Automated Scientific Narrative Generation Through Computational Provenance and Dynamic Authoring Frameworks

Abstract. Despite the growing quantity of born-digital research, scholarly articlesremain tethered to flattened PDFs—unable to expose or verify the computations theysummarize. We present a publication container that couples provenance-generatingeScience infrastructures with Dynamic Authoring Frameworks (DAFs). Using this method,an article is modelled as a set of symbolic operations over provenance. The result isa procedural narrative whose claims are anchored to deterministic combinations ofoperations and provenance, enabling verification. To enhance this method we place LLMinference tasks inside the DAF for small, constrained tasks. This mitigates hallucination,supports granular attribution, and enables researchers to move “warrant” making awayfrom LLM inference, to the explicit operations within the DAF. Using a remote-sensingcase study (CoastSat), we show how this method can produce consistent, accurate,and generative methodological descriptions from provenance. We argue that for modernscholarly communication to support generative text, it must move beyond “flat” scholarlyarticles towards more formal representations of authorship.