Head-to-head
45 curated matchups between flagship models, benchmark by benchmark — each score tagged by evidence status, with harness caveats so a win never reads as decisive on its own.