The capability I’m asking for is an end-to-end model-research loop: understand what a model needs to learn, propose a suitable architecture and training objective, design informative experiments, diagnose failures from the evidence, and revise the design…