Changelog

What shipped, dated.

A product that monitors changes for a living should publish its own. Every entry below corresponds to code that was live on that date.

2026-07-16

The portfolio release — tell us what you make, we watch FDA for it

  • Product profiles from one identifier: type a product code or K-number and FDA's classification database fills in the class, regulation, pathway, panel, and device flags — confirmed by you, never silently saved.
  • Identifier matching: every ingested item joins against your portfolio on product codes, CFR citations, K-numbers, and firm names. The "why this affects you" on every match is the join key itself, shown verbatim.
  • The recall gap, closed honestly: FDA publishes no identifiers on recalls, so our reviewer supplies the product code while reviewing — disclosed on the match as reviewer-keyed, never passed off as FDA's.
  • Accounts and the portfolio dashboard: passwordless sign-in, a since-your-last-visit view grouped by urgency, per-product pages with the derived profile and the clearance corpus in your product code (labelled unreviewed).
  • Action tracking: recommended actions carry an open/done state per product — the same guidance stays open on your other products until you close it there too.
  • Per-subscriber digests, including the quiet one: "We checked N FDA documents. None touched your products" — with the real count.
  • Item permalinks with the corrections record rendered above the fold, dated, permanent.
  • Row-level security enabled across the entire database.
2026-07-15

First light — the pipeline and the review discipline

  • Four sources live: Federal Register, CDRH warning letters, device recall enforcement, and 510(k) clearances — all official APIs, no scraping.
  • Immutable raw archive: every document is stored as fetched and never edited, so an FDA page that silently changes becomes detectable — and FDA does silently revise guidance.
  • The review rule, enforced by the database: no item can publish without a human-assigned urgency, a human-written recommended action, and the reviewer's name. The model drafts summaries; it is structurally unable to publish advice.
  • Volume discipline: routine Class II/III recalls and the 510(k) firehose route to the searchable corpus instead of the review queue, keeping the queue at roughly eighteen material items a week — small enough that a person genuinely reviews all of it.
  • The digest renderer, including the quiet-week path, with every colour contrast measured rather than assumed.