What Jev cannot do: five limits to understand before building

Jev is specialized, not universal

Jev cannot write free-form text or directly read images, audio, or video. Its structured answers can still be wrong; non-English tasks need separate evaluation, and open-ended planning needs another system. Use it for bounded judgments, with application code controlling consequential actions.

Reviewed September 27, 2026 against TypeSafe’s model reference, known Jev 1.13 failure modes, and confidence documentation. The implementation examples below are our integration recommendations, not measured benchmarks.

It does not generate free-form text

Jev returns predefined choices, ordered scores, and probability-like judgments. It does not write explanations, articles, code, or customer replies.

For a support workflow, let Jev choose between billing, technical support, and manual review; let an LLM or a template write the reply. Include a review option for requests outside your supported categories. A valid category is not permission to issue a refund: application code must check the account and authorize that action separately.

It is text-only today

Images, audio, and video must be converted into text or structured fields before Jev can evaluate them.

A scene controller can send named objects and allowed actions, then let a game engine apply the selected action. An image classifier needs a separate description or OCR step. If that step loses a warning label or misreads a number, Jev cannot recover the missing evidence. Our real-time decision examples illustrate the surrounding workflow, rather than a native video-input capability.

How much context can Jev read?

The current model reference lists 64k tokens across the state and all questions, with a separate 32k limit for the state plus the longest question. These are request limits, not a guarantee that filling the window improves accuracy. Filter irrelevant material first and check the limits again when upgrading model versions.

Type safety is not factual correctness

Jev can guarantee that an answer follows the requested structure. It can still choose the wrong category or assign an unhelpful score.

TypeSafe’s known failure modes include unreliable counting and date comparison, distraction from irrelevant context, and adversarial text influencing decisions. Keep sums and deadline comparisons in code. Supply only the evidence needed for a judgment, and test messages that try to override your classification rules.

For Choice and Score, confidence describes the output distribution; it is not proof that the answer is correct. Noul returns a yes-probability without a separate confidence field. Choose review thresholds using labeled examples from your own task, and inspect confident mistakes as well as uncertain answers.

English is strongest

TypeSafe says other languages, including CJK scripts, are supported but not equally accurate. Test Chinese and multilingual workloads separately.

Keep separate evaluation sets for English, Chinese, and mixed-language requests. Include abbreviations, negation, and local product names. If you translate first, test the whole translation-and-decision pipeline: a good result on the English translation alone does not establish quality for the original input.

It needs bounded decisions

Jev performs best when the possible outputs are known and the surrounding code owns control flow. Open-ended planning and creation still need an LLM, rules, or human judgment.

For document review, ask distinct questions about relevance, missing evidence, and whether review is required. Combine those answers in code rather than asking one vague question to approve the entire document. Use Jev vs LLM to decide which stage belongs to each model.

A practical acceptance check before deployment

1. Write down the allowed outcomes and what should happen when no outcome fits.
2. Label a representative sample, including ambiguous and adversarial inputs. Keep a separate set for final evaluation.
3. Measure errors by category and language. For example, a missed urgent ticket and an unnecessary review should be counted separately because their consequences differ.
4. Compare automatic decisions with the share sent for review. A threshold that improves accuracy by reviewing nearly everything may not meet your operational goal.
5. Record the model version, rules, and thresholds. Re-run the same evaluation before changing any of them.

These are proposed acceptance checks, not evidence that this site has benchmarked your workload. The official model reference explains why version aliases can change behavior. Estimate the calls needed for evaluation and ongoing review with the Jev API cost calculator, then follow the integration guide.