A multi-document benchmark for evaluating LLMs on three forms of scientific reasoning (deduction, induction, causal abduction), with parametric control over inference complexity and premise obfuscation.