Provenance of AI-assisted analysis: a disclosure and reproduction protocol for research involving language models

Duplicate deposit — do not cite this record. The same manuscript was deposited twice within two minutes on 24 August 2026 by two parallel automated sessions of the same author. The record of reference is 10.5281/zenodo.22077699; please cite that identifier. The content of the two records is the same manuscript. This note is added rather than the record deleted, because a published Zenodo record cannot be withdrawn by its owner; a request for removal has been recorded on the author’s side. Abstract. Language models entered research practice faster than the rules for disclosing them. The typical declaration today is a single sentence — "a language model was used in preparing this manuscript" — from which a reader can determine nothing: not where the model was applied, not what it did, not whether it touched the result, not whether any of it can be repeated. Alongside this sits a requirement nobody disputes, that human authors bear responsibility and that a model cannot be an author. That requirement is satisfiable, and it does not help a reader check the work. This paper argues that the impasse rests on a conflated notion of reproducibility, and separates it into two: audit reproducibility — that a stated result was in fact obtained in the manner and at the time claimed — and procedural reproducibility — that the procedure can be re-run to yield a substantively comparable result. Bit-identical output is the wrong demand for a generative step and the right demand for a computation; the consequence is a division of labour in which published quantities are never produced by generation. On that basis the paper proposes a disclosure ladder (L0–L5) with working thresholds, a step log whose fields make audit possible, a contribution classification (presentation · search · construction) requiring authors to name the highest class reached, and a rule that rejected steps are retained on equal footing with accepted ones. Limits are stated plainly: provenance does not make a result correct, and the withdrawal of a model version destroys procedural reproducibility while leaving audit reproducibility intact.