poorjev

implements Jev's typed `Choice`/`Score`/`Noul` interface on commodity zero-shot NLI models and makes the confidence honest with temperature scaling and conformal abstention, shipping a reproducible calibration eval (ECE 0.170 to 0.071, cross-validated) that runs offline with no API key.

THE SHAPE OF A SCORE CALL

A generic skeleton for this question type, not this project's actual code.

from typesafe import TypeSafe

ts = TypeSafe()
result = ts.evaluate(
    state=candidate,
    questions={
        "relevance": {
            "type": "score",
            "levels": ["low", "medium", "high"],
            "instructions": "How relevant is this to the query?",
        }
    },
)
if result["relevance"] >= 0.7:
    keep(candidate)