When the record is silent

Silence is a result, and the one people are worst at writing down. A question comes in, the register does not list the substance, the annex has no row, the guidance stops one case short of yours. The accurate sentence is 'not in the text we have'. The tempting one reconstructs what the authority probably meant.

That reconstruction is where organisations get hurt. Not in the flashy hallucination with a fake OJ number. In the quiet backfill: a deadline inferred from a neighbouring article, a duty-holder borrowed from a different instrument, a 'generally accepted' reading with no paragraph behind it.

Three kinds of silence

• The instrument is in the corpus and the row is not there. That is a finding. Write it, cite the annex you searched, stop.
• The instrument is not in the corpus at all: a coverage gap. Write it. Do not answer from a neighbouring market as if it were the same law.
• The instrument is there and the text is ambiguous. That is a judgement call. Flag the ambiguity, quote both readings, and hand it to a human.

Mixing those three is how a research note becomes a liability. A coverage gap dressed up as a negative finding will be quoted later as 'we checked, it is not listed'. An ambiguity dressed up as a clean no will be quoted in the other direction the day an authority takes the other reading.

Negative findings need cites too

'Not listed' is a claim about a list. It needs the list, the version, and the identifier you searched with. If you searched a name and the list is keyed by a number, you have not searched it. If you searched an old consolidation, you have not searched it. The trail should show the query, not just the conclusion.

This is also why a general assistant is the wrong tool for a negative finding. It can say 'I don't see it' because it does not see much. That is not the same as having opened Annex VI and walked the rows.

Do not launder judgement as research

Some questions are not research questions. 'Should we notify' when the text gives a test that depends on facts you do not have is not a research question. 'Is this the same product as the one in the register' is often a facts question. The honest output is the test, the missing facts, and a stop.

Consultants exist for the judgement that remains. The research product should make their job smaller, not impersonate them. If an answer reads like advice, it has gone past the record.

What we do with silence

RegAffairs AI answers from indexed official text. Where the record is silent, the answer says so. The research trail stays on the page, including the searches that returned nothing. We will not invent a paragraph to keep the prose moving. That is a design choice, not a slogan.