sunghunkwag/rsi-metaforge-core ? reverse-engineered prompt

Reverse engineered prompt

Build me a Python research runtime that can safely let a program propose its own improvements, test them in a sandbox, and only keep changes that pass hidden verification.

I want the system to feel like a controlled self improving loop, where candidates can be synthesized, adopted, or rolled back, but nothing gets accepted unless a sealed evaluator says it is valid. Please include support for a few different demo modes, like a simple benchmark mode, a file task mode for things like CSV or config repair, and a small self curriculum style loop where the system can generate its own tasks and solve them under strict checks.

Make it deterministic, with frozen baselines, hidden evals, anti cheating checks, and clear evidence logs so results can be reproduced. If you need to look up current docs or best practices online, do that. Keep the code organized and make sure the main script can run the core loop end to end.

Are you gonna build this?

make sure you review the code using coderabbit

Try freeSponsored — opens CodeRabbit in a new tab