The complete set of working documents of a project that develops, and simultaneously documents, a long-term collaboration between a human author and a large language model. The project is preparing a manuscript about that working architecture; its empirical basis is a second, independent project, a literature database in a medical domain, which serves as the case under study. The deposit is a provenance record rather than a dataset in the usual sense: it contains no measurements to reanalyse, but the complete version history of every governing document, every decision, and every recorded error. 170 files covering seven continuously maintained artefacts, 45 versions of a structured interview protocol, and two SHA-256 manifests. Files are restricted because some documents name a living person who has not consented to publication. Access is granted on request. The accompanying README.md describes the document classes, the citation convention, and the known limitations of the archive, including two missing files and the limits of what the checksums establish.