Who We Are

Legal-10 is a production-grade, chained benchmark designed to evaluate the limits of reasoning in Large Language Models. The project emerged from a critical realization: while models have become fluently articulate in legal prose, their ability to maintain substantive logical integrity over complex, multi-step workflows remains fragile.

Unlike existing public benchmarks that test skills in isolation—testing whether a model can identify a hearsay exception or summarize a clause—Legal-10 measures the connective tissue of legal practice. We evaluate whether a model can chain ten distinct reasoning steps without succumbing to the "Cascade Penalty."