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

One Retrieval, Two Directions

Chapter summary: ch.32 governs a source coming IN, ch.40 a claim going OUT. Built separately, they grew near-identical shapes — a work, a locator, a quotation, a receipt — and three different receipt types for the same act. The visible consequence is a writer being asked to retype a source Reframe has already fetched, quoted and receipted, and a reader having no way to tell which of the two systems is authoritative. This chapter rules that there is one retrieval, recorded once, and two directions of use. A retrieval receipt is a fact about what the app fetched; it does not know which direction it will be used in. An accepted reference is therefore a citation candidate that already carries its check, and the writer's act is to decide the claim it supports — never to re-key the evidence.

Purpose — the failure this exists to end

The same fetch, recorded three times, in three shapes, with no path between them.

Reframe currently holds three receipt types for one act — the app went and got something, and here is what came back:

And two near-identical carriers: GazetteerCitation (work · locator · url · quotation · retrieval) and OutgoingCitation (claim · claimAddress · work · quotation · verification). Both say: here is a source, here are its own words, here is proof someone actually looked.

Three properties of this decide the chapter:

1. The duplication is invisible to the writer and expensive to them. Driven on Romeo and Juliet, recording a citation meant typing a URL and pasting a quotation by hand — for a class of source the app already knows how to fetch, quote and receipt. The work was done and then thrown away.
2. Authority is genuinely unclear from the outside. Two things named "citation" with overlapping fields, and a third (CopilotCitationLink, a wire between a beat and a sentence) sharing the word for something that is not a source at all. A reader of this repository cannot tell which is the record.
3. It will get worse under its own rules, not better. Each chapter correctly forbade the others' shortcuts — ch.32 forbids recalled sources, ch.40 forbids asserted verification — so each grew its own honest machinery. Nothing in the existing doctrine says they are the same act.

The principle — a retrieval does not know what it is for

When Reframe fetches a source and finds words in it, that event is a fact about the world: at this moment, this URL returned this text, and this quotation was at this offset. It is true regardless of whether the app is about to believe something with it, or the writer is about to cite it.

Direction is a property of use, not of retrieval:

• Incoming — the app uses the retrieval to justify what it believes about the world. Governed by ch.32: retrieved, never recalled; the writer judges the inference.
• Outgoing — the work uses the retrieval to support a claim a reader may follow. Governed by ch.40: checked, or visibly not checked.

One record, two readings of it. The obligations differ and both remain in force; what may not differ is the evidence, because it is the same evidence.

What follows: an accepted reference is a citation candidate

If a reference has been retrieved, quoted and receipted for the incoming lane, then everything ch.40 requires of a checked citation has already happened — the source was fetched, the quotation is the source's own words, and a receipt records it. Making the writer retype it is not caution; it is discarding evidence and asking them to reproduce it less reliably.

So the writer's act at the citation boundary is to choose the claim it supports — the one judgement that is irreducibly theirs (ch.32 rule 7). The evidence travels; the judgement does not.

This does not weaken ch.40 rule 2. Verification remains an act, and the act is the one already performed and receipted. What is forbidden is a citation reaching checked with no receipt — not a citation reaching it with a receipt written by the reference lane.

Naming, because the confusion is partly ours

"Citation" currently names three different things, one of which is not a source. A wire between a beat the writer picked up and the sentence they wrote about it is a link on a screen; calling it a citation costs this repository the word it needs for the thing a reader can follow. The wire keeps its behaviour and loses the name.

What this does not license

• Not merging the ledgers. The Gazetteer answers "why does the app believe this about the world"; the work's citations answer "what is this sentence standing on". Different questions, different lifetimes, different owners. They share evidence, not storage.
• Not automatic citation. A retrieved reference becoming available as a citation is not it becoming one. No claim acquires a source without the writer attaching it to that claim.
• Not a weaker receipt. Unifying the three types means the strictest of them wins: a receipt must name where the text came from and what was found in it, or it is not evidence (ch.40 rule 2).
• Not retiring ch.32 or ch.40. Both stand. This chapter governs only what they share.
• Not a migration mandate for stored records. Existing Gazetteer and citation documents remain readable; the unification is of the type, and old records are read forward, never silently rewritten (ch.04: the canonical source remains immutable under analysis).

Rules

1. One retrieval receipt type records the act of fetching a source and locating text in it: when, from where, what was sought, what was found, and how much was read. The three current types converge on it.
2. A retrieval receipt is directionless. It records what happened, never what it will be used for.
3. The incoming and outgoing obligations both stand (ch.32, ch.40); this chapter unifies their evidence, not their duties.
4. An accepted reference is offered as a citation candidate carrying its existing receipt. The writer supplies the claim; the app never supplies it (ch.32 rule 7).
5. A citation created from an accepted reference is checked from the first moment, because the check already happened and is receipted. A citation with no receipt is never checked, whatever created it.
6. No source may be re-keyed by hand when the app already holds its receipt. Manual entry remains available for a source the app has never fetched.
7. "Citation" names a source a reader can follow. A wire between a beat and a sentence is a link and is named as one.
8. Every citation keeps its address in both directions (ch.36): claim ↔︎ receipt ↔︎ source, and from the uncertainty score to all three.
9. Re-checking is always available and always recorded. A receipt is evidence that the source said this then; the web is not immutable, and a re-check writes a new receipt rather than editing the old one.
10. A withdrawn or contradicted citation keeps its receipt (ch.40 rule 7) — the evidence of what was actually fetched is what makes the withdrawal auditable.

Acceptance

The doctrine is met when:

1. One receipt type is used by the Gazetteer, the reference ledger and the work's citations, and no second type records the same act.
2. A writer can turn an accepted reference into a citation on a claim without typing the source, the identifier or the quotation.
3. A citation so created reports checked and names the receipt that checked it — the same receipt the reference lane recorded.
4. A citation with no receipt cannot report checked by any path, including decoding a hand-edited document (ch.40 rule 2).
5. Following a citation reaches its receipt, and following the receipt reaches the fetched source; following the uncertainty note reaches the claim.
6. No type or identifier in the codebase uses "citation" for an on-screen wire.
7. Existing stored Gazetteer and citation records are read successfully after the unification.

Governing sentence

Reframe shall record a retrieval once and use it in either direction, so that evidence the app has already gathered is never discarded, never re-keyed by the writer, and never in two shapes that disagree about what was actually fetched.