Fact-grounded presentations from a spreadsheet or statement. The model never sees a number.
Language models are good at deciding what is worth saying and at saying it well. They are unreliable at reproducing numbers, and the failure is silent: a fabricated figure reads exactly like a correct one. In an end-to-end generator the document is flattened to text, the model writes slides containing digits, and there is no point at which a figure could be checked against its source.
Deck Studio is arranged differently. Numbers are extracted deterministically into typed facts (value, unit, period, entity, source cell) before the model is invoked. Changes, percentage changes and shares are computed in code with their inputs recorded. The model is shown fact identifiers, entities, periods and directions, never values, and returns sentences with slots. Code maps every slot to a fact, rejects any sentence with a digit outside a slot, a slot used as a name, a change verb attached to a level, or a comparison the numbers do not support. Values are substituted at render time from the fact store, charts are drawn from typed values, and a verification stage refuses to deliver a deck in which any rendered figure fails to resolve. Figure resolution and source coverage are reported together, because either alone is easy to game.
It runs on a small open-weight model on a desktop Mac, which is the practical point: the architecture, not the size of the model, is what makes the numbers trustworthy. Supported inputs are spreadsheets (xlsx, csv), Word and HTML documents, Markdown, plain text and PDFs with tables. Documents without numeric tables are read for sentences containing figures, which become page-sourced quotes. Uploads are kept for the session only.