SDOH false-positive stress gate¶

openmed.eval.sdoh_false_positive_stress measures patient-level SDOH false positives on repository-authored synthetic hard negatives. The matrix covers five Social History categories and seven patterns per category: screening, education, unanswered boilerplate, third-party language, out-of-section text, negation, and historical mentions. It never loads a restricted corpus or calls a remote model.

The default predictor applies OpenMed's local section scope and experiencer filter. Non-assertive screening, education, and unanswered template clauses are excluded before a finding enters the patient-level view. Third-party findings remain available to the existing experiencer review layer but are not counted as patient findings. Negated and historical use remains noncurrent. Double-negated unemployment stays unknown for review rather than becoming an automatic positive finding.

from openmed.eval.sdoh_false_positive_stress import (
    assert_sdoh_stress_gate,
    run_sdoh_false_positive_stress,
)

report = run_sdoh_false_positive_stress()
assert_sdoh_stress_gate(report)
print(report.to_dict())

Every category has its own false-positive rate and ceiling. The default ceiling is zero for each category; a caller may supply an explicit per-category ceiling for evaluation. The count of attempted automated eligibility actions must always remain zero and cannot be relaxed. The report contains only controlled category names, counts, rates, gate results, and the action count. It contains no source text, finding values, identifiers, or spans.

This synthetic gate covers the included patterns; it is not a clinical decision guarantee or an estimate on a real population. Permissioned corpora must remain eval-only, user supplied, and outside the repository. Keep review and eligibility decisions with qualified humans.