darwinian-evolver

Evolutionary optimization automates iterative prompt engineering by mutating a starting template toward higher fitness under a user-defined evaluator. It uses trainable failure cases to guide mutations and holdout cases to detect overfitting, improving exact-output constraint reliability.

Evolving regex patterns involves defining a runnable artifact and an evaluator that returns pass/fail scores in the range of zero to one. The evolutionary loop applies LLM-driven mutation guided by trainable failure cases to iteratively refine regex extraction correctness.

SQL query tuning is supported by treating the query as an organism and defining an evaluator that scores correctness from zero to one. The evolutionary loop mutates the SQL query based on trainable failure cases to improve fitness while detecting overfitting via holdout cases.

Overfitting detection uses holdout failure cases alongside trainable test cases to catch solutions that only work on training inputs. The evolutionary loop validates optimized artifacts against these holdout cases, ensuring the evolved prompt, regex, or code generalizes beyond training data.

Python 3.11 or higher is required, along with OpenAI, Jinja2, Pydantic, UV, and Git dependencies. The evolutionary loop runs via CLI or subprocess, requiring you to define an Organism with a runnable artifact and Evaluator logic that safely handles provider failures.