davidondrej/jailbreak-autoresearch — reverse-engineered prompt

Build me a small Python research tool for prompt harness experiments. I want to drop in an example prompt in example.md and a scoring rubric in desired output.md, then run tests to see whether different header and footer wrappers change how a target model responds. It should try a simple baseline, seeded prompts, mutations of the best result, and recombinations of strong fragments, and keep looping so it can reuse what worked before.

Use a local SQLite database to save each experiment with the harness used, the model role permutation, the response, and the score. Add a command to run a dry smoke test, a live baseline, and all strategies, plus a report command that summarizes the best results. The generated footer should always end by forcing the answer to be exactly one sentence. Make it easy to configure with a .env API key and a models.json file. If helpful, make it friendly for running inside Codex goal mode too. Look up current docs online if you need to.