What a citation actually proves

The last item on our seven limits reads: a citation proves where a sentence came from, not that it says the same thing. People skim past it, because it sounds like the kind of hedge every vendor writes. It is not a hedge. It is the precise boundary between two guarantees that look identical from the outside, and only one of them is absolute.

What does a Nemosyne citation guarantee?

It guarantees the passage exists, that it is a real span of a real document your firm connected, and that the engine itself put that passage in front of the model before the sentence was written. That last clause is the whole mechanism. The model is not asked to remember a source or to produce one afterwards; it is handed the evidence and the citation is a pointer to what it was handed.

The consequence is worth stating plainly: a reference to something that does not exist cannot survive. There is no path by which a plausible-looking document name is generated alongside a claim, because the name is not generated at all — it is carried. This is the failure mode people mean when they say an assistant “made up a citation”, and it is structurally absent here rather than trained out.

It also means a citation resolves. Every agent-facing answer ships a list of the source items it drew on, because an agent cannot name a document it was never told about. You can open the thing. That is a low bar and most systems do not clear it.

What does it not guarantee?

Support. A passage can be real, correctly attributed, genuinely relevant, and still weaker than the sentence built on top of it.

Take a concrete case. An investment committee memo contains the phrase “customer concentration remains elevated at 38%”. A fact derived from it might read: concentration was named as the principal risk. The passage is real, the figure is right, the topic is right — and “the principal risk” is an inference the memo may or may not support. It might have listed four risks and put concentration third. Provenance is intact; support is an open question.

A citation tells you where to look. It does not tell you that looking will agree with you.

This gap is not unique to us and it is not closed by any amount of prompt engineering. It is the residual of turning prose into typed claims at all. What differs between systems is whether the gap is named.

Why publish this rather than imply the stronger claim?

Because the stronger claim is the one a reader will assume by default, and an assumption you benefit from silently is a debt. A buyer who believes citation means verification will eventually find a sentence that overreaches its passage, and at that moment the question stops being “is this claim right” and becomes “what else did they let me assume”. That second question is expensive, and no answer given afterwards is as cheap as having said it first.

There is a version of this page that says “every answer is fully verified and traceable to source”. It would convert better. It would also be false in a way nobody could correct once it was believed, and the buyers this is built for — the ones in audit, credit and legal work — are professionally trained to find exactly that kind of gap.

What should a reader do with a citation?

Treat it as an invitation, not a receipt. The value is that checking is now cheap: instead of asking whether a claim is plausible, you open the passage and read it. A fifteen-second check is a different thing from a thirty-minute reconstruction, and the whole point of holding provenance is to make the check affordable enough that people actually do it.

The stakes decide the threshold. A background figure in a portfolio review probably does not need opening. A number going into a committee paper does. The system does not make that judgment for you, and it is not trying to — it makes the judgment possible.

What does the agent receive?

Not just a claim with a footnote. Every agent-facing slice ships a sidecar naming the source items it drew on — the file, the record, or the statement, named the way a person would recognise it rather than by an internal identifier. The reason is blunt: an agent cannot cite a document it was never told the name of, and a system that returns claims without that list forces the assistant to either omit the source or invent a plausible one.

There is a distinction inside that which took us a while to get right. The name of the bytes and the name a person reads are frequently not the same thing — “Project Falcon IM v7 FINAL(2).pdf” against “Meridian Health information memorandum”. Both are kept, because a citation that names only the filename is unreadable in an answer, and one that names only the title cannot be found in the folder.

Evidence can also be fetched in bulk rather than one citation at a time, which matters more than it sounds: if checking ten claims means ten round trips, nobody checks ten claims.

Where does this leave the design?

It sets what a fix would have to clear. Closing this gap does not mean a confidence score, which is a number nobody can act on. It means an entailment check: does this passage actually support this sentence, and by how much. That is a real piece of work and it is not shipped, so the limit stands as written.

In the meantime the design leans the other way wherever it can. Contradicting facts are kept rather than averaged, so a weak inference tends to sit beside a stronger one instead of replacing it. Reconciliation is deterministic, so a value that won did so by authority and recency rather than by persuasion. And nothing is overwritten, so the passage that produced a claim is still there in a year when somebody asks where it came from.