TypeSafe’s failure-mode guide for jev-1.13 is more useful than another launch comparison. The vendor identifies limitations that can survive a perfectly valid typed response. The page says it was last reviewed on October 2, 2026; that is a review date, not a verified original publication date.

I read the complete guide. These are vendor-described failure modes, not failures independently reproduced here; it provides no measured error rates or task-specific calibration results.

Keep exact operations in code

  • Counting is unreliable: count parsed items yourself, or classify each item and aggregate in code.
  • Numeric representations can trip it up: convert colors, encoded values, and quantities before asking semantic questions.
  • A Score is a judgment against a rubric—not a way to reconstruct an exact number by interpolation.
  • Extract uncertain date components if necessary; validate, compare, and calculate durations with ordinary date libraries.

Make the question easier to answer correctly

  • State literal conditions and boundary cases; the model may not infer what you intended.
  • Reduce double negatives and multi-hop reasoning; point directly to the relevant state fields.
  • Filter unrelated context before submitting it; more text can reduce decision quality.
  • Align instructions with answer criteria, rather than reversing what “true” means.

Test the decision surface, not just the schema

  • Adversarial state can steer the answer. Clear criteria and edge-case tests are recommendations, not a security guarantee.
  • Choice option order can matter. TypeSafe reports a tendency toward the first option in some cases; permute options while preserving their meanings.
  • Generation is the wrong workload. Chaining choices to produce text is discouraged; use a generative model or construct extraction candidates first.

The practical checklist: test distractors, missing evidence, contradictory criteria, reordered candidates, and inputs that argue for their own classification. Keep permissions and irreversible actions outside the model’s authority.

“State is data, and jev-1.13 does not treat it as hostile by default.” — TypeSafe

Typed output is not a correct decision. A closed answer space removes one parsing problem; it does not remove arithmetic errors, prompt injection, positional bias, or the need to measure mistakes on your own workload.