GOVERNANCE CHAPTER · 34

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

A Question That Leaves the Work

Chapter summary: When a want reaches outside the manuscript (ch.33), the question that travels must carry the work it came from — the title, the author, and the passages that raised it — because a question stripped of its work is not a smaller question, it is a different one, and it will be answered correctly about something else. Measured: asking "Kinch" returned a Canadian football player who died in 2022, cited and followable and entirely right about a different Kinch. This chapter governs what a question must carry, what may accompany it out of the machine and what may never, how an engine that answers differs from one that retrieves — we may take what it FOUND and never what it SAID — and what must be true of a document before it may be quoted. The writer's own composition never leaves; the source's published words may, bounded and shown verbatim before they go.

Purpose — the failure this exists to end

A correct answer to a question nobody meant to ask.

Measured on a virgin reading of Telemachus, 2026-07-31. The uncertainty map reported Kinch as a name the narrator never uses — a gap a source could close (ch.33). The writer authorised one lookup. The app searched for Kinch, fetched, quoted and presented:

John David Kinch (December 20, 1954 – December 23, 2022) was a Canadian football player who was a running back and fullback for five seasons in the Canadian Football League (CFL) for the Hamilton Tiger-Cats, Toronto Argonauts, Ottawa Rough Riders and Saskatchewan Roughriders.

Every rule of ch.32 held. The work was named, the revision was the locator, the sentence was verbatim, the URL was followable, the retrieval was recorded, nothing entered the world, the decision stayed the writer's. The chapter's guarantees are not what failed. The search did not fail either: it answered what it was asked, accurately, with evidence.

What failed is that the question left the work. Reframe had Ulysses open. corpus:meta held the title. The eighteen passages where the name occurs were in hand, all of them dialogue. None of it was sent. A bare surname went out, and a bare surname belongs to thousands of people.

So the finding generalises past search: context is not an improvement to a question, it is part of the question's identity. Two questions that differ in what they carry are different questions however alike their wording, and presenting one's answer as the other's is how a footballer comes to stand where Stephen Dedalus should.

And the opposite failure is worse, which is why this cannot simply be "send more". The consent card in front of the writer today reads "I would send only the two names in this question." That promise is exact, and it is the reason a writer can say yes without weighing what they are giving away. Everything this chapter adds to a question is something that leaves their machine.

The principle — a question carries its provenance, or it is a different question

A want names a gap in the work (ch.33). The question that leaves must carry enough of that work for an outside source to recognise which gap is meant. Not the manuscript — the provenance of the doubt: what the work is, who made it, what the text does with the name in question.

This is measured material, not reasoned material. The title is stored, the author is stored, the passages are counted, the shape of the usage is countable (ch.27). Assembling them into a sentence is a template. No model composes the question, because nothing about composing it requires judgement, and a composed question is one more place a fabrication could enter.

What a question must carry

The work, as a reader would name it — title and author. "Kinch" is a surname; "Kinch" in James Joyce's Ulysses is a question with one answer.
The subject, as the source spells it.
What the text does with it — the measured shape that made this a gap at all: eighteen occurrences, every one inside dialogue, none in narration.
The evidence — the source's own words, bounded (below).
What would count as an answer, where the want knows: a byname, a place, a date. A question that does not say what it wants invites an answer about anything.

What may leave, and what may never

Reframe reads two very different things, and they do not have the same rights.

The source — Ulysses, the Metamorphoses, a play from an archive: published, in libraries, already in the world. Quoting one of its lines in a search query gives away nothing that is not already given.
The writer's own composition — their draft, their beats, their cut script, their notes. Unpublished, theirs, and the reason this application reasons on the device at all (ch.20).

The writer's own words never leave the machine for a reference lookup, and there is no consent that permits it — not because consent could not be obtained, but because it could never be useful. A reference answers questions about the world the source belongs to (ch.32). No external work knows anything about an unpublished draft, so nothing about that draft could ever be settled by sending it out. A request that would carry it is malformed, not merely risky.

What accompanies a question is therefore the SOURCE's words, and:

bounded — the passages that raised the gap, not the chapter and never the work;
shown verbatim before it goes, in the consent card, exactly as it will be sent. The writer approves the question they can read, never a description of one. A card that says "the two names" while sending three passages has broken the only promise that made the yes meaningful.

Retrieval and answering are different acts

An engine that answers — a model with a search tool — does two things at once: it finds documents, and it writes prose about them with citations attached to its own sentences. These halves have opposite standing under ch.32 rule 1:

What it FOUND is retrieval. The URLs are real, were actually fetched, and can be followed. Discovery is a legitimate service and finding the right document is precisely where the naive search failed.
What it SAID is recall. However well-cited, the prose is the model's account of sources, generated — the thing rule 1 prohibits outright, and the more dangerous for being fluent and footnoted.

So: take the URLs, discard the prose, fetch the document, quote it ourselves. The generated text is never stored, never shown as a source, and never becomes a citation. What the writer reads is the work's own sentence, as ch.32 rule 2 already requires.

This is not a grudging use of such an engine. Reformulating a hard question and finding the right document is real work, and doing it well is the whole difference between the right entry and a footballer.

The return journey — what must be true before a document may be quoted

A document that comes back is a candidate, not a source. Beyond ch.32 rule 2's requirements:

It must be about the work. A page that never mentions the title or the author is not about this Kinch, however exactly it matches the name. This is countable, and it is the check that was missing: the footballer's page says nothing of Ulysses or Joyce. A document failing it is a lead (ch.32 rule 9), never a reference.
The quotation is still selected mechanically — the sentence bearing the subject, verbatim, chosen by measurement and not by a model, because the one thing a citation must never be is our own prose wearing a source's authority.
Nothing promotes without the writer. Unchanged (ch.32 rule 7): the source is shown, the identity is theirs to state.

What the model may contribute, and what it may not

Rule 1 of ch.32 already draws this line and this chapter puts it to work:

It may say what kind of source would settle a question — "an annotated edition would list this byname" — which is a research plan. Where a search finds nothing citable, that answer is honest and useful, and it is recorded as a named but unconsulted work: it may be stated, and it may never confirm anything.
It may not produce the source. A remembered citation is prohibited outright, and is indistinguishable in form from a real one.
It may not decide whether a fetched source answers the question. That is the inference the writer is shown the quotation in order to judge (ch.32 rule 7).

Spend

A free lane and a paid lane are different consents. Asking an open encyclopedia costs the writer nothing; asking a model with a search tool is their money, per act (ch.32 rule 12, ch.20). The card must therefore name which lane it would ask and that it costs, because a yes given to a free lookup is not a yes to a paid one. Measured: a running app resolved its lanes to planner → openai … falling back to your OpenAI API key while its consent card still named a free encyclopedia.

Rules

A question that leaves the work carries the work — title and author — and never travels as a bare term.
It carries the measured shape of the doubt and the source passages that raised it, so an outside source can tell which gap is meant.
The question is assembled by template, never composed by a model (ch.27).
The writer's own composition never leaves the machine for a lookup, under any consent: no external work can settle anything about an unpublished draft, so a request carrying it is malformed.
What accompanies a question is bounded — the passages that raised the gap, never the chapter, never the work.
The writer sees what will be sent, verbatim, before it is sent. The card shows the question, not a description of it.
What an engine FOUND may be used; what it SAID may not. URLs are retrieval; generated prose is recall and is never stored, shown as a source, or cited (ch.32 rule 1).
A returned document must mention the work or its author, or it is a lead and not a reference (ch.32 rule 9). Extended by ch.35: a document that passes this test may still not be the answer — measured, the cited page was an INDEX whose link label was the subject, with the annotation one hop away. A cited page is where reading starts.
The quotation is selected mechanically and verbatim from the fetched document.
Where nothing citable is found, the model may name the kind of work that would settle it, recorded as named but unconsulted — stated, never confirming.
The consent names the lane and its cost. A yes to a free lookup is not a yes to a paid one.
Every attempt is recorded — what was asked, where, what came back, and what was discarded and why (ch.32 rule 13, rule 11).
The lane is removable. Deleting every search adapter leaves the manuscript readable and every reading performable (ch.33 rule 12).

Acceptance

The doctrine is met when:

A question leaving the work names the work and its author, and no lookup is performed on a bare term.
The consent card displays the exact text that will be sent, and sending anything else is impossible.
No passage of the writer's own composition can be made to leave, by any command or consent.
A generated answer appears nowhere in the store and on no surface as a source; only fetched documents are cited.
A document that does not mention the work or its author is reported as a lead — demonstrated on the standing case: a page about a different bearer of the name is refused.
Kinch asked with his work reaches a source about Ulysses, and the writer is shown that source's own sentence.
Where nothing citable is found, the writer is told which work would settle it and that it was not consulted.
The consent names the lane and its cost, and a paid lane is never entered on a consent given to a free one.
Deleting the search adapters leaves reading whole.

Governing sentence

Reframe shall let a question leave the work only in the company of the work — its name, its author, and its own words, bounded and shown before they go — shall never let the writer's unpublished composition travel at all, and shall take from any engine what it FOUND and never what it SAID.