Signals

Operator notes.

How we watch FDA, and how to act on what the watching returns — the method behind the product, written down. Regulatory intelligence and practice notes, never regulatory advice.

Method·Aug 1, 2026 · 3 min

A Class I device, a Class I recall

FDA assigns a device its regulatory class once, as a permanent attribute of the device type. FDA assigns a recall its class per event, as a snapshot of one defect's severity. Both scales use the identical three labels, and a device's own class predicts nothing about the class of a recall against it.

Read the post →
Practice·Aug 1, 2026 · 4 min

21 CFR 820 didn't move. Its contents did.

FDA's Quality Management System Regulation, effective February 2, 2026, left the address unchanged and emptied most of what used to live there. A citation to Part 820 written last year and one written this year can point at the identical number and mean two different things.

Read the post →
Method·Jul 31, 2026 · 4 min

The product code that doesn't exist yet

FDA's classification database has no row for a device type until a De Novo request is granted — the product code, the regulation number, even the classification itself are created by the grant order, not looked up before it. A watch keyed to "your product code" cannot fire on a category that is still being invented.

Read the post →
Practice·Jul 31, 2026 · 4 min

The classification posts. The 483 behind it doesn't.

FDA scores every closed device inspection No Action, Voluntary Action, or Official Action Indicated — a free, weekly, non-FOIA database. The Form 483 narrative behind that score is a different document, gated behind a records request FDA answers on its own schedule. A firm watch that conflates the two is missing the earlier, public half of the signal.

Read the post →
Method·Jul 30, 2026 · 3 min

A recall has three dates, and the pipeline had to pick one

A device recall record carries three separate date fields — when the firm told the public, when FDA finished judging the hazard, and when the report posted — and they can be months apart. Read the wrong one as "when this happened" and a portfolio timeline is quietly off by however long the classification took.

Read the post →
Practice·Jul 30, 2026 · 3 min

The count is real. The rate is not.

A MAUDE search returns a number that looks like a fact you can compare — this many adverse event reports, this device, this window. FDA's own documentation says that number cannot become a rate, because nothing in the reporting system tells you how often the device was used.

Read the post →
Method·Jul 29, 2026 · 4 min

The firm on the record is not the firm on the box

A device label is allowed by rule to carry a brand's name while a different company actually made the thing. FDA's enforcement records follow the maker, not the brand — so a firm watch keyed only to your own name has a blind spot built into it by design.

Read the post →
Practice·Jul 29, 2026 · 4 min

Low confidence is an instruction, not a hedge

Our classifier can say it doesn't know. That single field changes what happens to an item next — and it's a habit worth borrowing the day you're reading a draft this product, or anyone else's AI, handed you.

Read the post →
Method·Jul 28, 2026 · 3 min

The same company, spelled three ways

FDA's own filings can't agree on how to spell a firm's name — a comma here, no comma there, Inc. versus INC. FDA Radar's matcher corrects for that, then stops on purpose, because the one join in this product that names a company is the one it refuses to make clever.

Read the post →
Practice·Jul 28, 2026 · 3 min

A device that was never a medical device

FDA published a Federal Register notice with the word Device, capitalized, in its own title. It has nothing to do with CDRH. Word-boundary matching fixes accidental substrings — it does nothing for a real word doing its ordinary job in the wrong industry.

Read the post →
Method·Jul 27, 2026 · 3 min

FDA's warning letters, addressed by column number

The data behind fda.gov's warning-letter search has no field names — company, date, and subject are positions in an array, and nothing FDA publishes promises they'll stay in order. What a monitoring pipeline does about a contract nobody signed.

Read the post →
Practice·Jul 27, 2026 · 3 min

Your regulation has a decimal. FDA's watch doesn't.

A device classification regulation reads down to the section — 21 CFR 890.5740, powered heating pad. The Federal Register indexes rule changes only down to the part — 21 CFR 890. Watch at the wrong granularity and a real change to your regulation never matches.

Read the post →
Method·Jul 26, 2026 · 3 min

Three letters is too short to search for

FDA Radar's own Federal Register filter once matched 'udi' inside 'studies' and nearly filed a drug guidance about antibodies as a device document. The failure mode behind it, and why the fix is a stricter filter, not a smarter one.

Read the post →
Practice·Jul 26, 2026 · 3 min

Most warning letters never close

FDA issues a close-out letter only after it has verified — usually by reinspection — that a firm's corrective actions actually worked. Violations that can't be corrected, and firms that no longer exist, get none. What the absence of one does and doesn't tell you.

Read the post →
Method·Jul 25, 2026 · 3 min

The guidance that skipped the draft

FDA's Good Guidance Practices rule lets a significant guidance document skip the draft entirely — issued final, in effect, on the day it posts. The record's title doesn't tell you which path a document took to get there.

Read the post →
Practice·Jul 25, 2026 · 3 min

Ongoing, as of when?

openFDA's recall status field reads like a live indicator — Ongoing, Completed, Terminated. FDA's own documentation says it isn't one. What gets corrected on a recall record after it posts, what doesn't, and the practice for telling the difference.

Read the post →
Method·Jul 24, 2026 · 3 min

The least interesting fact in a 510(k)

Cleared or not is a one-bit answer. The record underneath encodes the pathway, the conditions, and a disclosure choice the applicant locked in before submitting — and those are the fields worth reading across a product code.

Read the post →
Practice·Jul 24, 2026 · 3 min

No recall arrives with a product code attached

openFDA's recall data has a field for it. Across the last 300 device recalls we ingested, that field came back empty every time. The practice for finding your portfolio in a feed that isn't indexed by the thing you'd search for.

Read the post →
Practice·Jul 23, 2026 · 4 min

The deadline that doesn't email you

A Federal Register comment period can run ten days or nine months, and nothing on the page warns you which. The field that has the real answer, and the habit that checks it every week.

Read the post →
Method·Jul 23, 2026 · 4 min

The warning letter has two dates

One starts your compliance clock. The other is what the public sees, and it can lag by weeks. Confuse them and the subject line's real signal goes unread too.

Read the post →
Method·Jul 23, 2026 · 8 min

How to monitor FDA without drowning

Four official sources cover the material risk to a device portfolio. The hard part is not finding them — it is surviving their volume.

Read the post →
Practice·Jul 23, 2026 · 9 min

Your predicate was recalled. Now what?

A recall on the device your clearance compares against is not automatically your problem — and is never automatically not your problem. The assessment that tells you which.

Read the post →
Method·Jul 23, 2026 · 6 min

Why no item reaches you without a human

The model drafts. It is structurally unable to publish. The database constraint that makes our review rule a fact rather than a promise — and why this audience should demand one.

Read the post →

The posts explain the method. The product runs it.

Enter your products once; FDA Radar watches the same four sources daily and joins every item against your portfolio — reviewed by a person before it reaches you.

Open portfolioHow it works