This isn't a vibes-based review. Every test follows the same seven steps, in order, so results can be compared across tests and, eventually, repeated by anyone.
Publish the exact brief first. Before any AI touches it, the brief goes up publicly. No editing after the fact — what you read is what every model got.
Identical source material. Every AI gets the same starting content, constraints, and context. No model gets a hint the others didn't.
No cross-contamination. One model never sees another's output before producing its own — separate sessions, separate context windows.
Compare on five axes. Strategy, writing, design, code, and usability — every test gets scored on the same structure, not whatever felt notable that day.
Blind human read. Real people look at the results with no model name attached before any final verdict is published.
Publish everything, including losses. A model's weak result runs in full — cherry-picking only the winner isn't a test, it's an ad.
Build the combined version. Once everyone's scored, a final build takes the strongest ideas from each — credited, not blended anonymously.
Right now, most "scores" you'll see on this site are self-assessments by the model that built the thing — labeled as claims, not verdicts, because that's what they are. The blind human read is the part that actually settles a test, and it hasn't run at scale yet. Until it does, treat every score here as provisional.