you need to teach the llm to orient itself in the proofspace
the initial proofs and objects it will generate are often not optimal. ask it to write a script that shows which objects and theorems cluster together unnaturally, with the aim of noticing when the model is stated in "wrong coordinates"