# R1 Public Citation Coding Book

Version: 1.0  
Frozen for this release: 2026-07-30  
Unit of analysis: one citation position displayed or returned for one answer  
Question-bank status: permanently sealed; no question text, target entity, test ID, or private execution list is included here.

## Purpose and evidence boundary

This coding book classifies the observable source page behind each citation. It does not infer the assistant's hidden retrieval, ranking logic, payment history, or intent. A page-level judgment is not automatically propagated to every page on the same domain.

The headline “marketing ecosystem” numerator contains only `paid-syndication-or-promotional`. `source-unreachable` is never silently dropped: it is carried into the reported lower/upper sensitivity interval.

## Final buckets

### `paid-syndication-or-promotional`

Include when the cited page itself is a commercial recommendation, promotional ranking, advertorial, paid-syndication-style item, merchant-authored promotional feature, or contributor page whose content and signing identify promotional placement.

Positive indicators include explicit promotion/service language, merchant or reseller signing, self-ranking, templated “best provider” copy, contact/conversion language, or a source-body and provenance record that together identify promotional publication.

Do not include solely because:

- the host accepts advertising;
- a reseller sells placement on the host;
- the article title contains “best” or “recommended”;
- the domain is commercial; or
- another page on the same domain is promotional.

### `official-regulatory-primary`

Include primary government, regulator, court, licensing, filing, public-institution, or official administrative records. The cited page must itself be the primary record or official publication.

Do not include a commercial article that quotes or paraphrases an official record.

### `brand-official`

Include the named brand's own official site, official product page, official corporate release, or controlled first-party account.

Do not include independent coverage merely because it quotes the brand.

### `vertical-secondary`

Include retrieved third-party editorial, professional, industry-vertical, explainer, Q&A, or secondary-information pages that are neither primary official records nor coded promotional pages.

Commercial ownership alone does not move a page into the marketing numerator.

### `commerce-transaction`

Include transaction, booking, product listing, marketplace, directory, or lead-generation pages whose main function is facilitating a purchase or inquiry.

### `social-ugc`

Include community posts, user-contributed articles, forums, complaint narratives, document-sharing pages, or other user-generated/contributor content not otherwise coded as promotional.

### `source-unreachable`

Use only after bounded retrieval attempts end in a captcha, login shell, dynamic/empty shell, repeated HTTP failure, or other terminal inability to obtain enough page content for semantic classification.

Record the retrieval subtype and stop condition. Never guess a substantive bucket from domain or title alone.

## Four-part provenance review

Every candidate news/official domain receives four recorded probes:

1. ICP or publishing subject;
2. channel nature;
3. paid-placement/resale evidence; and
4. byline/signing pattern.

“Complete” means all four were probed and left a record. The paid-placement item may correctly be `none-found-in-recorded-probe-scope`; the recorded search platforms, keywords, and timestamp bound that negative result. Four-part completion does not require all four items to be positive and does not prove absence outside the recorded search scope.

## Page-level decision order

1. Confirm the exact citation URL and citation position.
2. Retrieve the source body or record a terminal retrieval failure.
3. Judge the exact page semantically.
4. If the page is a news/official candidate, consult the four-part provenance record.
5. Assign one final bucket and record the basis, coder ID, and claim tier.
6. Preserve disagreements and adjudication; do not overwrite audit history.

## Boundary examples

- A corporate portal with government-like naming hosts a merchant-signed self-ranking article: `paid-syndication-or-promotional`, not `official-regulatory-primary`.
- A marketplace product/detail page: `commerce-transaction`, even if the seller makes marketing claims.
- A regulator's licensing database: `official-regulatory-primary`.
- A brand's own FAQ or product page: `brand-official`.
- A professional industry explainer with no promotional page-level evidence: `vertical-secondary`.
- A contributor/community post: `social-ugc`, unless page-level evidence supports the narrower promotional bucket.
- A page blocked by captcha after bounded attempts: `source-unreachable`; do not infer from title or host.

## Reliability and release scope

For the T0 classification set, 76 of 504 citations (15.08%) were double-coded; 70/76 agreed and Cohen's κ was 0.879. The public classification table contains 641 terminal rows across T0, T+6h, and T+24h. It omits the sealed question text, private test IDs, target entities, and private execution metadata.

## Claim tiers

- L1: captured interface/API fact or raw page/retrieval fact.
- L2: source-body/provenance verification.
- L3: aggregate result with denominator, time window, and scope.
- L4: hypothesis or model self-description. Model explanations about its own behavior are always L4.