GlossaryEval (LLM evaluation)

Eval harness

An eval harness is the runnable infrastructure that, on every model bump, every prompt change, and every release, automatically runs a fixed test set, computes metrics (accuracy, factuality, refusal rate, latency, cost), stores results as a time series, and blocks release if any threshold drops.