darwinian-evolver

Define an Organism holding your prompt template, an Evaluator that scores outputs on a 0-1 scale with failure cases, and an LLM-based Mutator that proposes fixes. Then run EvolveProblemLoop for several iterations and inspect the ranked population snapshots.

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 OpenRouter provider.

Yes, the Organism is a Pydantic model that can hold a regex_pattern, sql_query, or code_block field instead of a prompt_template. You implement a run method that exercises the artifact and an evaluator that scores the results.

Azure-backed OpenRouter models reject phrases like "ignore previous instructions" via content filters. 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 runs can cost a few dollars.