The same company, spelled three ways

Three FDA records, three ways of writing the same company: ZIIP, Inc. on one filing, ZIIP INC on the next, Ziip Inc on a third. None of the three is wrong. FDA does not require a firm to spell its own name identically across a warning letter, a recall report, and a 510(k) applicant field, and firms routinely don’t. A monitoring product that wants to tell a subscriber “this recall names your manufacturing partner” has to close that gap itself — and exactly how far it goes to close it is a real design decision, not a rounding error.

What gets stripped, and what doesn’t

The normalization FDA Radar applies before comparing two firm names is short and specific: lowercase the string, strip punctuation — periods, commas, quotation marks, parentheses, ampersands — and remove a fixed list of legal-entity suffixes as whole words: inc, incorporated, corp, corporation, co, company, llc, llp, ltd, limited, gmbh, plc, sa, ag, kg, bv, srl, pty, pvt. Run ZIIP, Inc., ZIIP INC and Ziip Inc through that and all three collapse to the same string. That is the entire transformation. Nothing phonetic, no edit-distance score, no similarity threshold to tune.

Where it stops, on purpose

A wider net was available and wasn’t taken. The matcher’s own code comment is explicit about why: a fuzzy match on a firm name, in this context, is not a convenience feature. It is an accusation. If “Zip Innovations” and “ZIIP Inc.” score close enough on a similarity threshold, a looser matcher would tell a ZIIP subscriber that a warning letter issued to an unrelated company was about their own manufacturing partner — a specific, wrong sentence, not a vague miss.

A false positive in a keyword filter costs someone three seconds discarding an irrelevant document. A false positive in a firm match names the wrong company to a person deciding whether to act.

The other filter in this codebase makes the opposite bet

FDA Radar’s Federal Register adapter runs a device-relevance filter with the stated bias pointed the other way entirely: toward catching too much, not too little. The reasoning, in the code’s own words, is that a false positive there costs a reviewer a few seconds discarding an irrelevant notice, while a false negative means a real device guidance never reaches anyone at all. That trade is correct for that filter, precisely because every match it produces still passes in front of a person before it reaches a subscriber. It is the wrong trade for a firm match, because a firm match doesn’t route through a queue asking “is this relevant” — it routes straight into a sentence asserting whose company a record names.

The line the sentence has to hold

That sentence is not generated prose. The sentence a subscriber reads — “Names your firm — ZIIP INC” — is the literal join key, rendered back verbatim, on a separate code path built for exactly one reason: it cannot hallucinate a match it is never asked to compose. Loosen the join upstream and that guarantee doesn’t break loudly. It breaks quietly, at the one place a wrong sentence lands with a specific company’s name already attached to it.

None of this is a claim that a name match settles what a given warning letter or recall actually means for the firm it names — that reading is still work for a person, on the file, every time. It is regulatory intelligence and method: a description of what “your firm was named” is allowed to mean before FDA Radar will say it. Per what we monitor and how often, the join runs on every item against every portfolio profile; what changes is only ever the confidence required before a basis for that join is allowed to exist at all.