Editorial policy

How this content
gets made.

Health content is only as good as the process behind it. This page sets out that process plainly enough to hold us to it.

  1. Retrieval before generation

    A medical claim is produced from retrieved sources, not recalled from a model’s parameters and cited afterwards. Where retrieval finds nothing usable, the system says so rather than filling the gap.

  2. Credibility is a property of the source

    Every passage is graded when it enters the corpus, across six levels. Grading is fixed at ingestion, so it cannot be adjusted to make a particular answer look better supported than it is.

  3. Commercial content is labelled and bounded

    Manufacturer-published material is retrievable and labelled as commercial. It sits at 0.2% of retrievable passages and is never the sole basis for an answer.

  4. Numbers are generated, not written

    Every corpus figure on this site is produced from the database by a script and imported from one file. If a number is not in that file, it does not get published. Current values were derived 2026-09-14.

  5. Review is recorded, not asserted

    Content changes move through pull-request review, so every published claim carries a reviewer, a timestamp and a diff. Guides making clinical claims additionally require a named medical reviewer, enforced at build time.

  6. Corrections are changes, not quiet edits

    Where something is wrong we correct it and update the review date. Guides show when they were last reviewed so you can judge how current they are.

Medical reviewers

No medical reviewer is currently listed, and consequently no content claiming medical review is published. The build enforces this: a guide referencing an unlisted reviewer fails rather than publishing with an unverified byline. The access glossary is published because it defines administrative terminology against cited public definitions and makes no clinical claim.

What the corpus currently holds

7,500+
sources indexed
11,000+
retrievable passages
25+
therapeutic areas

Published figures are rounded down from measured values (7,620 sources, 11,383 passages, 25 areas) so they stay true between regenerations. 0 rows without a usable embedding are excluded throughout — they cannot be retrieved, so counting them would overstate what the system can actually reach. See the knowledge base for the full source mix.