LangFair Evaluation Card
This card describes LangFair as an evaluation toolkit, not as a benchmark or evidence of clinical readiness. It records the capabilities and boundaries reviewed at commit afd6dc0.
Status and identity
- Name and version: LangFair, repository state at
afd6dc0(2026-06-29). - Stable identifier: Git commit
afd6dc0b053515e2795e5f0dc2f039bac491e668. - Maintainers and contact route: CVS Health; GitHub issues.
- License: The repository contains a
LICENSEfile; confirm the applicable terms for each dependency and artifact before redistribution. - Lifecycle status: Experimental for clinical evaluation. LangFair itself is a general-purpose evaluation toolkit and does not claim clinical validation.
Intended use and decision context
- Evaluation objective: Measure selected group-fairness disparities for binary classification and selected toxicity, stereotype, counterfactual, and recommendation-system behaviors.
- Intended users: Model and application teams performing exploratory fairness evaluation.
- Intended decisions: Identify candidate disparities for investigation and compare a fixed system under a fixed protocol.
- Out-of-scope decisions: Clinical deployment, patient-level decisions, certification, causal fairness claims, or proof that a system is equitable.
- Unit of analysis: Depends on the metric: labeled classification records, generated texts, counterfactual text pairs, or recommendation/search outputs.
- Deployment context represented: User-supplied data. The toolkit does not provide a representative clinical population or healthcare deployment context.
Toolkit composition and coverage
| Component | Construct | Evidence type | Important boundary |
|---|---|---|---|
ClassificationMetrics |
Predicted-prevalence, false-negative, false-omission, false-positive, and false-discovery rate parity | Deterministic metric calculation | Binary labels, binary predictions, and exactly two groups; multi-group and intersectional audits require external orchestration |
| Text-generation metrics | Toxicity, stereotype, and counterfactual similarity or sentiment | Metric and model-assisted evaluation | Outputs depend on metric/model choice and do not establish downstream clinical harm |
| Recommendation metrics | Jaccard similarity, search-result misinformation, and pairwise ranking gaps | Deterministic and model-assisted evaluation | Does not establish clinical utility, calibration, or patient outcome equity |
- Covered failure modes: Selected between-group rate disparities and selected harmful-language or output-consistency signals.
- Explicitly uncovered failure modes: Calibration within groups, decision-curve utility, treatment-effect heterogeneity, access barriers, workflow effects, causal discrimination, and patient outcomes.
- Domain relevance: False-negative-rate gaps can be clinically important, but clinical relevance must be established for the specific endpoint, subgroup definition, prevalence, and care pathway.
- Coverage gaps: No bundled healthcare dataset, clinical benchmark, clinical validation study, or patient-safety adjudication protocol.
Evaluation protocol
- System boundary: Metric layer applied to user-supplied predictions or generated outputs; not an end-to-end clinical system evaluation.
- Prompt and interaction protocol: User-defined and therefore not standardized by this card.
- Scoring procedure: Use the exact LangFair commit, metric configuration, subgroup definitions, label semantics, and input transformation in every report.
- Reproducibility artifacts: Retain the code commit, environment lock, metric configuration, subgroup counts, missing-result policy, and de-identified inputs or a reproducible substitute.
Measures and interpretation
- Primary measure: No universal primary measure. Select measures before evaluation based on the clinical error costs and decision context.
- Subgroup reporting: Report both group denominators, outcome prevalence, confusion-matrix counts, point estimates, and uncertainty. LangFair’s point estimates alone are insufficient when groups are small.
- Thresholds: No general clinical pass/fail threshold is established.
- Comparability conditions: Changes to cohort construction, subgroup definitions, prevalence, thresholds, model version, prompt, or metric configuration invalidate direct comparison unless explicitly adjusted.
- Known metric failure modes: Small denominators, base-rate differences, threshold dependence, multiple comparisons, binary-group reduction, and counterfactual substitutions that alter clinical meaning.
Validation evidence
- Content validity: Not established for a specific clinical use by the toolkit itself.
- Reliability: Must be assessed for model-assisted metrics and repeated generation settings; not reported by this card.
- Construct validity: Must be demonstrated in the target clinical task; metric availability is not construct validation.
- External validity: Not established for healthcare deployment.
- Open validation gaps: Multi-group and intersectional evaluation, uncertainty estimation, calibration, clinical utility, representative cohorts, and clinician or affected-community review.
Safety boundaries and governance
- Unsafe inference risk: A small disparity estimate does not show that a model is clinically safe or equitable, and a large estimate does not by itself identify the causal mechanism.
- Potential harms from use: False reassurance, unstable subgroup rankings, stigmatizing group definitions, privacy leakage from small cells, and clinically invalid counterfactual substitutions.
- Mitigations: Pre-specify endpoints and groups, suppress unsafe small-cell disclosures, attach uncertainty, review substitutions clinically, and use domain-expert and affected-stakeholder review.
- High-stakes boundary: LangFair results must not be used alone for clinical deployment or patient-care decisions.
- Issue and correction route: Report toolkit defects in LangFair issues; report errors in this card in the Trust and Safety Evaluations repository.
Taxonomy recommendation
Catalog LangFair under Ethical and Bias Testing → Fairness Evaluation as a use-case-level evaluation toolkit. Do not list it as a standalone benchmark, and label clinical use as experimental until a target-specific validation record is available.
