rti-ethics-engine 2.0.0

India's first granular AI ethics framework for RTI applications (36 metrics, 9 dimensions)

RTI Ethics Engine (v2.0.0)

India's first granular AI ethics framework for RTI applications - A production-ready Python package implementing 36 metrics across 9 ethical dimensions with granular formula access and dynamic data computation.

---

🚀 Installation

pip install rti-ethics-engine

⚡ Quick Start (Dynamic Mode)

The recommended way to use the engine is to compute metrics directly from your application data:

import numpy as np
from rti_ethics_engine import evaluate_from_raw

# 1. Prepare your data
query = "What is the budget for the Metro project?"
response = "The budget for the Metro project is 5000 Crores..."
predictions = np.array([1, 1, 0, 1])  # Model predictions
labels = np.array([1, 1, 1, 1])       # Ground truth
sensitive_attr = np.array([0, 1, 0, 1]) # Demographic group (e.g. Urban/Rural)

# 2. Evaluate with one call
result = evaluate_from_raw(
    query=query,
    response=response,
    predictions=predictions,
    labels=labels,
    sensitive_attr=sensitive_attr,
    latency_ms=450.0,
    grounding_score=0.92
)

# 3. Get results
print(f"Label: {result.final_label}")  # "ETHICAL"
print(f"S_LLM Score: {result.s_llm}")  # 0.942
print(f"Risk Band: {result.risk_band}") # "LOW"

📊 The 3-Layer Audit Protocol

The RTI Ethics Engine uses a rigorous 3-layer protocol for every dimension:

1. Layer 1: Hard Gates (L1) - Pass/Fail compliance checks (e.g., Statistical Parity < 0.1).
2. Layer 2: Risk Scoring (L2) - Normalized risk vectors weighted by policy importance.
3. Layer 3: Forensic Evidence (L3) - Detailed context and reasoning for the scores.

Granular Access

# Check if a specific dimension passed its hard gates
if result.get_l1_status("fairness") == "FAIL":
    print(f"Fairness fail reasons: {result.get_l1_reasons('fairness')}")

# Get granular metric values (36 metrics available)
spd = result.get_l1_metric("fairness", "spd")
psi = result.get_l1_metric("drift", "psi")

# Export complete audit trail
audit_json = result.to_json()

🏗️ 9 Ethical Dimensions

Dimension	Description	Key Metrics
Legal Compliance	RTI Act 2005 & Data Policy alignment	Lawfulness, Consent Validity
Bias & Fairness	Demographic equity across groups	SPD, DI, EOP, AOD
Explainability	Transparency & Model Confidence	FI Sum, Counterfactual Distance
Security	Robustness against prompt injection	PISR, Adversarial Accuracy
Monitoring	Audit trail & SLA integrity	Audit Completeness, SLA rate
Performance	Statistical accuracy of responses	Recall, Group-F1, Accuracy
Human-AI UX	User friction and frustration levels	TCR, Cognitive Load
Drift	Model/Data stability over time	PSI, KS-Statistic
Environment	Carbon footprint & compute cost	GPU Hours, Inference Cost

📖 Documentation

For detailed input/output schemas and advanced configuration, see:

• INPUTS_OUTPUTS.md - API Reference
• FORUMULAS.md - Mathematical Manifest

🧪 Testing

pytest tests/ -v

🇮🇳 Why RTI Ethics Engine?

Designed specifically for Right to Information (RTI) workflows, it ensures that AI systems responding to citizen queries are:

• Legally Sound: Compliant with Section 4/8 of the RTI Act.
• Fair: No bias between rural/urban or language variations.
• Auditable: Providing forensic evidence for Central Information Commission (CIC) audits.

---

License: MIT | Support: github.com/rti-ethics/rti-ethics-engine