The last piece argued that most of an investigation is retrieval, and that an agent should be doing it. The obvious next question is what you get back.
You get a case note, and it is worth being precise about how that differs from a summary, because the difference is the whole design.
A summary asks to be trusted. A case note asks to be checked. Anything that writes prose can summarise a case. The harder problem is producing something you can disagree with productively at three on a Thursday afternoon, without having to start again from the beginning.
So here is what is in one, and why.
Findings, with their sources attached
The note is built from findings, and each finding is linked to the sources it came from. A finding might rest on several sources, and one source might support several findings, so the link is explicit rather than implied by where things sit on the page.
The reason is practical. If you doubt one finding, you can go straight to the record behind it and settle the question. Without that, doubting one line means re-checking all of them, and a note you have to re-verify in full has not saved you anything.
It also changes what disagreement looks like. Instead of "I am not sure that is right", you get "that screening result is from March and the counterparty re-registered in June". The second one is a conversation. The first is a stalemate.
And if a finding cannot point at a source, it says so. That one is worth dwelling on for a moment.
Gaps, and the difference between two kinds of missing
Often the most useful thing in the note is what is not there. But there are two very different kinds of missing, and they support completely different decisions.
The first is that we looked and there was nothing. An address with no transaction history is a finding, and a useful one.
The second is that we never looked. A provider was down, a lookup failed, or no source was configured for that check.
In a document that reports both as silence, those two are indistinguishable, and that is the failure that actually matters. "No on-chain activity" and "we did not check" are not the same sentence. The retrieval layer records which of the two happened, per source, and the note carries that through rather than flattening it.
There is a related rule the agents work under. If an identifier arrives already shortened, it gets reproduced exactly as given, never expanded or reconstructed. A wrong wallet address in a compliance record is worse than an incomplete one.
The same instinct runs through the whole thing. An agent that quietly fills gaps with plausible reasoning is worse than no agent, because now the hole is invisible.
The recommendation, and why it says it is one
The note does reach a conclusion. It proposes a disposition, explains the reasoning, and names the findings that reasoning rests on.
That surprises people who expect a product built around human judgement to stop short of having an opinion, so it is worth saying what the recommendation is and what it is not.
It is an advisory. That is the word used in the code, not something added later for a marketing page. There is one current advisory per case, it refreshes when the evidence changes, and it carries no authority of its own.
Your decision ends it. Once you have made a call, the system stops recommending on that case. A fresh opinion after the fact is treated as noise, not as a second view worth raising. Your decision is final in a way the recommendation never is.
It forms its own view rather than agreeing with the last one. Where a human decision already exists as context, the agent assesses independently instead of deferring to it. An advisory that quietly reproduces whatever was decided last time is worth nothing.
The recommendation also carries the version of the rubric it was produced under, and the note carries the version of the template it was rendered from. Two years later you can establish not just what was recommended, but which encoded standard was in force at the time. That is the difference between a record and an anecdote.
What the recommendation will not do is close anything. It is an input to your decision, and the reasons that boundary sits exactly there are in Where the agent stops.
It is longer than a summary
That is the trade, and it is worth naming.
We could hand you three sentences and a disposition. It would be quicker to read and worth less, because you would either accept it without grounds or rebuild it from scratch. The note is sized to be decided on rather than skimmed, and the time it saves is the two hours of retrieval that would otherwise have come first.
The short version
Findings with their sources attached, and the uncited ones flagged. Gaps named, with "nothing found" kept separate from "not checked". A recommendation that tells you it is only a recommendation and stops the moment you decide.
Everything you need to make the call, and nothing that makes it for you.