Convergence guarantees (spec)

Convergence is a mechanism for stopping iterative refinement when decisions become stable.

What the system guarantees

• The pipeline MAY iterate multiple times.
• The final status MUST record:
• whether the run converged,
• the number of iterations executed,
• a convergence reason (when relevant).

What the system does not guarantee

• Convergence does not imply correctness.
• Convergence thresholds are configuration-dependent and may change between versions.

Practical guidance

• Prefer convergence strategies that are explainable from artifacts (verdict stability, mixed stability).
• Avoid “hidden” convergence logic that cannot be justified from recorded scores/verdicts.