Capability and Verification Claims in LLM Outputs: A Technical Note on Annotation Targets

Large language models are increasingly used as simulated reviewers, editors, and professional advisors. These uses can be useful, but LLM outputs may also present evaluation as completed, verification as performed, or professional judgment as settled. This technical note introduces capability and verification claims as an annotation target beyond ordinary hallucination detection. Capability and verification claims are output-level statements where an LLM appears to claim or imply that it has checked, verified, assessed, judged, or can perform something. The proposed scheme separates primary claim types from output-side markers. Primary types include verification, capability, judgment, and role-invocation claims. Output-side markers record role-conditioned response kata, completion and boundary markers, epistemic-basis markers, and optional boundary disclosures or verification prompts. The aim of this note is not to test user trust directly, infer internal model mechanisms, or evaluate whether a given answer is factually correct. Instead, the aim is to clarify how LLM outputs may present checking, judgment, or professional assessment beyond what is explicitly grounded. This version is a preliminary technical note intended to define the annotation target and provide an initial annotation structure for future empirical work.