Summary omission budgets¶

openmed.clinical.summary_omission_budget provides a deterministic gate for checking whether a generated summary omitted policy-important source facts. It is an offline review control, not a clinical decision or compliance certification.

Each caller-defined importance class has a positive severity weight and its own omission limit. Limits are evaluated independently, so strong coverage in one class cannot compensate for an over-budget class. A mandatory class always has a zero limit and any omission from it returns mandatory_class_omitted, regardless of aggregate coverage.

import hashlib

from openmed.clinical.summary_omission_budget import (
    ImportanceClassPolicy,
    SummaryEvidenceCoverage,
    evaluate_summary_omission_budget,
)


def opaque_id(value: str) -> str:
    return "sha256:" + hashlib.sha256(value.encode()).hexdigest()


mandatory = ImportanceClassPolicy(
    class_id=opaque_id("mandatory"),
    severity_weight=10,
    mandatory=True,
)
report = evaluate_summary_omission_budget(
    [
        SummaryEvidenceCoverage(
            evidence_id=opaque_id("synthetic-fact-1"),
            importance_class_id=mandatory.class_id,
            represented=False,
        )
    ],
    [mandatory],
)
assert not report.passed
assert report.refusal_code == "mandatory_class_omitted"

Evidence and policy classes use opaque sha256:<hex> identifiers. The report contains only identifiers, counts, weights, limits, and categorical refusal codes; it never carries source facts, summary text, or evidence identifiers. Invalid-input exceptions are categorical and do not echo caller values.

The gate does not generate text, call a model, make network requests, or claim that a summary is clinically complete. Callers remain responsible for defining and reviewing appropriate importance classes and limits.