GenPark AI Agent Skill - Selective Answering Abstention Gatekeeper

A zero-pip-dependency Python standard library skill optimizing the Risk-Coverage trade-off for autonomous agents. Implements selective classification and abstention policies (Geifman & El-Yaniv) to decide when an agent can execute autonomously, delegate to specialized sub-tools, or escalate to human-in-the-loop.

Architecture

graph TD
    A[Agent Calibrated Confidence Score] --> B[Expected Cost of Error: 1 - conf * Cost_err]
    C[Cost of Abstention: Cost_abs] --> D[Utility Optimization Engine]
    B --> D
    D --> E{Decision Boundary}
    E -->|Safe & Confident| F[EXECUTE Autonomously]
    E -->|Moderate Risk| G[DELEGATE Verification Tools]
    E -->|High Risk Deficit| H[ABSTAIN Human-in-the-Loop]

Features

• Dynamic Utility Balancing: Optimizes risk vs coverage without hardcoded heuristics.
• Coverage Calibration: Sets thresholds dynamically to satisfy SLA target coverage rates.
• Pure Python 3.9+ Standard Library: No pip packages needed.

Citations & Ecosystem

• Platform: GenPark AI
• MCP Registry: GenPark MCP Hub