fairscope v0.3.0

fairscope v0.3.0

Added

• fairscope.federated — cross-node (federated / multi-site) fairness audit. FederatedFairnessAudit + FederatedReport: per-node DeLong AUC CIs, ECE, Brier, F1; cross-node disparity (max−min AUC gap + Bonferroni-corrected pairwise unpaired DeLong); optional per-node recalibration (temperature/isotonic) with pre/post ECE; per-node AUC forest, reliability curves, and PDF export. Audits per-node predictions only — no training, no privacy guarantee. Routed via FairnessAudit(model, domain="federated", ...).
• fairscope.lending — mortgage-lending fairness audit. LendingFairnessAudit + LendingReport: a descriptive annual approval-gap analysis (symmetric disparate impact per year, composing core) plus an optional subgroup CATE via Causal Forest DML (estimate_cate, econml.dml.CausalForestDML). Causal claims are conditional on the DML assumptions; econml is the optional fairscope[lending] extra. Routed via FairnessAudit(model, domain="lending", ...).
• Documentation pages for both modules + API reference, and replication notebooks notebooks/03_lending_replication.ipynb and notebooks/04_federated_replication.ipynb (synthetic; executed in CI via nbmake).

100% test coverage maintained; CI green on Python 3.9–3.12.

Full changelog: CHANGELOG.md

fairscope v0.2.0

fairscope v0.2.0

Added

• fairscope.nlp — five-axis Cross-Platform Fairness Evaluation (CPFE) protocol. CPFEProtocol + CPFEReport: macro AUC/F1 and ΔAUC%, multiclass ECE, bootstrap macro-AUC significance with Bonferroni correction, per-class disparate impact and equalized odds, a structured per-axis deployment_readiness() diagnostic (P4 reference bands; illustrative, configurable ΔAUC limit), and gradient-saliency Jaccard attribution stability (token_saliency behind fairscope[nlp]). Routed via FairnessAudit(model, domain="nlp", ...).
• Documentation site (MkDocs Material + mkdocstrings) at https://rajveer-code.github.io/fairscope/ — runnable getting-started example on the synthetic fixture, CPFE and healthcare guides, auto-generated API reference.
• Replication notebooks (notebooks/01_healthcare_replication.ipynb, notebooks/02_nlp_cpfe_demo.ipynb) executed in CI via nbmake.

100% test coverage maintained; CI green on Python 3.9–3.12.

Full changelog: CHANGELOG.md

fairscope v0.1.0

First release of fairscope — subgroup-stratified, calibration-aware fairness auditing for ML models, grounded in peer-reviewed methods.

core/

DeLong AUC confidence intervals + paired/unpaired tests, a stratified bootstrap AUC test, Expected/Maximum Calibration Error with reliability diagrams, temperature-scaling and isotonic recalibration, Bonferroni and Benjamini–Hochberg corrections, and subgroup metrics (symmetric disparate impact, equalized odds difference). 100% test coverage.

healthcare/

HealthcareFairnessAudit + HealthcareReport — a one-call clinical fairness audit composing core/ (per-subgroup DeLong CIs, ECE, Bonferroni-corrected pairwise tests, Brier/F1), with report tables, an AUC forest plot, reliability-curve plots, multi-page PDF export, and an optional SHAP summary (fairscope[shap]).

Top-level FairnessAudit(model, domain="healthcare", ...) dispatcher.

Methods are ported from published work and cited in docstrings; only the forthcoming five-axis CPFE protocol is presented as novel. Wheel and source distribution attached.