A six-level scale for verification claims, and the finding that completeness is only reachable for formally specifiable properties

arXiv 2608.19009·medium signalXBlueskyLinkedInCopy link

Verification Autonomy Levels run from L0, the model declaring its own answer correct with no deterministic anchor, through L2, objective ground truth for correctness only, to L3 and L4, decidable systems with single-property or domain-level completeness. Reviewing 17 papers across symbolic math, behavior monitoring, medical diagnosis, and code generation, the author shows the field conflates these levels, and argues empirical open-world verification such as fact-checking caps at L2. Useful as a blunt audit question for any pipeline that claims its output is verified, which level is the anchor actually at.