learns your feedback.
- Not official TypeSafe fine-tune documentation.
- Reason codes are yours; we do not ship a taxonomy.
FAQ
Thumbs-down in Playground?
Useful vibe; not a production loop.
AutoResearch?
Official cookbook: Jev features into a classical model you train. That’s your feedback sink.
Will TypeSafe apply our gold as a LoRA?
Not in the public models story we used. Assume shared weights. Confirm docs.
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: production feedback, evaluation guide, glossary human handoff. Canonical: https://docs.typesafe.ai.