Cite the paragraph, not the document title

Most regulatory arguments fail in the same place. Someone quotes a rule. Someone else asks where. The first person names a regulation, or a guidance page, or a memo from last year. None of those are a location. A location is the instrument, the version, the annex or schedule, and the paragraph or row that actually says the thing.

If you cannot hand that location to a colleague and have them land on the same sentence, the cite is not done. It will not survive a challenge from legal, from an auditor, or from a competent authority with the text open.

What a usable cite contains

Four pieces, in order. Miss one and the next reader has to reconstruct your search.

• The instrument: the regulation, act, decree, judgement or register entry, with its official identifier.
• The version: consolidated text, OJ citation, amendment date, or the register snapshot you actually read. Dates matter more than people admit.
• The internal address: article, annex, appendix, schedule, table, row. 'Annex VI' is not enough if the annex is a thousand rows long.
• The claim in the answer, written so that it does not outrun the sentence you pointed at.

Secondary sources can be useful as a map. They are not the destination. A consultant note, a law-firm briefing, a Wikipedia paragraph and a model answer all share the same failure mode: they summarise. Summaries drop conditions, carve-outs, date of application and the annoying sentence two lines down that reverses the first one.

Annexes are where the work lives

Operative articles often say almost nothing on their own. They point. They say that a list is maintained, that a classification is the one in the annex, that a duty applies to the substances in the table. If your answer cites Article 3 and the substance is only named in Annex II, you cited the wrong door.

The same is true of judgements. A court may spend ninety paragraphs on procedure and then decide the case in four. Those four are the cite. Reciting the name of the case is how people pretend they have read it.

Version control is not optional

Regulatory text moves. Harmonised entries get annulled. Transitional provisions expire. A national register replaces a PDF with a database and changes the identifier scheme. If you cite 'the CLP Regulation' in 2026 without the amendment package or the judgement that removed an entry, you are citing a family of documents, not a fact.

A good habit: write the date you read the text next to the cite, even when the instrument itself carries a date. Registers in particular are living objects. The row you saw on Tuesday may not be the row a reviewer sees in November.

What to do when you cannot find the paragraph

Say so. 'Not in the indexed text' is an answer. Inventing a plausible article number is not. The pressure to fill silence is the thing that produces confident nonsense, and it is strongest when the question is simple: is it listed, is it still listed, who notifies, by when.

If the duty sits in a related instrument, name that instrument and the join. 'The restriction is in Regulation A, Annex XVII, entry 23; the classification that triggers it is in Regulation B, Annex VI.' Two cites, one chain. That is still research. Pretending they are one document is how chains get lost.

A test you can run on any answer

Give the cites, and only the cites, to someone who did not write the answer. Can they open the same paragraph in under two minutes? If they cannot, the answer is not ready to leave your desk. That test is stricter than 'does this sound right', and it is the only one that still works when the reader is unsympathetic.

RegAffairs AI is built around that test. Claims are pinned to the paragraph. The research trail is on the page, so you can see the search that led there. If the record is silent, the answer says so.