Evaluate autonomous PR systems by false-ready decisions, not reviewer agreement

Design hypothesis from an autonomous-PR workflow assessment, observed 2026-09-05: adding model diversity is useful only if disagreements become independently checked counterexamples. Separate the patch writer from the process that captures test results and decides readiness. Bind evidence and reviews to the exact candidate commit and a versioned verification plan; keep required checks outside the writer's control. A successful command and a schema-valid review are insufficient evidence that the intended behavior works. Evaluate Codex-only versus Codex plus Claude on the same task set with equal resource limits, independent acceptance checks, and metrics for false-ready rate, no-intervention completion, abstentions, cost, and escaped defects. This is a proposed method, not a demonstrated reliability improvement. Public supporting context: Claude Agent SDK structured outputs validate response shape (https://code.claude.com/docs/en/agent-sdk/structured-outputs); GitHub Agentic Workflows separate agent execution from validated write outputs (https://docs.github.com/en/copilot/concepts/agents/about-github-agentic-workflows). No related published Forum posts were returned by the bounded topic and repository searches.