https://github.com/SoumilB7/depth-eval Take a look at this My plan was to give this type of instruction set to an llm with a harness just having a calculator and no code execution environment. The most beneficial part of it is it can be solved using a dag easily so the thing that it really challenges a model on is pure instruction depth and overlap maintenance in its reasoning. I have tried to make the readme as intuitive as possible, lmk if you think approaching it like this is a good idea..