Filtered. Not discarded.

Open FDA Radar’s review queue on an ordinary week and it holds a handful of items — a Class I recall, a warning letter, a guidance notice. Query the underlying corpus the same week and it holds hundreds. Both numbers are correct. They are not measuring the same thing.

One field decides the boundary

Every item FDA Radar ingests gets a review state of either pending (it reaches a human) or filtered (it does not, though it stays searchable and feeds the corpus a watchlist queries later). For most sources the answer is always pending — a warning letter or a Federal Register notice is rare enough that every one earns review. Device recalls are the exception, and the rule there is narrow: a recall queues only if openFDA’s own classification field reads Class I. Class II, Class III, and recalls FDA has not yet classified at all stay in the corpus, filtered.

The volume that forced the rule

The classes are not a stylistic choice. Under FDA’s own definition, a Class I recall is one where there is a reasonable probability that use of, or exposure to, the product will cause serious adverse health consequences or death; Class II covers temporary or medically reversible harm, or a remote probability of anything worse; Class III covers harm that is not likely at all. The three labels sit on the same enforcement record, and nothing about the record separates a single Class I event from the ordinary Class II volume around it — a team has to read the field.

A 30-day measurement taken during development put a number on that volume: 327 items across every source, and recalls alone accounted for 300 of them — 250 Class II, 50 Class I. Add in the roughly 250 510(k) clearances FDA publishes in an average month, almost all irrelevant to any one company, and an unfiltered queue approaches 500 items a month against a review model costed for about 18 a week. A queue built to be read cannot also be a queue that repeats everything FDA published; something has to be corpus instead of alert, and class is the field that tells the two apart.

Filtered is not a euphemism for discarded. A Class II recall a reviewer never saw is still in the database, still tied to its product code, still there the day a watchlist or a search asks for it — it was simply never pushed at anyone.

The practice

Reading FDA Radar’s digest as a complete recall feed is the mistake the classification field exists to prevent. The digest is a triage decision, not a transcript: it surfaces what FDA itself reserves for a reasonable probability of serious injury or death, on the assumption that a Class II correction to a labeling issue does not need the same day’s attention that a Class I recall on a device cited as predicate does. Whether a specific Class II event still deserves a closer look is a judgment for whoever owns that product code — that reading, like every reading in Signals, is regulatory intelligence and method, never regulatory advice about a recall of your own. The corpus is built for exactly that judgment: search it, or let a watchlist ask it the narrower question the daily queue was never meant to answer.

FDA Radar states the boundary on the page that lists what we monitor and how often rather than leaving it to be discovered from a recall that never arrived. It is the same principle editorial standards states directly: being comprehensive is FDA’s job, and being clear about what does not reach a human is ours.