judge

To evaluate convergence across multi-agent debate refinements, use the convergence_check mode to generate per-round verdicts. It reads agent refinements and configuration from the workspace, outputting results to debate/round-N/verdict.json for each round.

Final synthesis generation works by reading all refinements and verdicts from the debate workspace to produce a comprehensive solution document. It outputs the final synthesized solution to debate/synthesis.md, including any open issues when needed.

The best way to track per-round verdicts is using the convergence_check mode, which evaluates agent refinements against configurable criteria. It writes structured per-round verdict files to the debate/round-N/verdict.json path within your workspace.

Yes, you need to define convergence criteria in your configuration before checking convergence. The judge reads this configuration from the active debate workspace to evaluate agent refinements and determine if convergence has been reached.

Yes, you can generate a final synthesis document that explicitly includes open issues. When the debate ends, the synthesis mode reads workspace refinements and writes the final solution to debate/synthesis.md, noting unresolved items when needed.

Convergence evaluation requires workspace-aware processing to ensure traceability across rounds. By reading refinements, verdicts, and configuration directly from the debate workspace, the judge maintains a consistent context for accurate per-round and final synthesis.