Turning model structure into prose without inventing findings.
Elizabeth Sramek
Published 8 September 2026
Structure-first: fit the model, decompose it, map each component to a templated sentence with its estimated quantities. Prompt-first: hand a chart and some summary statistics to a language model and ask for commentary. The first is auditable. The second reads better and can assert things the model never found.
An executive summary of the discovered structure. Each component with its estimate and interpretation. Charts of fit and forecast with intervals. A model-criticism section naming where the model and data disagree. Methodology and dataset provenance.
Smoothing templated sentences into readable prose, and adapting register for the audience — provided the underlying claims come from the fit. Used that way the model is a copy editor, not an analyst.
A generated paragraph that says “revenue growth accelerated in Q3” when the fitted model contains no changepoint and no interaction term. It is fluent, plausible, and unsupported by anything that was computed.