Llm-Inference

Pre-release: measures deployed model behavior with deterministic grading instead of treating a model name as a benchmark.