tylergarlick/abraxas/guardrail-monitor

Guardrail Monitor Skill

The Guardrail Monitor is a tripartite engine designed to prevent AI drift and ensure that outputs align with both user values and empirical ground truth. It consists of three specialized modules: Pathos, Pheme, and Kratos.

Core Modules

Pathos — Value & Saliency Tracking

Pathos monitors the "emotional and ethical salience" of the interaction. It identifies what the user cares about and flags when a proposed output might conflict with those priorities.

• Function: Extracts value categories (Safety, Accuracy, Privacy, Ethics, Autonomy) and calculates their salience.
• Saliency Score: 0-1, where 1.0 indicates a critical, non-negotiable priority.

Pheme — Ground Truth Verification

Pheme provides a mechanism for verifying claims against trusted external sources.

• Function: Cross-references claims across multiple sources and weights them by reliability.
• Statuses: VERIFIED (sufficient agreement), CONTRADICTED (high-reliability contradiction), or UNVERIFIABLE.

Kratos — Authority & Conflict Arbitration

Kratos resolves contradictions between conflicting sources using a hierarchical authority model.

• Function: Assigns precedence based on source type (e.g., Peer-Reviewed $\rightarrow$ Government $\rightarrow$ Encyclopedia $\rightarrow$ Social Media).
• Domain Rules: Applies specialized precedence rules for specific fields like Medical, Legal, or Scientific.

Commands

check_value_saliency

Analyzes a topic or decision context against tracked user values.

• Arguments: topic (required), decision_context (optional), user_values (optional).
• Behavior: Identifies relevant values, calculates a saliency score, and detects value-level conflicts.

verify_ground_truth

Verifies a claim using a list of sources.

• Arguments: claim (required), sources (optional), require_min_sources (optional).
• Behavior: Returns a confidence score and a status based on source agreement and reliability.

arbitrate_conflict

Resolves a conflict between two competing claims.

• Arguments: claimA, claimB, sourceA, sourceB (required), domain (optional).
• Behavior: Uses authority precedence and domain-specific rules to determine the "winner" and provide reasoning.

Implementation Details

• Architecture: Two-tier Python implementation (FastMCP $\rightarrow$ GuardrailLogic).
• Integrated Logic: Combines value extraction, reliability scoring, and authority hierarchies.

MCP tools

Available MCP tools: check_value_saliency, verify_ground_truth, arbitrate_conflict.