/ 01
Short Answer
Jev 1.13 has nine documented failure modes. It reads instructions literally, cannot count or do arithmetic reliably, treats dates more like text than ordered values, and becomes less reliable with irrelevant state. TypeSafe publishes these limits alongside workarounds. Most fixes keep precise computation in code and reserve Jev for a narrow semantic judgment.
/ 02
Why We Publish This
This list comes from TypeSafe’s own jaggedness page, which applies to jev-1.13 and was last reviewed September 17, 2026. It matters when deciding where the model belongs in a production workflow.
/ 03
The Nine Failure Modes at a Glance
| # | Failure mode | Do this instead |
|---|---|---|
| 1 | Literal reading | Write exact conditions and boundary cases in instructions and criteria. |
| 2 | Math and numbers | Count, convert and calculate in code. |
| 3 | Date and time comparison | Extract parts with Jev; compare dates in code. |
| 4 | Indirection | Reduce reasoning hops and name the relevant state. |
| 5 | Large state with irrelevant detail | Filter before sending the state. |
| 6 | Adversarial content | Use precise criteria and test hostile inputs. |
| 7 | Contradictory instructions and criteria | Align criteria with instructions. |
| 8 | Common-sense structural invariants | Ask each decision one way; enforce identities in code. |
| 9 | Generation | Use a generative model for prose. |
/ 04
1. Literal Reading
Jev answers the question as written, including qualifiers, negations and implied conditions. State the exact condition in instructions and put boundary cases in criteria. If you find yourself explaining what you meant after a bad result, add that missing condition or split it into smaller questions.
/ 05
2. Math and Numbers
Jev is not a calculator. It may recognize the shape of a plausible count rather than actually tallying characters, words or list entries; errors grow with size. Count in code. Semantic representations work better than numeric encodings such as hex color values. A Score can be fractional, but that does not make it a precise measurement scale for reconstructing numbers.
/ 06
3. Date and Time Comparison
Jev does not reliably order dates, compute durations or handle windows, especially with mixed formats or relative references. Use Choice to extract bounded date components, including an explicit “not mentioned” option. Validate and compare the assembled date in code.
/ 07
4. Indirection
Double negatives, properties of properties and multi-hop references cost accuracy. Point instructions directly at the relevant named state and reduce the number of reasoning steps.
/ 08
5. Large State Full of Irrelevant Detail
Irrelevant material interferes with decisions and makes failures harder to debug. Filter the state to fields needed for the question. TypeSafe’s direct API model page lists 64k tokens for the state plus all questions combined, with a separate 32k limit for the state plus the longest question. OpenRouter’s endpoint metadata lists context_length: 32000 for the route used by this playground. The two published figures use different descriptions; we have not tested whether OpenRouter accepts requests whose combined questions exceed 32k. Plan against the listed 32k window until that behavior is confirmed.
/ 09
6. Adversarial Content
The state can contain injected instructions or self-serving claims. Jev 1.13 does not always treat that material as hostile. Define precise criteria and test adversarial examples before using it for moderation, guardrails or RAG verification.
/ 10
7. Contradictory Instructions and Criteria
Criteria that reverse or conflict with the instruction confuse the model. Make true and false descriptions extensions of the same plain-language question.
/ 11
8. Common-Sense Structural Invariants
Related question forms need not satisfy simple arithmetic identities. TypeSafe shows that a Noul and a Choice about the same support ticket can assign different comparable yes probabilities. A question and its negation may produce Noul values that do not sum to 1. Do not transfer a Noul threshold to Choice or require separate outputs to obey an equation. Choice is relative among options; separate Nouls are absolute checks.
/ 12
9. Generation
Jev is not trained to generate prose. Chaining choices to force text generation is slow and weak. Use Choice for a bounded extraction or a generative model for open-ended writing.
/ 13
Other Hard Limits
| Constraint | Jev 1.13 as reviewed |
|---|---|
| TypeSafe direct context | 64k tokens across state and all questions; state plus longest question limited to 32k. |
| OpenRouter context | 32k window listed for typesafe/jev-1.13; this playground uses OpenRouter. |
| Input | Text and text-valued JSON, no image, audio or video. |
| Language | English is best supported; test other languages on your data. |
| Rate limits | TypeSafe direct API lists 250,000 tokens/second and 1,200 requests/minute, subject to change. OpenRouter paid Jev has no published fixed platform request cap; provider limits and credits can still cause 402/429. |
| Price | $0.042 per million input tokens and free output on the current OpenRouter model listing. |
The 50 or 1,000 requests/day figures in OpenRouter’s free-model guidance apply to free model variants, not paid Jev. See OpenRouter’s limits guide and its plan comparison.
/ 14
The Official “Avoid” List
- Do not assign exact computation to Jev.
- Do not hide several judgments in one question.
- Do not use it for multi-hop System Two tasks.
- Do not add irrelevant state context.
Keep semantic judgment and deterministic calculation separate.
/ 15
How We Handle This in the Playground
The form uses typed Choice, Score and Noul questions, not a chat box. Example probabilities are marked as demo values. You can add multiple questions to a request, and live calls are identified separately from previews.
/ 16
Try It in the Playground
Explore typed questions in the Jev playground. See Noul thresholds and JSON output before integrating.
/ FAQ
Frequently asked questions
Does Jev have known limitations?
Yes. TypeSafe publishes nine failure modes for Jev 1.13.
Can Jev count or do math?
Not reliably. Perform counting and arithmetic in code.
Can Jev compare dates?
Date extraction may be structured, but comparisons and calculations belong in code.
Why do Noul and Choice give different answers to the same question?
They are different question shapes; TypeSafe does not promise cross-type probability identities.
Can Jev be prompt-injected?
Adversarial content in state can influence Jev 1.13. Test hostile inputs and use precise criteria.
What is Jev’s context length?
TypeSafe’s direct API lists 64k tokens for state and all questions together, plus a 32k cap on state and the longest question. OpenRouter metadata lists 32k for this playground’s route. We have not empirically tested the boundary, so use 32k as the conservative planning figure.
Will these issues be fixed in a later version?
TypeSafe says many may be fixed later. This page is explicitly about jev-1.13, reviewed against official notes from September 17, 2026.