AI provenance labels are increasingly used in scholarly publishing, but most disclosure systems still compress a complex workflow into a sentence, badge, or stage summary. Earlier Reflexive Laboratory papers established two parts of this problem separately: Bell (2026a) argued that AI labels are most useful when treated as structured provenance objects rather than cosmetic disclosures, while Bell (2026b) showed that provenance judgments under a shared human/cyborg/AI (H/C/A) scheme diverge systematically across coders and therefore cannot be treated as neutral descriptors. This paper supplies the missing methods bridge between those results. It verifies the exact seven-stage, 56-task questionnaire preserved in the archived AI Usage Label Generator interface snapshot, formalizes transcript-first coding rules for applying that instrument inside a provenance-rich research environment, and defines a multi-layer provenance object composed of workflow description, evidence anchor, rationale, label, and uncertainty flag. The paper then demonstrates the method on a bounded three-paper pilot drawn from the Reflexive Laboratory corpus (P2, P3, and P9). The pilot does not claim a full-series recoding, universal validation, or author-side ground truth. Its narrower contribution is to show that transcript-anchored, question-level coding is operational, that not-applicable handling is indispensable for conceptual and methodological papers, and that uncertainty clusters—especially in literature-review rows—can be made explicit rather than hidden inside a single paper-level summary label. The broader claim is that AI provenance ranking becomes methodologically stronger when it moves from single-label disclosure toward instrumented, evidence-anchored, comparison-ready coding that can later be transferred to external corpora. Keywords: AI provenance; research disclosure; transcript-first research; provenance coding; H/C/A; workflow transparency; research metadata; Reflexive Laboratory
Peter Bell (Wed,) studied this question.