This paper describes a governed AI writing system for health-AI research teams that operates through six layers and combines protocol-based writing, issue-based task tracking, citation-audit checkpoints and researcher review gates before submission. It is a retrospective description of two manuscript-production cases, and the unit of analysis is the writing workflow rather than model performance.
Case 1, a medical billing audit manuscript, went through 22 versions across 12 dedicated writing tasks with researcher corrections in every section. Case 2, a clinical AutoML methodology manuscript, reached submission in 12 versions, reusing what was learned in Case 1 and including a technical review by the engineering team at version 10. A structured self-review found nine system gaps, and eight were corrected within the same development cycle. The work is a preprint not yet peer reviewed, and the authors call for prospective studies of manuscript quality and adoption.
