The agents proposed 397 experiments, each with a design, independent review, correctness gates, then timing against the best build. 98 produced improvement and 289 retired with a written reason. Every outcome went into a shared record the next proposal reads first.