Laurie Voss, writing for Arize and republished by O’Reilly, asks what changes when a model can classify and score but cannot write an explanation. Jev takes a state—a ticket, output, or agent trace—plus typed questions and returns choices or probabilities. The case is not that every LLM should be replaced; it is that many LLM calls end in a small decision and pay for text generation they do not need.

The early case for a separate decision layer

  • TypeSafe reports substantial speed and cost gains over generative-model judges. Voss cites its four-workflow evaluation: Jev averaged 68% accuracy at roughly $0.0004 and 0.4 seconds per case; GPT-5.6 Terra also reached 68% at a higher reported cost and latency. These are vendor comparisons, not an independent benchmark of every use case.
  • A small independent spam test put zero-shot Jev at 98.3% accuracy versus 98.4% for a trained TF-IDF classifier. Their disagreements were not statistically decisive. The appeal is reaching a useful decision without collecting thousands of labels and training a task-specific model.
  • Voss also cites early moderation and bulk-judgment tests by other developers, but calls the evidence small and recent. Arize had not yet completed its own benchmark at publication.

The probability is as important as the label

An overall accuracy score says how often a judge errs, not which cases need a person. A well-calibrated probability—one that corresponds to the observed frequency of being right—can support an action threshold, a human-review band, and drift monitoring. In the cited spam test, sending the uncertain middle band to a reviewer let the remaining messages reach a reported 99.5% accuracy. That result belongs to that dataset; it does not establish calibration for every Jev task.

“Can’t hallucinate” is narrower than it sounds: Jev cannot return a value outside the supplied answer schema, but it can still choose the wrong allowed answer. Valid output and correct judgment are different properties.

The trade-off: cheap coverage, less explanation

A typed label does not say why a case failed. That missing rationale matters when developers use an LLM judge’s explanations to improve prompts, code, or policies. Voss proposes a hybrid loop: let Jev screen every output or agent step, then send uncertain cases and samples of failures to an explanatory LLM judge or a human. Separate the decision layer from the model that writes text, version its questions, and measure it on the actual workload.

The practical starting point is to count LLM calls that end in one of a handful of labels. Those are candidates for a controlled comparison—not an automatic migration. The article is strongest as an architecture proposal; its dramatic cost and calibration claims still need broader independent replication.