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

A Citation Is a Promise Someone Can Check

Chapter summary: ch.32 through ch.35 govern knowledge coming IN — retrieved, never recalled. This chapter governs the claim going OUT: a sentence in the work that cites a source, and what the app owes a reader who tries to follow it. The failure is already on the record, and it is the author of ch.39 committing it: a citation with authors, a venue and a footnote marker, formatted like scholarship, supporting a claim the cited paper does not make — because it came from a search engine's summary and nobody read the source. Nothing in the artifact distinguished that from a checked one. So: verification is an ACT, never a state somebody asserts, and a citation the app has not checked is rendered as a BREAKDOWN — loud — not as a quiet footnote. A citation is a promise that someone can check; until it has been checked, it is a claim about a source and not a source.

Purpose — the failure this exists to end

A sentence that looks checked and is not.

On 2026-08-02 this repository published a governance chapter stating that "an explicit reject/OOS class performs worse than the alternatives", footnoted to Cavalin et al., EMNLP 2020.¹ The footnote was complete: four authors, a venue, a year, a resolvable link. The paper does not make that claim — it concerns word-graph embeddings for class labels. The sentence had come from a search engine's summary of several papers and was published without the source being read.

The chapter it appeared in was about not trusting fluent, confident sentences you have not checked.

Three properties of that failure decide this chapter:

1. It was indistinguishable from good work. Every visible attribute of a sound citation was present. There is no rendering fix for this, because the artifact was not malformed; it was unverified.
2. It was caught by reading the paper, and only by that. Not by review, not by formatting, not by the model.
3. It propagated at the speed of copying. The claim reached the chapter, the README narrative, a commit message and PLANS.md before anyone opened the PDF.

This is not a local embarrassment. Measured on GPT-4o across six simulated literature reviews, 19.9% of generated citations were entirely fabricated, and of the 141 that referred to real work, 45.4% carried bibliographic errors — most often an incorrect or invalid DOI (37.8% of those given one).² The same study found fabrication rising sharply with unfamiliar topics: 6% for major depressive disorder against 28–29% for less visible ones.³ Our incident is the ordinary case, not the exception.

The principle — checked, or said to be unchecked

A citation makes a promise on behalf of the work: a reader who follows this will find it says what I said it says. The promise is what gives a cited sentence its authority, and the authority is what makes an unverified citation harmful rather than merely incomplete — it borrows the credibility of a check nobody performed.

So the app may hold exactly two honest positions about any citation:

• Checked. The source was retrieved and the quotation was found in it. The claim may stand on it.
• Not checked. Said so, visibly, in the same breath as the claim.

There is no third position, and in particular "the writer says it is verified" is not one. A verified flag that a person sets is precisely the artifact that failed above: the ch.39 footnote was, in effect, asserted verified by its author, and the assertion was worth nothing. Verification is an ACT — fetch the source, find the quotation — and its absence is a fact about the app's knowledge, not a preference.

Where this sits

ch.32 governs knowledge entering the work: retrieved, never recalled, with the source's own words. This chapter is that doctrine turned outward. The two share a spine — a claim, a locator, a quotation, a receipt — and differ in direction and therefore in obligation: an incoming reference must be admissible before Reframe believes it; an outgoing citation must be checkable before a READER is asked to.

An outgoing citation is also an outward act, so it inherits what already governs those: what leaves is bounded and shown (ch.34 rules 4–6) — the identifier and the source's own quoted words may travel, the writer's unpublished composition never does — and whose lane and whose money is stated before it runs (ch.20).

Unverified is a breakdown, and must be rendered loud

The scoring kit already draws the distinction this needs. Its .ambiguity is a result — the material genuinely supports more than one reading — while .failure is a breakdown: not assessed, ungrounded, or at risk of being invented, and a renderer "is expected to make .failure louder, never to launder it into a calm open question".

An unverified citation is a .failure. It is not an open question about the source; it is an absence of knowledge, wearing the costume of scholarship. A withdrawn one is the same. Only a checked citation is settled, and a citation whose source was read and does NOT support the claim is a .failure that must be resolved by the writer, never quietly dropped.

This is the property that would have caught ch.39 before publication: the claim would have carried a loud lane saying no one has opened this, rather than reading as finished work.

Withdrawal is recorded, not erased

When a citation fails verification, the claim it supported does not silently lose a footnote. The withdrawal is part of the record: what was claimed, what was cited, what the source actually says, and when it was withdrawn. ch.39 keeps its own withdrawal in the chapter text for exactly this reason — a corrected artifact that hides the correction teaches nothing and cannot be audited.

Every citation therefore keeps its address (ch.36): the claim can be reached from the score, the source from the claim, and the withdrawal from both.

What this does not license

• Not a bibliography feature. Formatting a reference list is not verification, and a well-formatted unverifiable citation is the exact object this chapter exists to prevent.
• Not the model producing citations. ch.32 rule 1 stands: a source is retrieved, never recalled. A model may say what kind of work would support a claim; it may not supply the reference.
• Not automatic assent. A checked citation means the quotation is really there. Whether it SUPPORTS the claim is the writer's judgement, presented to them, never inferred by the app (ch.32 rule 7).
• Not a blocker on writing. A writer may make a claim with no citation at all. That is an honest, unscored sentence; the failure state is reserved for a citation that claims a check that did not happen.

Rules

1. A citation carries a resolvable identifier — DOI, arXiv id, ISBN, or a URL — and enough bibliographic detail for a reader to find the work without it.
2. Verification is an act performed by the app: the source is retrieved and the quotation is found in it. Nothing else may set a citation to checked.
3. An unchecked citation renders as a breakdown, at the same prominence as the claim it supports — never as a quiet footnote, and never laundered into an open question.
4. A citation whose source contradicts the claim is a breakdown the writer resolves. The app states the disagreement and does not choose (ch.32 rule 7).
5. A model never supplies a citation (ch.32 rule 1). It may name the kind of source that would settle a claim.
6. The quotation is the source's own words, selected mechanically, exactly as an incoming reference requires.
7. A withdrawal is recorded with its reason and remains reachable from the claim it once supported.
8. Every citation keeps its address (ch.36): claim ↔︎ score ↔︎ source, both ways.
9. Verification obeys the outward rules: bounded content leaves (ch.34), the writer's unpublished composition never leaves, and the lane and its cost are stated (ch.20).
10. An uncitable claim is allowed and unscored. This governs citations, not assertions.

Showing evidence — the check the writer can make herself

Status: not implemented. This section governs a Copilot feature that does not exist yet. It is written before the code, and the capability identity it names is declared unavailable in the registry (ch.37) until it satisfies what is below. Nothing here describes current behaviour.

The writer must be able to inspect the bounded evidence that supports a citation or research result without being sent into an open web. The evidence surface is a structured, read-only record: provider, work or record identity, locator, retrieval time, content digest, the source's own quotation or extracted passage, provenance, and checked or unchecked state. It is not a page browser and it is not a second reading surface.

The textual evidence and the rendered evidence are one bundle. When the provider returns a source that can be rendered, WebKit may produce a read-only snapshot of that exact retrieved page state. The snapshot is evidence of the retrieval, not a new source. It must remain addressable beside the quotation and receipt, so the writer can see the words that were cited in the source presentation that produced them.

What this must not become

• Not a browser. The view exists to display a bounded evidence record, not to navigate. Following a link is not available from Reframe. A new retrieval requires a new explicit, mediated research or citation request, governed by ch.34 rules 4–6 and ch.20.
• Not a way to set checked. Looking is not verifying. Rule 2 is unchanged: the act is retrieve-and-find, and a writer who reads the page has still not performed it.
• Not a service reporter. ch.48 governs what happens when a call fails; a rendered page and a dead transport share an idiom and nothing else. A spawned CLI failing over MCP has no page to render, and treating that resemblance as a mechanism is the error ch.48 exists to stop.

Rules for the reading surface

11. What is shown is what was retrieved. The evidence record is the retrieval that produced the receipt. If it must be fetched again, the app says so and re-performs the check.
12. The quotation is located in the returned evidence. The record names the locator and preserves the source's own words; the writer is not asked to search an unbounded page.
13. Retrieval is an explicit outward act and is bounded as one. The provider, purpose, scope, lane, and cost are stated before it runs. The writer's unpublished composition never leaves.
14. The evidence surface is read-only. Nothing may be edited, submitted, navigated, or authenticated into from it. It is evidence, not a web client.
15. Evidence that cannot be returned says so, in place — offline, refused, unavailable, or without a checkable quotation — with the same prominence as an unchecked citation. "Could not retrieve" is a state, never a blank pane.
16. Text and snapshot share one evidence identity. A citation bundle binds the claim, provider, locator, retrieval receipt, timestamp, response/content digest, exact quotation, and WebKit snapshot to one retrieval identity. A screenshot from another request, a quotation copied from another response, or a page rendered after navigation cannot satisfy the bundle.
17. WebKit is an evidence renderer, never a browser. The snapshot is read-only and non-navigable: no links, forms, authentication, redirects into a new source, or arbitrary URL entry. A new source requires a new mediated retrieval and a new evidence bundle.
18. The bundle is atomic in presentation. Reframe must not present the quotation as checked while omitting its matching receipt or snapshot when one was available. If any required member is missing or mismatched, the bundle is visibly unchecked or failed, never silently repaired from a different fetch.

Acceptance

The doctrine is met when:

1. No citation can reach the checked state without a retrieval receipt and a quotation located in the fetched source.
2. An unverified or withdrawn citation is visible at the claim, and is rendered in the breakdown register rather than the settled one.
3. Following a citation from the uncertainty display reaches the claim, and following it from the claim reaches the source — demonstrated on the standing case, ch.39's withdrawn footnote.
4. A withdrawal remains in the record with its reason after the claim has been corrected.
5. No model-supplied reference can enter the work.
6. Verifying a citation sends the identifier and the source's own words, and nothing of the writer's composition.
7. A citation evidence bundle can be opened and shows its exact quotation, locator, retrieval receipt, content digest, and matching WebKit snapshot, all resolving to the same retrieval identity.

For the reading surface, when it is built (it does not exist today, and until it does the capability stays declared unavailable per ch.37):

• A checked citation can be opened and the writer sees the bounded retrieved evidence with its quotation and locator.
• When renderable evidence exists, the writer sees the matching WebKit snapshot in the same evidence bundle; its receipt and digest agree with the quotation, and the snapshot cannot navigate.
• A new retrieval states its purpose, provider, lane, and any cost before it runs, and never carries the writer's composition.
• Evidence that cannot be returned renders its reason at claim prominence, never an empty pane.
• Inspecting evidence does not change any citation's checked state, and no web page can be navigated from Reframe.

Governing sentence

Reframe shall let the work cite outward only in a form a reader could check, shall perform that check itself rather than accept anyone's word that it was performed, and shall show an unchecked citation as the breakdown it is — so that a sentence never borrows the authority of a verification that never happened.

Sources