An agent writes to a client about their contract, and the contract it describes belongs to someone else. Nothing broke. Nothing was hacked. The transcript shows a search that returned a document, a model that read it, and a reply built on it, each step working as designed.
You would fire a person for that, and the reason would be plain: they used the nearest file rather than the right one. What you saw was a fluent reply built on a near miss. The Library, the search across a room’s stored material, is built in ArkOne’s reference design for an executive agent so that the nearest file is not enough. It has to be near enough, and there is a number that says how near.
A search never comes back empty
Searching by meaning, which is how an agent finds documents, works by distance. The question and every stored passage become positions in a space, and the passages closest to the question come back first. It is a good way to find the contract when you asked about “last year’s renewal terms” rather than the contract’s title. It has one property people forget: it always returns something. Ask for a client the Library has never heard of, and the closest passage is another client’s file. Ask about a policy your company never wrote, and the closest passage mentions the same words for a different reason.
Nearest is a ranking. It says nothing about whether the top result is any good, only that everything else was worse. A ranking handed to the Chair, the one voice that speaks for the whole room, arrives looking like an answer, and the model reads it as one. A weak match read confidently produces a confident wrong reply, and a confident wrong reply is the kind that leaves the room.
So the reference design measures the distance and refuses to hand over anything beyond a cutoff. The setting is recorded on the Register: a result at a distance beyond 0.55 is discarded before the Chair sees it. Where a near miss would have come back, nothing comes back, and the Chair is told so. The sentence to carry: a weak match is worse than no match, because no match can be reported and a weak match gets used.
One companion rule sits beside it. Messages of one or two characters never reach the Library; there is nothing in “ok” or “y” for a search to find, and the nearest passage to nothing is noise. A three-letter business term still goes through, because “VAT” or “NDA” is a real question.
One search, four results, two thrown away
Suppose the renewal that runs through this series, the invented client asking to keep last year’s price. The Chair asks the Library for the client’s contract and last year’s pricing. Four passages come back ranked, each with its distance.
| Result | Distance, illustrative | Kept or discarded | Why |
|---|---|---|---|
| This client’s contract, current term | 0.18 | Kept | Close: the same client, the same subject |
| Last year’s rate card for this account | 0.24 | Kept | Close: the figures the question asked for |
| A renewal email from a different client, invented for the exercise | 0.61 | Discarded | Beyond the line: right subject, wrong company |
| A web article on discount strategy, fetched last month | 0.72 | Discarded | Beyond the line: the words match, the question does not |
The first two results are the two documents the Chair received in the first paper’s trace, and now you can see why only two. The third and fourth were found; a search that finds four returns four. The cutoff stopped the Chair reading a different client’s renewal terms as this client’s, and stopped a web article standing in for your company’s pricing.
Look at the third row, because it is the one that costs money. The other client’s email is about a renewal, a discount, holding a price. On every surface measure it is a fine match, and without a line to fall beyond it would have arrived in the Chair’s hands beside the real contract, in the same voice. A model handed two contracts for one client will reconcile them, and the reply would have carried terms your client never agreed. That is the failure in the opening paragraph, and the model read exactly what it was given.
Two limits sit beside the capability. The cutoff is a number, and any number is a judgement. Set the line too tight and a badly worded question finds nothing when the right document was there; too loose and near misses come through. The reference design puts the line where a discarded document is the cheaper mistake, on the reasoning that a search reporting nothing sends a person to look, while a search reporting a near miss sends nobody anywhere. That position is written down so you can argue with it. And a distance measures similarity of meaning, not truth: a document close and wrong still comes through, which is why the four shelves label where each passage came from.
Fifteen is the cap on the Register, not a target. Press run and count them.
A meeting has not been run yet. Press run to watch one count, fan out to the specialists, come back, and stop.
What this arms you to ask
Ask the vendor what decides whether a document the agent found is relevant enough to use.
A good answer names a threshold, says where it is set, and shows you a search that returned nothing. That last part is the tell; a system that can come back empty has a line somewhere. Three thinner answers come back, and each describes a ranking rather than a line: it uses semantic search, it returns the top results, the model judges relevance itself. The top result of a search over the wrong material is still the wrong material, delivered first.
The follow-up is one question. Ask them to search for a client the system has never heard of, and show you what the agent does with what comes back. If the reply describes a file, ask whose.
A wrong document is how a bad reply gets built. The fourth kind of tool, from the fourth paper, is how it leaves, which is why the cheapest place to stop it is before the Chair ever reads it.
Next paper: The Cabinet remembers nothing. It reads the Minutes.
Asked plainly
How does an AI agent decide whether a document is relevant?
Searching by meaning turns the question and every stored passage into positions in a space, and the passages nearest the question come back first. Nearest is a ranking rather than a judgement. The reference design described here adds a judgement: each result carries a distance, and any result farther than a fixed cutoff is thrown away before the agent reads it.
Why would an AI agent answer using the wrong document?
Because a search always returns something. Ask about a client the agent has no file for and the nearest passage is another client's file, handed over with the same confidence as a perfect match. If nothing measures how near the match was, the agent reads it as the answer. A distance cutoff turns that weak match into no match, and the agent says it found nothing.
Can an AI agent tell me it found nothing?
Only if the design lets it. A search with no threshold never returns empty; it returns whatever was nearest. The reference design discards results beyond its cutoff, so an empty result is possible and honest, and the agent can say so rather than improvise from a near miss. Very short messages, one or two characters, skip the search altogether.
Ask the vendor
The questions this part arms, each with the good answer and the answer that arrives instead.