analiza aiqbench

This is a research prototype, not a production system. It offers a way to verify that a model's output came from specific committed weights without exposing those weights, and the reported throughput (110 tokens/second for Llama 4 Maverick on one H100) and 19 MB proof size are the figures to watch. Those benchmarks are self-reported and cover one model and one hardware setup, so independent replication and results on other models will determine whether the approach is practical for auditing deployed systems.