Ask with citations: every claim, traceable

The real problem with LLM answers is not that they are wrong. It is that you cannot check whether they are wrong.

From plausible to findable

Ask inside LLM-notebook and every claim carries a marker: [1], [2]. Click it and you jump back to the exact paragraph, highlighted.

That changes two things:

1. You verify in five seconds instead of re-reading for five minutes.
2. You confirm a claim was not taken out of context before it goes into a deck.

How you ask decides whether the answer holds

Same material, different question, very different verifiability:

• ❌ "What's wrong with this contract?" — too broad, so the answer stays vague.
• ✅ "Do clauses 4 and 9 conflict on liability? Quote the text." — checkable in one click.
• ✅ "Copy the paragraphs that support this, with page numbers." — forces the model back to the source.

A useful rule: ask it to quote, not to summarize. A summary cannot be verified. A quote can.

Citations are a system property, not a prompt trick

Plenty of tools claim citations. The difference is whether the citation is generated by the model or resolved by the system.

We do the second: parsing keeps paragraphs and page numbers, and answering binds markers to real paragraphs. A marker either jumps somewhere real, or it does not appear at all.

Citations are not there to make the answer look good. They are there to make it safe to use.