PyTorch benchmark harness comparing full / fine-tuned / quantised NLP models on accuracy, latency, memory, and energy per 1,000 predictions. Produces accuracy-vs-emissions trade-off curves for stakeholder consumption.

Totsu (凸): Optimization structure toolkit: measure constraint tension, visualize MILP behavior, and explore negotiable trade-offs.