Worked example, 26 Aug 2026. Claude and OpenAI models, judged from screenshots with the names hidden
Four AI drafts (Claude Opus 5, Claude Fable 5, and one OpenAI model, GPT-5.6 Sol, run twice with different instructions) were given the same brief and built a landing page each, without seeing one another's work. Three judges then scored them from anonymised screenshots (the two outside judges blind; our own Claude judge could recognise its own page, a risk written down before the run), and Jody picked one with the labels off. These are the real pages, live and scrollable, not pictures of them. Scroll inside any frame.
Coastline Groundworks Ltd is a fictional worked example built by ConstructionX AI. No real firm, person or job is depicted. No client data appears anywhere in this test.
Test A
Unanimous: all three judges put the OpenAI run 1 page first. Jody picked the Claude Opus 5 page, which both outside judges ranked last.
Test B
Three judges, three different first places. No majority. This is the leg that decided the verdict.