darwinian-evolver

Define an Organism holding your prompt template, an Evaluator that scores outputs in [0, 1], and a Mutator that asks an LLM to fix failures. Then run the evolution loop for several iterations and inspect the ranked population in the output snapshots.

Copy the custom problem template, set the artifact field to your regex pattern or SQL query, implement a run method that tests it against labeled examples, and provide trainable and holdout failure cases. The mutator proposes improved versions based on observed failures.

Yes, the upstream CLI is hardcoded to Anthropic, but the shipped parrot_openrouter.py driver routes LLM calls through OpenRouter using the OpenAI SDK. Set OPENROUTER_API_KEY and EVOLVER_MODEL to use any model on OpenRouter.

Provider content filters, especially Azure-backed OpenRouter models, reject phrases like "ignore previous instructions". Wrap the LLM call in try/except and return an error string so the evolver scores that organism 0 and continues instead of crashing.

Avoid it when the optimization target is differentiable (use gradient descent or DSPy), when you only need 2-3 variants you can write by hand, or when the fitness signal is purely subjective with no measurable criterion.

A typical run uses 50-500 LLM calls depending on iterations, parents per iteration, and concurrency settings. On gpt-4o-mini that costs pennies, while Claude Sonnet can cost a few dollars per run.