samfair 0.1.2

SamFair: Algorithmic bias audit and Post-Prediction Neural Linking library

SamFair

SamFair is an algorithmic bias audit and explainability library designed to meet the strict regulatory requirements of the upcoming DPDP Act and India AI Bill 2025.

It provides an end-to-end suite for auditing AI recruitment tools (AEDTs) and extracting readable logic from opaque AI models using our proprietary Post-Prediction Neural Linker (PPNL) algorithm.

🚀 Features

• Synthetic Golden Data Generation: Creates robust, culturally-aware synthetic candidate profiles containing intersectional protected attributes (e.g., Gender, Religion, Caste proxy).
• Adverse Impact Engine: Computes the 4/5ths Rule (80% Rule) across intersectional slices to detect hidden bias.
• Post-Prediction Neural Linker (PPNL): Mines surrogate decision tree rules that lead to rejection for flagged groups. Translates opaque predictions into actionable, human-readable logic (IF feature X THEN reject).
• Immutable Evidence Logging: Every audit and rule extraction is hashed (SHA-256) into a tamper-proof JSONL trail.
• Automated DPIA Reports: Generates ready-to-share PDF Data Protection Impact Assessments outlining the found biases and remediation steps.

📦 Installation

pip install samfair

💻 Quick Start

1. Generating Golden Sets

Generate a synthetic dataset of Indian candidate profiles with embedded protected attributes.

from samfair_lib.synthetic_data import generate_golden_set

df = generate_golden_set(n=1000)
print(df.head())

2. Bias Auditing (4/5ths Rule)

Pass your model's predictions into the audit engine to calculate adverse impact ratios.

from samfair_lib.audit import compute_adverse_impact

# df is your DataFrame containing protected columns
# predictions is a list/array of model outputs (1 for accept, 0 for reject)
protected_cols = ['gender', 'religion', 'caste_indicator']

results = compute_adverse_impact(df, predictions, protected_cols)
flagged_groups = results[results['flagged'] == True]
print(flagged_groups)

3. Explainability with PPNL

If biases are found, use PPNL to extract the logic leading to rejections for flagged intersectional groups.

from samfair_lib.ppnl import ppnl_explain

ppnl_output = ppnl_explain(df, predictions, flagged_groups)
print(f"Extracted Rule: {ppnl_output['rule']}")
print(f"Feature Contributions: {ppnl_output['feature_contributions']}")

4. Logging & Reports

Create tamper-proof audit trails and PDF reports.

from samfair_lib.evidence import log_evidence
from samfair_lib.reports import build_report

evidence_hash = log_evidence("AUDIT_RUN", {"flagged_count": len(flagged_groups)})
report_path = build_report(results, ppnl_output, evidence_hash)

🏗️ Architecture

SamFair operates entirely locally without sending data to third-party APIs. It uses lightweight Scikit-Learn surrogates for high-fidelity rule mining, ensuring you have total control over the audit workflow.

📄 License

MIT License.