LLM evals test outputs. Rarely whether the model understood first

(github.com) by noxion | Mar 20, 2026 | 0 comments on HN