Organising Accountable Scholarly Work Across AI Systems: The AIMan and JOAK Cases

AI systems can contribute to research, review, and writing, yet a scientific organisation must still know who or what performed each task, which evidence was used, what a review covered, and who accepted the result. This exploratory case study examines how those responsibilities can persist across changes in AI systems and document form. AIMan models the organisation explicitly: persistent Personas carry attribution across replaceable technical Instances; Roles define responsibility; versioned Messages, Decisions, and Artifacts form recoverable memory. A Human Sponsor retains scholarly and release authority. The Journal of Artificial Knowledge (JOAK) extends the design through review, editorial judgment, and publication. Two connected cases provide evidence. In the first, ChatGPT and Codex Instances completed a repository-mediated handover and contributed to a source-grounded scholarly study without shared conversational memory. In the second, JOAK used version-bound AI review, Sponsor decisions, a computational supplement, and checked article production to publish Volume 1, Issue 1. The contribution is an operational account of accountable scholarly work across the research-to-publication cycle: organisational identity, source state, criticism, acceptance, and public evidence remain distinguishable as technical Instances and document forms change. JOAK makes the journal's distinctions among a submitted version, a review recommendation, editorial acceptance, and public release explicit in AI-supported work. The cases show recoverable coordination in one scholarly cycle and one published issue. They do not establish superior scientific quality, independent human peer review, provider-neutral interoperability, or sustainable journal operation. These remain measurable questions for subsequent work.