Three layers, and the screen says which one
A rule is a document about a subject. A supplier works in a code. Nothing in the federal government connects the two, which is why regulatory tracking has never been much use to somebody deciding whether to bid. Here is exactly how the connection is made, and how confident each one is.
who already passes work down in your code from subaward reporting, filed by the prime
Why the layer is on the tag
Three different methods put a code on a rule, and they are not equally strong. A code quoted in the document itself is a fact: the agency wrote it down. A code reached by matching a phrase against the official NAICS title is an inference: the words line up and the subject probably does too. A code an AI suggested from the title and abstract is a guess worth checking.
A screen that renders those three identically is quietly lying about the two weaker ones. So every tag carries its layer, in words, everywhere it appears: on the list, on the rule page, and in the Excel and CSV export, because a spreadsheet that drops the layer turns a guess into a fact the moment it leaves this product.
The three layers, in the order they run
| Layer | What it does | How much to trust it |
|---|---|---|
| 1. Quoted in the document | Six-digit codes written in the rule's own title or abstract are pulled out and checked against the real NAICS table. A number that is not a code is dropped rather than shown. | A fact. The agency named the code. This is the only layer that can be wrong about the subject and right about the number, and that is rare enough to be worth reading when it happens. |
| 2. Matched on a phrase | The document is searched for the official titles of NAICS codes, on whole phrases of twelve characters or more, using the government's own wording rather than a paraphrase of it. | An inference, and a conservative one. Short titles and common words are excluded because "services" would otherwise tag half the federal register. |
| 3. Suggested by AI | Runs only when the first two layers found nothing AND the rule is about buying in the first place. Every code it returns is validated against the real table, so an invented code never reaches the screen. | A guess worth thirty seconds. It is the last layer, it is never the only evidence for a claim on this site, and with AI switched off the first two do the whole job. |
An invented code is dropped, not shown
Both the explicit layer and the AI layer are checked against the 2022 NAICS table before anything is stored. A model asked for codes will happily return a plausible six-digit number that does not exist, and a code page for a code that has never existed is worse than no tag at all: it is a dead end that looks like a finding.
The same check catches a document that happens to contain a six-digit number which is not a code, which is common in a rule that cites a docket number or a public law.
Which agency a rule reaches
A rule that changes the FAR binds every agency, because it changes the regulation itself, so it appears on every agency page. A rule that changes an agency supplement, DFARS being the one most people meet, reaches that agency. Anything else reaches the agencies that published it.
The publishing agency is named the way the Federal Register names it and the agency pages here are keyed the way the award record names them, which are two different spellings of the same department: defense-department against department-of-defense. Comparing them directly returns nothing and looks exactly like an agency that has published no rules, so both are reduced to the same comparison key before they are matched, with a small hand-written table for the cases a key cannot reach, such as the Coast Guard, which publishes in its own name and buys under Homeland Security.
Re-tagging removes
A rule whose text changes is tagged again from scratch, and its old tags are deleted rather than merged with the new ones. Merging can only ever add, so a rule that stops mentioning a code would point at it for ever with nothing saying why.
That matters more than it sounds. Proposed rules are amended, abstracts are rewritten, and a tag that survives its own evidence is the kind of error nobody finds, because the only symptom is a rule on a code page where it no longer belongs.
Where to disagree
Hover any tag on the rules screen and it says which layer made it and what it matched on. If a tag is wrong, that is worth telling us through the contact page with the rule number, because a bad phrase match is usually a bad phrase rather than a one-off and fixing it fixes every rule it touched.
Related
- Rules and fundingWhat the tagging is for.
- Entity resolutionThe same discipline, applied to company names.
- Choosing NAICS codesWhich codes are yours in the first place.
- Commenting on a proposed ruleWhat to do once a rule reaches your code.
See who was last awarded in your NAICS
Every code, every buying office, every vendor and every recompete, organized from the federal record and read nightly. Open it with a plan.