Evals for LLMs: the layer after tracing that decides whether your model performs or only looks like it doesMay 25, 2026