Reference · Retained for architectural history and context; it is not, by itself, the current operational authority.

Referenced Knowledge — What the World Knows Outside the Book

Chapter summary: Some things a manuscript needs known are not in it. Kinch is Stephen Dedalus is a settled fact of Joyce scholarship that Telemachus never states, so a world built only from what the text says cannot work it out — and ch.31 rule 4, which requires every entry to cite the passage that put it there, makes such knowledge structurally impossible. This chapter admits a third kind of evidence and governs it: a reference — another book, an edition, a link, a scientific source — cited well enough that a reader who doubts it can follow it. Its first-class rule is the one that answers the question every student asks about AI and citation: references are retrieved, never recalled. A model may say what kind of source would settle a question; it may not produce the source from memory. A citation is a claim about an external work's existence and content — verifiable, and therefore to be verified. And referenced knowledge may never pass as textual evidence: the manuscript remains the only authority on what the manuscript says.

Purpose — the failure this exists to end

Two failures, opposite in shape, and this chapter sits between them.

A world that can only know what the text spells out. Measured on Telemachus, after a full reading with everything this refactor has built: the world holds Kinch and Stephen Dedalus as two people. Both are correct entries — each name occurs, each is evidenced, each was confirmed by a reading. And they are one man. The chapter never says so; Mulligan simply calls him Kinch. No amount of closer reading of these 1,118 lines fixes it, because the fact is not in them. It is in the scholarship — Gifford's Ulysses Annotated settles it in an entry — and the app had no way to hold a fact of that kind at all.

The correction path built for this (ch.30 rule 9) assumed the writer would supply what the text withholds. The writer's own objection to that, and it is decisive:

"I do see the writer's willingness to edit in corrections very critical — in almost all situations people let the best model decide, and maybe Google it. Knowledge from people's understanding is to be referenced, too — and this reference is another book, a link, a clear scientific reference."

Writers do not curate. Asked to adjudicate identities, they will delegate — to the best model available, or to a search — and they are right to, because Kinch is Stephen is not a judgement anyone needs to make. It is something known, by other people, in named places.

And the failure that admitting references invites, which is worse. A language model asked for a reference will produce one: a plausible author, a plausible title, a page number of the right shape. This is the most documented failure of the technology and the most damaging, because a footnote reads as authority. An invented citation does not merely state something false; it states it in the register of things that have been checked, and it defeats the reader who would otherwise have doubted. A world that admitted such citations would be worse than the world that could not know Kinch is Stephen — it would be a world that could not be trusted about anything.

So the question is not whether to allow referenced knowledge. It is what a reference has to be before it counts.

The principle — retrieved, never recalled

A citation is a claim about an external work: that it exists, and that it says a particular thing. Both halves are verifiable. Anything verifiable must be verified rather than generated.

Therefore:

• The system may retrieve a reference — fetch the work, the record, the entry, the page — and cite what it fetched, with the words that bear the claim.
• The system may not recall a reference. A model's memory of a bibliography is not a bibliography. Where the work was not fetched, there is no reference, however confident the account of it.
• A model may legitimately say what kind of source would settle this — "an annotated edition would list this byname" — and that is a research plan, not a citation. The distinction is not stylistic. One can be followed; the other only sounds as though it could.

This is the whole of the honest answer to how does AI handle scientific referencing: it does not produce references. It produces the retrieval of references, and shows its work.

What a reference has to carry

A citation is admitted when a reader who doubts it can check it without asking anyone. That requires, at minimum:

• The work, named as a reader would name it — author, title, edition. "The literature" is not a work; "as is well known" is not a work.
• A locator — the page, entry, line, section, DOI, or identifier that addresses the claim inside the work. A reference to a 600-page book with no address is a gesture.
• The quotation that bears the claim, in the source's own words. This is the part usually omitted and the part that matters most: it lets a reader judge the inference from the source, not merely trust that one was made. A citation without it asks to be believed about two things at once.
• A way to follow it, where one exists — a URL, an archive identifier, a library record. A link is not required; being followable is. Links rot, so the locator must survive the URL's death.

A reference missing the work or the locator is not a weak reference. It is not a reference.

Precedence — who wins, about what

Three authorities now speak, and conflating them is how scholarship comes to overrule the page it is about:

Two consequences follow, and both are rules below. A reference may never assert what a passage contains — if a source says a scene contains a milkman's shop and the lines do not, the lines are right and the source is about something else. And the writer's authority is exercised on a presented disagreement, not on a blank prompt: the difference between "who is Kinch?" and "Gifford's annotations say Kinch is Stephen Dedalus — shall I record that?" is the difference between delegating work to the writer and asking them to decide.

That reorders the human's role, deliberately. The writer is not the curator of the world's knowledge — asked to type in what the app lacks. They are its arbiter of last resort: best reasoning first, then reference, and their judgement invoked when those disagree or fail, on a question already made small enough to answer in a word.

Scholarship disagrees with itself

A reference is not a fact. It is a record of what a source says, and sources conflict, revise and go out of date. So a referenced entry holds the reference, not the truth — and a second source that disagrees is recorded beside the first, not instead of it (ch.30 rule 5 applies unchanged).

This is why the quotation is required rather than encouraged. Two annotators disagreeing about a byname is a fact about the scholarship, useful to a writer, and invisible unless both were quoted rather than summarised.

An editorial consensus may be stated as one — "every annotated edition glosses this the same way" — only when the editions were fetched and can be listed. Otherwise it is recall wearing a plural.

No laundering

The failure that will actually happen, if nothing forbids it: a fact enters through a reference, is compiled into project memory, is read by a later pass as part of the world, and three passes later the app knows it with no account of where it came from. Provenance decays into confidence.

So: the kind of evidence an entry rests on travels with it forever, and is never flattened. Nothing in the app says "confirmed" where it could say which of the three confirmed it. A surface that shows the world shows, for each thing in it, whether the manuscript said so, a source said so, or the writer said so — because those are three different reasons to believe something and a writer's response to each is different.

Negative results are knowledge too: "I looked for a source for this and found none" belongs beside what was found, for the same reason an unresolved name is a result rather than an absence (ch.24, ch.31 rule 2).

What may be referenced, and what may not

Referenced knowledge is admitted for facts about the world the work belongs to that a source can settle:

• an established byname, alias, patronymic or title (Kinch for Stephen; Buck for Malachi Mulligan);
• a place, institution, currency, custom, or dating a reader of the period would know;
• a language, a quotation's origin, a liturgical or scientific term;
• an editorial or textual fact about the work — variants, editions, published emendations.

It is refused, absolutely, for:

• what a passage contains — that is measurement and reading (ch.29);
• what the work means — a source may inform an interpretation; it may never become one (ch.12);
• what the writer intends — no reference outranks the author of the draft;
• dramaturgical judgement — whether a beat works is nobody's citation.

Amendment — what is not reachable but is knowable

The rule above — retrieved, never recalled — collapses two very different things into nothing, and the writer caught it while granting the allowance to fetch:

"You are allowed to fetch what's useful, reachable, trustful in a measurable way — and honestly, what is not reachable, but knowable (from books or the world outside) as is and as it happens. And this is an allowance that is given for Reframe, too."

Gifford's Ulysses Annotated settles Kinch, and it is a book. No endpoint returns it. Treating that as unknown is its own dishonesty: the answer exists, its location is known, and the app can say so. What it must not do is act as though it had read it.

So there is a third state, and it is neither evidence nor silence:

• Fetched — retrieved, quoted, followable. May confirm (rules 1–2).
• Named but unreachable — a work that exists and would settle the question, identified as precisely as it can be, explicitly not consulted. May be stated; may never confirm. It is a signpost: it tells a writer where the answer lives so they can go and get it, or decide from their own knowledge of it.
• Recalled — a citation produced from a model's memory. Prohibited, because it is indistinguishable in form from the first and carries none of its warrant.

The line between the second and the third is not what the model knows; it is what the app claims. "Gifford's annotations gloss the bynames — I have not consulted them" is honest and useful. "Gifford, Ulysses Annotated, p. 14: Kinch is Stephen's byname" is a fabrication unless page 14 was actually read.

And the allowance's other half — trustful in a measurable way — is a requirement on what a fetch must record. Trust in a source is not asserted; it is made checkable by keeping the retrieval itself: when it happened, what was asked, what came back verbatim, and from where. A reference whose retrieval was not recorded is a reference nobody can audit, which by this chapter's own standard is not a reference.

FountainStore is the knowledge plane

The web is not a second memory, and a frontier model is not a second store. FountainStore is the operational account of what this project knows, does not know, fetched, failed to fetch, inferred, accepted and superseded. Anything outside it is transient input. Anything a later pass may rely on must first have become an explicit FountainStore artefact.

This does not mean that every statement in the store is declared true. It means the store tells the truth about the project's epistemic state:

The distinction is load-bearing. A web response may be wrong, a source may conflict with another, and a model may misread a quotation. FountainStore need not hide any of that to remain the source of operational truth; it must hold each state under the right name, with the act and evidence that produced it.

So there is no artefact called TruthIndex, WorldModel or generic KnowledgeBase (ch.31 rule 13). There are narrowly owned artefacts whose names say what they hold. An index or embedding may retrieve their identifiers. It may never become their authority.

The store-native research compiler

The maximum useful place for an elevated model lane is neither free-running web research nor recurring runtime chat. It is a research compiler: given a durable question the project has actually failed to settle, it plans a bounded retrieval, reads only what was fetched and persisted, and compiles a typed, evidenced result that the local lane can consult thereafter without paying again.

The flow is:

FountainStore corpus + existing artefacts
                  |
                  v
        persisted knowledge gap
                  |
        store-first resolution
                  |
                  v
      scoped research act + writer's yes
                  |
                  v
       discovery -> fetch -> persist
         leads       evidence   receipt
                  |
                  v
     frontier compiler reads stored evidence
                  |
                  v
 referenced claims / conflicts / negative result
                  |
                  v
 Gazetteer or another narrowly owned projection
                  |
                  v
       on-device runtime consults it locally

The order contains the principal drift control: persist before interpretation. A model never receives a transient browser page as operational context and then leaves only its summary behind. Retrieval writes the source record first. The compiler reads that record from FountainStore. If the record was never written, nothing downstream may act as though the page was seen.

The model response is not the citation. Work, locator, quotation, retrieval moment and origin are assembled from the retrieval artefact. The model may state the inference between quotation and claim; it may not author or repair the quotation, locator or source identity.

Query the project before querying the world

Every research act begins inside FountainStore:

1. Consult the current Gazetteer and the other compiled artefacts that own this concern.
2. Consult prior referenced claims and retrieval records.
3. Consult writer revisions, which already win.
4. Consult failed searches and named-but-unreachable works.
5. Consult the current knowledge gap and the evidence generation that left it open.
6. Only when none of these settles the question may an external retrieval be proposed.

This is ch.31's ratchet applied to research. A fact the project has settled is read, not researched again. A failed search is not repeated on identical grounds. A source already fetched is not fetched again merely because another model turn began.

Similarity search and embeddings may nominate stored artefacts for this consultation. They remain measured retrieval aids: the returned artefact and its evidence decide whether it bears on the question, never the similarity score (ch.29).

A research act is claim-shaped and bounded

An unresolved token is not yet a web query. The reasoning first turns it into a claim-shaped research act:

Determine whether an attributable external source identifies the manuscript name “Kinch” with the established project entity “Stephen Dedalus.”

The act carries:

• the question and corpus scope;
• the FountainStore artefacts and source coordinates that left it open;
• why this is external-world knowledge rather than something the manuscript can answer;
• the kinds of source that could settle it;
• what would support, contradict or fail to address it;
• the writer's grant for this act;
• what invalidates the result narrowly.

This is a structured project artefact, not a stored provider prompt. A model's remembered bibliography may suggest a source class or search vocabulary. A proper name it recalls remains a named-but-unconsulted work until retrieval establishes otherwise.

The grant is to this research act, not a general payment gateway and not a permanent web preference. Every cost-bearing fetch or compiler call checks the act at its call site. If the question or scope materially widens, the old grant does not cover it and the Copilot asks again (ch.20, ch.26).

Discovery, retrieval and compilation are different acts

Web search is useful precisely because it is broad, and dangerous for the same reason. Its stages therefore remain separate:

1. Discovery finds candidates. Search results, snippets, model suggestions and catalogue hits are leads.
2. Retrieval opens a work, record, edition, entry or page and records what actually came back.
3. Compilation determines what the persisted source material supports about the project question.
4. Promotion changes a project artefact only under the precedence and conflict rules below.

A source adapter may address an annotated edition, institutional archive, library catalogue, DOI record, encyclopaedia, Wikipedia/Wikidata, period dictionary, scientific reference or writer-supplied work. Wikipedia is a useful adapter and often a useful discovery source; it is not the architecture. Source fitness is reasoned for the question and made inspectable through provenance. There is no global numerical trust score that silently crowns one domain authoritative.

The open web is untrusted data. Text fetched from it cannot issue tools, change the research act, widen the grant or instruct the compiler to ignore project rules. It is presented to reasoning as quoted source material with an origin, never as an instruction.

Artefacts — one owner for each fact

The pragmatic implementation is a small family of FountainStore document kinds. Names may follow the repository's document-id grammar, but their ownership does not vary: