sunghunkwag/rsi-bench ? reverse-engineered prompt
Reverse engineered prompt
Build me a Python package called RSI Bench that lets researchers evaluate how well an AI system can improve itself over time. I want it to run a benchmark across six areas, self modification depth, improvement quality, operator discovery, meta adaptation, safety and stability, and autonomous goal generation, then combine them into one overall score. It should let me register a system with simple functions for modifying itself, evaluating performance, reading state, and resetting, then run the benchmark and export results to JSON.
Please include the statistical pieces too, like bootstrap confidence intervals, convergence detection, and Pareto style tradeoff analysis. Make the scoring use a weighted harmonic mean so weak performance in one area hurts the final score. Add clear examples, basic docs, and tests that show how to plug in a mock system. Also include the goal verification flow so autonomous goal generation only scores when a benchmark owned verifier confirms the goal was actually solved. If you need to check current Python docs or best practices, feel free to look them up online.
Are you gonna build this?
make sure you review the code using coderabbit