I designed and documented a model-assisted coding comparison with MiniMax M3.1 Flash: six tasks, two repetitions and three languages. The 36 episodes produced 56 attempts including repairs. Successful episodes were 6/12 for SOLONCODE, 8/12 for JavaScript and 6/12 for Python; the preregistered superiority gate was not met. Full task-level outcomes, token accounting and limitations are published.