Skip to content

Jev Criteria and Output Types

Describe finite choices, binary evidence, probabilities, and ordered score levels. Import metadata through Python aliases; the Kedi type grammar does not gain arbitrary function calls.

Typed Criteria

```
from typing import Annotated
from kedi.typesafe import Probability, Rubric

Correctness = Annotated[
    Probability,
    "Does the answer agree with the reference?",
]
Quality = Annotated[
    float,
    "Assess the answer against the reference",
    Rubric(["Incorrect", "Partly correct", "Fully correct"]),
]
```
> adapter: pydantic
> model: typesafe/jev-latest

~Assessment(
  correctness: `Correctness`,
  quality: `Quality`
)
>> The reference is Ankara and the answer is Ankara. The assessment is [assessment: Assessment].
= `assessment`

Probability is a finite float between 0 and 1. A rubric with N ordered levels produces a score between 0 and N-1; float scores retain their fractional part. Integer rubric outputs round to the nearest level, with ties rounded upward. Rubric bounds are automatically included and validated; no separate range is needed.

ChoiceCriteria supplies descriptions for each exact Literal/enum option. BooleanCriteria(true=..., false=...) supplies descriptions for positive and negative evidence. Import these from kedi.typesafe in the Python prelude and use them in Python Annotated aliases. This does not add function calls to native type syntax. Importing kedi.typesafe without the optional package fails with installation advice; ordinary import kedi does not load Jev.

Outputs and Limitations

  • Nested required objects are reconstructed after evaluation.
  • Lists of finite string choices are multi-label decisions, in declaration order.
  • Nullable finite choices include an explicit no-match option.
  • Email, phone, regex, and custom extractors produce finite candidates locally; Jev chooses among them. These are extraction, not arbitrary text generation.
  • Nullable extraction without candidates returns None without a provider request.
  • General unconstrained strings, free-form invokes, arbitrary numeric extraction, media input, recursive schemas, and arbitrary arrays are unsupported.
  • Independent fields share one request. Dependent decisions need separate calls.
  • Unsupported schemas and malformed provider output raise errors; they are not silently converted to false, repaired, or sent to another model.