The authors propose a framework where evaluation metrics and agent skills co-evolve to enable self-improvement in the absence of reliable pre-existing metrics. This approach utilizes an evolutionary lifecycle to search for compositions of small drawback detectors, creating a transparent and inspectable metric that agrees with anchored reference sets.
- The metric loop is trained to agree with a ten-item anchored reference set and regularized by consensus over unlabeled outputs.
- Double Ratchet co-evolves the metric with a lifecycle-managed skill loop across code generation (MBPP+), enterprise text-to-SQL (Spider~2.0-Snow), and report generation.
- The system retains 88--110% of the held-out lift achieved by skill loops driven by ground truth or the best available rubric.
- Safety is maintained through anchor discipline and outer audits, with an independent judge catching gamed rubrics in report generation tasks.
The authors argue that this failure-expecting architecture serves as the right default for applications where no reliable automatic verifier exists.