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.
By Connor Griggs — Regulatory & Quality Strategist
Every regulatory intelligence tool built since 2023 says some version of “AI-powered, human-reviewed.” The phrase is doing no work. It describes a workflow the vendor can quietly stop following on a busy week, and you would never know. So instead of asserting our review rule, this post explains how it is enforced — because “we promise” and “the database refuses” are different grades of claim.
The failure mode we are paranoid about
A generated recommended action that is plausible, fluent, and wrong is worse than no recommendation at all — because this audience acts. The reader of an FDA Radar item is not browsing; they are deciding whether to open a CAPA, review a risk file, or call their consultant. The failure we design against is not embarrassment. It is a professional acting on confident, wrong advice with our name on it.
What the constraint actually enforces
The language model in our pipeline drafts summaries. That is the entirety of its authority. An item cannot reach published state without three fields the model has no path to supply: a human-assigned urgency, a human-written recommended action, and the reviewer’s name. The rule lives in the database as a constraint, not in a process document — the publish simply fails without them. On a busy week the queue backs up; nothing unreviewed leaks past it. We consider that trade obviously correct, and we notice competitors making it the other way.
A monitoring product is a chain of claims. Ours ends with a person’s name on every link.
Two supporting habits
The raw archive is immutable. Every FDA document is stored as fetched and never edited. FDA does silently revise guidance pages; against an immutable archive, a silent revision becomes a detectable diff instead of a private memory. When the source changes under an item we published, that is a fact we can show you, dated.
Corrections are appended, never disappeared. When we get something wrong, the correction renders on the item’s permalink, above the fold, dated, next to the original claim. An intelligence product that silently rewrites its record is asking you to trust its present tense. We would rather show you our error rate and let the record argue for itself.
Why we publish the mechanism
Partly because this audience checks small things for a living, and deserves vendors who expect to be checked. And partly as a commitment device: having told you the constraint exists, we cannot remove it without the removal being a lie. The review rule costs us speed on exactly the weeks speed would be easiest to sell. That cost is the product.
Primary sources
Regulatory intelligence, not regulatory advice. This post describes method and published FDA records as of its date; decisions about a specific device belong with your regulatory professional.