Knowledge and Fact are different things
Knowledge names the whole supplied body — documents, fetched pages and typed entries, each with authority and scope. A Fact is one grounded statement drawn from that body. The distinction is not pedantry: a source can be entirely legitimate and still contain no statement that settles the question in front of it, and treating the presence of a document as proof of a claim is how an agent asserts something nobody agreed to.
The two words, used precisely#
- Knowledge
- Everything a workspace has supplied for answers to be grounded in. It has a size, a status per item, an authority per item and a scope per item.
- Fact
- One statement that some part of Knowledge actually supports, strongly enough to be said out loud to a customer.
- Memory
- Something Connect learned from work rather than being told. Never a substitute for either of the above. See Memory in Connect.
The relationship is one-way. Every usable fact traces to Knowledge; not every piece of Knowledge yields a fact. A forty-page installation manual is Knowledge, and it may contain nothing at all about the warranty period a customer just asked for.
Where the difference is enforced#
Retrieval and sufficiency are two separate steps, and the gap between them is exactly this distinction. Retrieval asks *what material relates to this question*. Sufficiency asks *does any of it carry the authority to state the answer*, and for a commercial, legal or support commitment it demands verified_fact or owner_instruction specifically.
| Question asked of the bank | Yes means |
|---|---|
| Is there relevant material? | Something will appear in the context block. Nothing more. |
| Does anything carry fact authority? | A statement may be made, as far as the text supports it. |
| Is this a critical commitment? | The bar rises: a topical policy is not enough, explicit owner authority is required. |
So a reply can quote a policy, describe how the business handles the topic, and still decline to give the number — not because the model was unhelpful but because the second question came back no. That behaviour is documented at claims Knowledge does not support.
The mistake, and what it looks like in the wild#
The failure runs like this. Somebody uploads the company brochure, marks it verified_fact because it is official, and assumes the bank is now authoritative. A customer asks about a delivery window. The brochure mentions delivery in a marketing sentence. It is retrieved, it carries fact authority, and the answer states a window nobody in the business would have committed to.
- The document was real. The authority was not wrong for most of it. The specific statement was never a fact anyone had approved.
- Authority belongs to material of one kind, and a mixed document is several kinds at once.
- The repair is to split it: the specification tables as
verified_fact, the marketing prose asreference, the commitments as short entries somebody deliberately wrote. - The check that catches it before a customer does is testing retrieval with the questions customers actually ask, and reading which source the top row came from.
Two things this is often confused with#
A fact and a memory. A memory is learned from a conversation and is scored against the request, decays in relevance and can be superseded by a later one. A fact is supplied, and does not change until somebody changes it. A caller saying their warranty runs three years does not make it so, and Connect writing that down as a memory is not the same as it becoming true.
A fact and a citation. A citation proves a passage was in front of the model when the answer was written. It does not prove the sentence followed from it. Auditing an answer means reading the cited passage, not merely observing that one exists — which is why the citation carries the source, the section and the position rather than just a tick.
Questions#
If a document is uploaded and marked verified, is everything in it a fact?
No. Authority is set on the source, but sufficiency is judged per question against the passages that actually matched. A verified document containing nothing about the question yields nothing, and a verified document containing a loose sentence about it is exactly the risk described above. Split mixed documents by the kind of claim their parts make.
Which word should I use when talking to my team?
Say *source* for a thing you added and *fact* for a statement you would be content to see in an email to a customer. If somebody asks whether Connect 'knows' something, the useful reply is which source would support saying it out loud — that question has an answer, and 'does it know' does not.
Does the model decide what counts as a fact?
No. Eligibility, authority and the sufficiency threshold are decided in code before the model sees anything, and the context block tells it in words which material may be stated and which is background only. See who decides what is true.