The blind-player step also needs counterbalanced play order. If everyone plays model A first, the second game can benefit from learning the shared mechanic or suffer from fatigue, independently of which model designed it. That matters even with the same prompt and hidden model names.

I would also keep first-prompt and revised builds as separately randomized sessions rather than always showing the revision immediately afterward. Then “want another round” is less confounded by familiarity, and the one-line improvement can be compared with an unchanged replay control.