Jev Limitations

This page exists because “fast structured decisions” is incomplete without the failure modes. Most items below are official information from TypeSafe's jev-1.13 jaggedness note, last reviewed there on 2026-09-17.

Generation

Jev does not write replies, code, or explanations. Forcing generation by chaining Choices is officially described as slow and poor. Extract candidates with regex or an LLM, then let Jev pick.

Math, counting, and numbers

Keep arithmetic in code. Official notes: counting characters, term occurrences, or long lists is unreliable. Hex colors and raw RGB compare worse than English color names. Do not interpolate an exact magnitude from a Score that sits between two levels.

# Count in code. Ask Jev only "is this item a fruit?"
items = ["apple", "typesafe", "banana"]
# one Noul per index, then sum(noul > threshold)

Dates and times

Jev reads dates as text, not ordered quantities. Extract parts as Choices (month, day, year, “not stated”), then compare in code.

Literal reading and indirection

It answers the words you wrote. Double negatives and “property of a property” questions lose accuracy. Name the state field you care about.

Large, noisy state

Accuracy falls as unrelated content grows. Filter first. Official context budgets: 64k for the whole request, 32k for state plus the longest question.

Adversarial content

State is not treated as hostile by default. Injected instructions can move the answer. Write precise criteria and test edges.

Contradictory criteria

A Noul where true means “no” performs worse. Align instructions and criteria.

Structural invariants you might assume

Official example: a Noul and a yes/no Choice on similar refund wording did not produce interchangeable probabilities. A question and its negation need not sum to 1. Do not carry a Noul threshold onto a Choice.

Language and input types

English is the primary training language. Other languages, including CJK, are officially “handled but not equally well.” Input is text only — no image, audio, or video.

Calibration is population-level

Official System One docs: calibration does not guarantee that one answer is correct.

When to use Jev anyway

Closed-set routing, binary gates, and single-dimension scores over short, relevant text.

When not to use Jev

Any System Two task: multi-hop reasoning, exact math, document-length generation, or “just figure it out.”

Common mistakes

• Hiding several judgments in one question.
• Sending the whole thread when one message suffices.
• Asking Jev for a number code can compute.

Pair this page with What is Jev and the examples that stay inside these rails.