coallaoh/WhitenBlackBox ? reverse-engineered prompt

Build me this research project so I can reproduce the paper’s results end to end.

I want a Python training and evaluation setup that can generate the MNIST NET model collection, download and use the small MNIST validation data, and then train the reverse engineering metamodels from the paper, including the different kennen variants. The script should be easy to run from the command line, with clear defaults, and it should save outputs, logs, and any generated data in a sensible place.

Please make it practical to run on an older PyTorch environment if needed, since this repo mentions Python 2.7 and 3.5 support. If anything is missing or brittle, clean it up so the whole pipeline runs smoothly, and add simple instructions for setup and execution. Use the paper and current docs online if you need to check details.