Jev is a discriminative model — it classifies and scores rather than generates — so the natural comparison is not GPT versus Jev but a trained text classifier versus Jev. Both return a label and a probability. The difference is where the knowledge of your labels comes from.

The core difference

A traditional classifier (logistic regression on TF-IDF, a fine-tuned BERT, a gradient-boosted model on embeddings) learns your label set from labelled examples. Change the labels and you collect data and retrain.

Jev learns nothing per customer. TypeSafe says the same weights serve every account and Jev is not fine-tuned on customer data (TypeSafe docs). You describe the options in plain language in the request, and it classifies zero-shot against those descriptions (Wikipedia).

Trained classifier Jev
Labels defined by Training data Natural-language criteria in the request
Changing labels Relabel + retrain + redeploy Edit the criteria
Cold start Needs hundreds–thousands of examples Works on day one
Probabilities Often over- or under-confident without extra calibration Trained for calibration (RLCD)
Context Usually short text Up to 32k tokens per question; JSON state with named fields
Latency Sub-millisecond to ms, self-hosted 70–500 ms, hosted (vendor figure)
Cost at huge volume Near-zero marginal cost $0.042 per million input tokens, output free
Hosting Anywhere, including offline Hosted API only

Where Jev wins

  • Labels change often. New product areas, new policy categories, new tools for an agent. Editing a criteria map beats a retraining cycle.
  • No labelled data yet. Jev gives a usable baseline immediately; you can use its outputs to bootstrap labels later.
  • Many small decisions. One request can ask a Choice, a Score and several Nouls in parallel. Building that as classic models means one model per question.
  • The decision needs context. "Is this refund request consistent with the order record?" depends on a JSON state, not a bag of words.
  • You need honest uncertainty. Calibrated probabilities make act / confirm / escalate routing practical.

Where a trained classifier still wins

  • A stable label set with lots of data — spam vs not-spam on millions of labelled emails. A small model will be cheaper per call and may be more accurate on that exact distribution.
  • Hard latency or offline constraints — on-device, sub-millisecond, or air-gapped.
  • Extreme volume where even a fraction of a cent per thousand calls matters.

A practical combination

TypeSafe's own docs describe using Jev's probabilities as features for a downstream classical model, and a hierarchical classification recipe that beam-searches over Choice probabilities (TypeSafe cookbooks). A common path is:

  1. Launch with Jev and confidence gating.
  2. Log decisions and human corrections.
  3. If one decision becomes stable and very high-volume, distil it into a small classifier — and keep Jev for the long tail and new labels.

Try it

The GitHub issue classification and research paper triage use cases show criteria-based classification with copyable code. For the basics, see What is Jev?.

FAQ

How is Jev different from a traditional text classifier?

A trained classifier learns your labels from labelled examples and must be retrained when they change. Jev is not fine-tuned per customer: you describe the options in the request and it classifies zero-shot against those descriptions.

When should I use Jev instead of a trained classifier?

When labels change often, you have no labelled data yet, you need many small decisions per item, the decision depends on JSON context, or you need calibrated uncertainty for act / confirm / escalate routing.

When is a trained classifier still better?

For a stable label set with lots of data, hard latency or offline constraints such as on-device or air-gapped use, and extreme volume where even tiny per-call costs matter.

Can Jev and a classifier be combined?

Yes. Launch with Jev and confidence gating, log decisions and human corrections, then distil stable high-volume decisions into a small classifier while keeping Jev for the long tail and new labels.