Agentic correctness can be trained. Small models are the goal, swarms of small model slots and multi model variants, combined with training I'm sure can achieve the correctness of much much larger ...