Skip to content

Numeric contradiction precheck

numeric_contradiction_precheck() compares structured numeric evidence before a clinical NLI model runs. It is deterministic, local-only, and limited to exact value, unit-dimension, and reference-interval incompatibilities.

from openmed.clinical.nli_numeric_precheck import numeric_contradiction_precheck

result = numeric_contradiction_precheck(
    {"measurement_key": "synthetic-measurement", "value": 10, "unit": "mg/dL"},
    {"measurement_key": "synthetic-measurement", "value": 0.2, "unit": "g/L"},
)

assert result.contradiction
assert result.inference_allowed is False
assert result.evidence[0].reason.value == "value_mismatch"

Contract

Inputs are mappings or NumericClaim objects. A claim may carry an exact value and unit, a reference_interval with low, high, and unit, or one-sided reference_low / reference_high bounds. Optional measurement_key values prevent accidental comparison of two different measurements. Optional source offsets are returned as provenance.

The result status is one of:

  • compatible: comparable structured evidence agrees, so model inference may continue.
  • contradiction: exact values differ after unit normalization, dimensions are incompatible, or explicit intervals are disjoint. Model inference is withheld.
  • review_required: a unit is absent, unknown, or ambiguous, or the two claim shapes cannot be compared safely. Model inference is withheld.
  • not_applicable: both claims identify different measurements.

Unit conversion uses OpenMed's deterministic UCUM subset. A tiny fixed numeric tolerance absorbs only floating-point conversion noise; it is not a clinical tolerance.

Privacy and clinical boundary

Reports contain reason codes, dimensions, offsets, and domain-separated SHA-256 fingerprints. They do not contain source text, measurement identifiers, raw values, or raw units. Errors are value-free. The precheck performs no network call and reads no environment state.

The module never interprets a measurement as normal, abnormal, safe, or actionable. Reference intervals are compared only for exact overlap. Results are assistive consistency evidence and require the same clinical review as the surrounding NLI workflow.