Grading the Graders: Verification Autonomy Levels (L0-L5) for LLM Reasoning
This work introduces Verification Autonomy Levels (VAL), a six-level taxonomy to classify LLM verification schemes based on their source and guarantee. It resolves confusion by focusing on the epistemic status of ground truth, arguing that the anchor source is key to trust.