Jev limitations, explained
Separate documented limits, conditional capabilities and unanswered questions before deciding what to trust Jev with.
Published capabilities
Reported demonstrations
Your task accuracy
Real-world reliability
These labels describe our conclusion. Source labels describe who published the evidence.
A decision model is not a general-purpose media or writing tool.
Documented limit
The provider documents that Jev is not trained to generate text. Chaining choices into prose is not a practical substitute.
Choosing a reply category and writing the reply are different jobs.
Use a generative model when the output needs to be new prose.
Documented limit
For jev-1.13.0, the model reference specifies text-only input: strings, JSON objects or arrays of text values. Images, audio and video are not accepted directly.
A system that converts a picture into text is not the same as Jev seeing that picture.
Inspect what a demo actually sends to the model before inferring a media capability.
A documented capability still needs the right task, surrounding software and checks.
Documented limit
TypeSafe documents unreliable counting and date comparisons, plus sensitivity to indirect questions, distracting context and adversarial content.
A structured answer is not proof that the arithmetic or interpretation is right.
Keep exact calculations in ordinary code; test judgment calls separately.
Conditional fit
The model reference says English performs best. Other languages, including CJK scripts, are accepted but do not perform equally well.
Do not transfer an English result to a different language without checking it.
Evaluate representative examples in the language your readers or users actually use.
Conditional fit
TypeSafe derives confidence from the output probability distribution. Its guidance makes action thresholds depend on the domain, risk and observed performance.
Treat certainty as a signal to evaluate, not permission to skip review.
Use labeled examples and a fallback path before letting uncertain decisions trigger actions.
Conditional fit
LangChain’s tool-risk middleware checks only the tools selected for it. A risky call is refused; the middleware does not ask a person to approve it.
This integration example does not establish protection for every action in an application.
Check which actions are covered, who approves them and what happens when a check fails.
Unverified does not mean impossible. It means the evidence here does not settle the question.
Not yet verified
Our review has not measured a failure rate for your tasks. The launch post’s type-safety guarantee is not an empirical decision-accuracy result.
Do not turn a format guarantee into a zero-error promise.
A representative labeled test set, including failure cases, is needed to answer this.
Not yet verified
The model reference says aliases can move to a new version. We have not established that a previously tested threshold remains suitable after an update.
A model name that stays the same can still point to different behavior.
Record the version that answered and repeat the relevant checks when it changes.
The homepage previously marked general image-input support as not yet verified.
Reclassified as a documented limitation for jev-1.13.0 after checking the explicit text-only model reference. This is a change in our verification, not a claim that the product changed today.
View sourceRead the current boundaryExplore the original documentation behind each entry. Publication dates are shown where available; source-check dates record when we reviewed the material.
See our access guide for availability, account requirements and published pricing.
Check access Read our correction principles