Executable Research Objects: Reproducibility, Canonicality, and State Suf- ficiency in AI-Assisted Research

AI-assisted research increasingly preserves transcripts, repositories, packages, and visible process traces, but preserved process does not by itself determine the scientific object on which later work should rely. This paper proposes the executable research object as the minimal governed unit of cumulative AI-assisted science. An executable research object binds a claim, an explicit falsifier, a replay path, declared outputs, validation conditions, authority state, and archival binding so that the claim can be rerun, checked, and situated within visible governance state. The paper’s intervention is therefore not another call for better documentation. It is a shift in what counts as the relevant object of scientific accumulation. Archives, persistent memory layers, repositories, and transparent traces each preserve something valuable, but none of them alone fixes a replayable and authoritative claim unit. The paper defines the executable research object formally, distinguishes it from these adjacent preserved forms, and argues that canonicality should attach only through enactment: replay confirmation, validation, and governance-visible authority transition. The institutional consequence is direct. Journals, repositories, laboratories, and AI-assisted research systems should increasingly organize themselves around replayable, validated, enacted claim objects rather than around preserved process or file bundles alone. Keywords: AI-assisted research; executable research objects; reproducibility; canonicality; state sufficiency; transcript-first research; research infrastructure; replayability; artifact validation