What problem does it solve?
When AI models in a deliberation system fail to converge or stop early as expected, it leads to wasted API calls, prolonged debates, and inconclusive results. This Skill provides a systematic debugging guide to diagnose and resolve these issues.
Core Features & Use Cases
- Transcript Analysis: Learn to interpret deliberation transcripts for convergence information, similarity scores, and voting outcomes.
- Configuration Checks: Verify key settings like
semantic_similarity_threshold, min_rounds_before_check, and early_stopping parameters.
- Backend Selection: Understand how the system selects similarity backends (SentenceTransformer, TF-IDF, Jaccard) and ensure the best one is installed for accurate semantic comparison.
- Voting vs. Semantic Conflicts: Diagnose conflicts between model voting outcomes and semantic similarity status, understanding precedence rules.
- Use Case: Your multi-model debate consistently runs for the maximum number of rounds, even when responses appear similar. Use this skill to check if
min_rounds_before_check is too high, if a less accurate similarity backend is being used, or if models are failing to signal continue_debate: false in their votes.
Quick Start
Begin by examining the deliberation transcript (transcripts/*.md) for the "Convergence Information" section to check the reported status and similarity scores.
