What a research trail is for

Regulatory work is reconstructable or it is theatre. A memo that does not show how it was built cannot be checked, only believed. Belief is a poor control when the cost of being wrong is a missed notification, a wrong classification, or a statement to an authority you cannot later defend.

A research trail is the method, written down as it happened. Which corpus was searched. Which documents were opened. Which paragraphs were kept. Which were discarded, and why. It is not an audit-log export or a performance. It lets a second person stand in the first person's shoes.

The trail is for the next reader, not for you

You already know what you did. The lawyer who will take this to a client does not. The colleague covering your leave does not. The future version of you, six months later, looking at a sentence you no longer remember writing, does not. The trail is a handoff.

That is why it has to sit with the answer. A method parked in another tool, or in a private notebook, or in a chat you cannot reopen, might as well not exist. When the question comes back, 'I think we checked this' is not a trail.

What belongs on it

• The question as asked, not the cleaned-up version you wish you had typed.
• The searches run, including the ones that returned nothing. Empty results are evidence.
• The documents actually read, with enough of an identifier that someone else can open them.
• The cites that made it into the answer, tied to the sentence they support.
• The gaps. If a jurisdiction was not in the corpus, say that. If an amendment was not yet consolidated, say that.

What does not belong: a vibe about confidence, a score, a claim that the model is sure. Certainty is not a method. If the text is clear, the cite will carry it. If the text is not clear, no amount of confident tone will make it so.

Why summaries erase the trail

A general assistant will happily give you a fluent paragraph and a list of sources that look official. Sometimes those sources exist. Sometimes they do not. Sometimes they exist but do not say what the paragraph claims. You cannot tell which of the three you have been handed unless you retrace the work, at which point you have done the work twice.

The failure is structural. A system that cannot show its steps cannot back work that a hostile reader will pick apart. And regulatory answers get hostile readers more often than people admit.

Shown, not exported

There is a temptation to treat the trail as a file you download and file. That is the wrong object. The useful trail is the one a reader can expand next to the sentence they are about to quote. If they have to leave the page, they will not. If they will not, the trail does not do its job.

RegAffairs AI keeps the trail on the answer. Every search and every document the agent read is visible there. We do not offer an exported audit package, because the point is not a zip file. The point is that you can see the method without leaving the claim.