convergence-debugger

What problem does it solve?

AI deliberations may not converge as expected, early stopping might fail, or semantic similarity scores could be incorrect, leading to wasted API calls and inefficient AI interactions. This skill provides a systematic guide to diagnose and fix these issues, ensuring your AI models reach consensus efficiently.

Core Features & Use Cases

• Transcript Analysis: Quickly identify convergence status, similarity scores, and voting outcomes from deliberation transcripts.
• Configuration Verification: Debug common misconfigurations in config.yaml related to thresholds, min rounds, and early stopping.
• Backend Selection & Debugging: Verify which semantic similarity backend is in use (SentenceTransformer, TF-IDF, Jaccard) and debug score discrepancies.
• Early Stopping & Impasse Debugging: Diagnose why early stopping isn't triggering or why impasse detection isn't working as expected.
• Use Case: If your AI deliberation runs for too many rounds despite models appearing to agree, use this skill to check config.yaml for early_stopping.enabled and respect_min_rounds, and examine the transcript for continue_debate: false flags in model votes.

Quick Start

Begin by examining the latest deliberation transcript in the transcripts/ directory. Look for the "Convergence Information" and "Final Voting Results" sections to understand the deliberation's outcome and status.