Does Jev Learn From My Feedback?

Not on its own. Each call to Jev, TypeSafe AI's decision model, is an independent verdict: it judges what you send in that request, and nothing in the ecosystem documentation we've seen describes it remembering your past calls or your corrections. For data-use and retention terms, docs.typesafe.ai is the only source that counts.

That's not the dead end it sounds like. "Learning from feedback" just moves out of the model and into your pipeline, where you can see it, version it, and roll it back.

Where your feedback actually goes

Into the questions. Most wrong verdicts trace back to a boundary the question never defined. When a reviewer overrides a verdict, the fix is usually a clause ("cosmetic damage counts; disliking the color does not"), and that's the whole craft. A reworded question is a new instrument, so version it.

Into an optimizer. The jev-align library formalizes this loop: you label examples, an optimizer (GEPA) rewrites the question, and what comes out is a Jev function with a known error rate, per its author's description. Your feedback trains the question, not the weights, which is the realistic shape of "Jev learning."

Into a log. Every verdict plus every human correction is a labeled example. That log is the raw material of the next section.

The log-then-retrain pattern

If you truly want a model that absorbs your corrections, build it from the log. This is the "start LLM, graduate the survivors" sequence from the LLM vs traditional ML guide: run Jev in production, log verdicts and overrides, and once a task proves high-volume and stable, train a classic classifier on that data. Jev then judges the edge cases and watches for drift.

The labeling side of that pipeline, spot-checks and adjudication included, is covered on the data labeling page. The short version: humans audit the labels before anything trains on them, because a model trained on unaudited verdicts faithfully learns their mistakes too.

Frequently asked questions

Does Jev remember previous requests?

Nothing we've seen in the ecosystem documentation describes memory between calls; each verdict depends on the question and evidence in that request. Confirm current behavior at docs.typesafe.ai.

How do I make Jev better at my specific task?

Improve the questions: add boundary clauses where reviewers disagree with verdicts, then re-run your labeled cases. The judge-questions guide is the method.

Can I train my own model on Jev's verdicts?

Yes, that's the graduate pattern: log verdicts, have humans audit a sample, and train a classifier on the result, as described in the LLM vs traditional ML guide.

Is "feedback" the same as fine-tuning here?

No. Fine-tuning changes weights; feedback in a Jev pipeline changes question text or trains a separate model. See whether you can fine-tune Jev for that distinction.