Clinical NLI verification¶

OpenMed exposes a small, backend-neutral natural-language-inference (NLI) stage for checking whether a generated or grounded claim is supported by a source span:

from openmed.clinical.nli import nli, verify

pair = nli(
    "Synthetic patient has pneumonia.",
    "The patient has pneumonia.",
)

checks = verify(
    ["The patient has no pneumonia.", "The patient has pneumonia."],
    "Synthetic patient has pneumonia.",
)

nli always returns a mapping with label (entailment, contradiction, or neutral) and a finite score in [0, 1]. The default backend is a deterministic, dependency-free heuristic for offline development and synthetic fixtures. It is intentionally conservative and is not a trained clinical model.

verify can be called from a future verify=True option on a summarization or grounding stage. It evaluates every claim and returns its original claim and source together with the NLI result and an explicit contradicted flag. A contradiction is surfaced for review; it is never silently dropped. A local trained MLX head can replace the backend through the backend= argument or set_default_backend() while preserving the nli and verify APIs.

MedNLI data policy¶

MedNLI is DUA-gated and eval-only. The BigBio mirror is represented by a gated stub; OpenMed does not bundle, download, or use MedNLI at runtime. Authorized evaluation code must provide its own approved local access through the existing eval-only dataset boundary. No MedNLI records or model weights belong in the repository.