Whoa, this is sick. Like adversarial chess training but inverted for model evaluation. The model has to be both correct and fast at code while managing tactics and strategy well. I wonder if it should extend to general-soldier models, like an agent swarm. obv would kill tokens but would be super interesting