convergence-debugger

Convergence failures stem from misconfigured thresholds, backend mismatches, or early stopping settings. Examine your deliberation transcript for the Convergence Information section and compare convergence_threshold, divergence_threshold, and early_stopping configuration against actual semantic similarity scores and voting outcomes to identify the root cause.

Verify early_stopping.enabled is true, respect_min_rounds is configured correctly, and convergence_threshold matches your semantic similarity backend output range. Check that min_rounds_before_check and consecutive_stable_rounds align with your expected deliberation length.

The skill validates three backends: SentenceTransformer, TF-IDF, and Jaccard. Inspect your configuration file and deliberation logs to confirm which backend is active, then cross-check semantic scores in your transcript against expected output ranges for that backend to detect score discrepancies.

Early stopping failures occur when continue_debate flags remain true despite agreement, when consecutive_stable_rounds haven't been reached, or when min_rounds_before_check hasn't elapsed. Analyze your transcript's Final Voting Results and convergence signals to pinpoint which condition is blocking early termination.

Yes. The skill detects conflicts between voting outcomes and semantic similarity signals across rounds by comparing scores, thresholds, and consensus flags in your transcript and logs, then outputs diagnostic steps to resolve misalignment or recalibrate thresholds.

The skill enforces functional validation and detects missing convergence sections, flagging them as errors in the diagnostic report with actionable remediation steps to restore logging or reconfigure your deliberation system to emit required metrics.