veritas-ml 1.0.0

AI Bias Detection and Ethics Compliance Agent

Veritas - AI Bias Detection and Ethics Compliance

Veritas audits sklearn-compatible models for group fairness issues, applies AIF360 mitigation when requested, retrieves AI ethics context from bundled knowledge bases, and produces deterministic or LLM-assisted markdown reports.

Features

• Canonical audit API for models and predictions
• AIF360 fairness metrics and threshold-based issue detection
• AIF360 mitigation registry with automatic strategy selection
• sklearn-compatible VeritasClassifier
• Bundled RAG knowledge bases with Chroma search and lexical fallback
• Import-time knowledge-base auto-initialization through _bootstrap.py
• CLI for knowledge-base init, status, and query

Installation

pip install veritas-ml

For local development:

pip install -e .

Create bias_params.json beside your script or in the working directory:

{
  "protected_attributes": ["sex", "race"]
}

Quick Start

Audit Predictions

from veritas import AuditConfig, audit_predictions

config = AuditConfig(protected_attributes=["sex"])

result = audit_predictions(
    y_true=[0, 1, 1, 0],
    y_pred=[0, 1, 0, 0],
    sensitive=[0, 1, 1, 0],
    config=config,
    X=[[1.0], [2.0], [3.0], [4.0]],
)

print(result.to_dict()["bias_issues"])

Audit a Trained Model

from sklearn.linear_model import LogisticRegression
from veritas import AuditConfig, audit_model

model = LogisticRegression(max_iter=500).fit(X_train, y_train)
config = AuditConfig(protected_attributes=["sex"])

audit = audit_model(model, X_test, y_test, sensitive_test, config)
print(audit.to_dict()["dataset_metrics"])

Run the Full Pipeline

from sklearn.linear_model import LogisticRegression
from veritas import AuditConfig, run_pipeline

config = AuditConfig(
    protected_attributes=["sex", "race"],
    dataset_name="training dataset",
)

result = run_pipeline(
    LogisticRegression(max_iter=500),
    X_train,
    y_train,
    sensitive_train,
    config,
    strategy="auto",
    generate_report=True,
)

print(result.to_dict()["final_report"])

Use the sklearn Wrapper

from sklearn.ensemble import RandomForestClassifier
from veritas import VeritasClassifier

clf = VeritasClassifier(
    base_estimator=RandomForestClassifier(n_estimators=100),
    protected_attributes=["sex"],
    run_audit=True,
    apply_mitigation=True,
    generate_report=True,
)

clf.fit(X_train, y_train, sensitive_features=sensitive_train)
print(clf.get_bias_summary())
print(clf.get_compliance_report())

Knowledge Bases

Veritas keeps import-time auto-initialization in _bootstrap.py. If ~/.veritas/vector_store/ does not exist, import-time bootstrap attempts to ingest the bundled knowledge assets. Search still works through bundled lexical fallback when Chroma is unavailable.

CLI:

veritas init --laws src/veritas/knowledge/ai_ethics_knowledge_base.pdf \
  --algorithms src/veritas/knowledge/aif360_algorithms_documentation.txt
veritas status
veritas query --db laws "gender discrimination in hiring algorithms"
veritas query --db algorithms "reweighing disparate impact"

Public API

• AuditConfig
• AuditResult
• MitigationResult
• PipelineResult
• audit_predictions(...)
• audit_model(...)
• mitigate_model(...)
• run_pipeline(...)
• VeritasClassifier
• BiasDetector as a thin compatibility adapter over audit_model

Testing

veritas_venv/bin/python -m unittest discover -v