Field note · Healthcare · November 2026
A report that cannot invent a figure
At a medical congress, a report with one wrong figure is worse than no report. We built a pipeline designed to write session reports while the congress is still running, and made it structurally unable to publish a number it cannot trace.
From the room to the text
The pipeline captures the video feed of the room, extracts the audio and transcribes it. A glossary, loaded from a file and from the slides, corrects technical vocabulary only: molecules, trial names, acronyms. Nothing else is rephrased.
Facts first, prose second
Writing happens in three separate passes: extract, enrich, structure. The first pass produces facts, each with its source: a timestamp in the transcript, a slide, or both. Any context the pipeline adds sits in a separate field and is never merged into the fact itself.
The editor is an AI, the checker is not
A writer agent and a reviewer agent then work like an editorial committee. The reviewer criticises without rewriting; a revision pass corrects only what was flagged, with two rounds at most to stay within the hour.
Behind them sits a check with no AI in it: every figure in the final text must exist in a fact cited by its paragraph, and every cited fact must exist. A revision that breaks this rule is thrown away, however well it reads.
References that exist
Citations at the bottom of slides are read and matched against PubMed, then Crossref. A candidate is accepted only if the year matches and the first author's name appears in the record. When there is no reliable match, the report says so and cites the slide. It never produces a plausible-looking reference.
What we take from it
Generative AI writes well and guarantees poorly. So the guarantee goes to plain code: the model writes, a deterministic check decides what may be published. The same split applies to financial reporting, regulatory documents and any text where a wrong number has a cost.