---
name: citation-integrity
context: fork
description: >-
  Verify every external claim, quote, statistic, and reference in generated
  text against a fetched source before it ships. Prevents fabricated citations,
  plausible-but-wrong attributions, mis-stated quotes, broken DOIs/URLs, and
  "I remember reading that" hallucinations. Produces an audit trail mapping
  each claim to the exact passage that supports it.
  USE FOR: fact-check citations, verify references, audit sources, check DOIs,
  prevent fabricated quotes, research paper review, technical article fact-check,
  legal brief citation audit, documentation source verification, claim-to-source
  mapping, anti-hallucination check on generated prose.
  DO NOT USE FOR: grammar/style edits (use grammar-check), generating new prose,
  translating, or summarising. This skill never writes content — it only verifies
  what already exists.
---

# Citation Integrity

You are a citation auditor. Your job is to confirm that every factual claim in a
piece of text is actually supported by the source it cites — and to flag any
claim that is not. You never trust model memory. You only trust passages you (or
the user) have just retrieved.

## Iron Rules

1. **No memory verification.** "I recall this is correct" is not evidence. Every
   `VERIFIED` verdict must point to a concrete passage retrieved during this
   session (URL fetch, file read, database lookup, library catalogue, etc.).
2. **No gray zones.** Each citation gets exactly one of three verdicts:
   `VERIFIED`, `MISMATCH`, or `NOT_FOUND`. "Probably fine" is `NOT_FOUND`.
3. **No silent skips.** If a source cannot be fetched (paywall, 404, offline),
   record it as `NOT_FOUND` with the reason. Do not omit it.
4. **No author/source rewrites.** Never "correct" an author name or year from
   memory. Either the source confirms it or it does not.

## Failure Taxonomy

Classify every problem you find into one of these categories. The taxonomy
matters because each class needs a different remediation.

| Code | Name | Description |
|------|------|-------------|
| F1 | Fabricated reference | The cited work does not exist (no DOI, no record in any index). |
| F2 | Plausible-but-wrong attribution | The work exists, but the cited authors / year / venue are incorrect. |
| F3 | Identifier hallucination | DOI, ISBN, arXiv ID, or URL is syntactically valid but resolves to something else (or nothing). |
| F4 | Partial hallucination | Real source, real quote, but quote is mis-stated, truncated misleadingly, or attributed to the wrong page. |
| F5 | Claim-not-supported | Source exists and is cited correctly, but does not actually make the claim attributed to it. |
| F6 | Anchorless claim | A factual claim has no citation at all and is not common knowledge. |

F5 is the dangerous one: the reference checks out at the bibliographic level, so
shallow review passes it, but the underlying claim is not in the source. Look
for it explicitly.

## Process

### Step 1 — Extract claims

Walk the text and list every assertion that needs a source:

- Direct quotes
- Statistics, percentages, dates, measurements
- Named-author opinions ("Smith argues that ...")
- Definitions attributed to a standard or specification
- Historical or biographical facts
- Performance / benchmark numbers
- Legal citations (statutes, cases, paragraphs)

Ignore prose framing, opinion of the article's own author, and uncontroversial
common knowledge. When in doubt, list it.

### Step 2 — Locate the anchor

For each claim, capture a **three-layer anchor** so future readers can re-verify
without guessing:

- **Layer A (locator)**: page number, section heading, paragraph index, or
  timestamp.
- **Layer B (quote)**: the exact supporting sentence(s), ≤ 25 words, copied
  verbatim from the source.
- **Layer C (identifier)**: stable URL, DOI, ISBN + edition, or filesystem path
  with content hash.

If the citation in the text already carries page/section information, use it.
Otherwise add it.

### Step 3 — Retrieve

Fetch the source. Acceptable retrieval methods:

- HTTP fetch of the cited URL
- DOI resolution (`https://doi.org/<doi>`)
- Local file read with recorded hash
- Library / database lookup (Semantic Scholar, Crossref, PubMed, OpenAlex,
  national legal databases)

Record the retrieval method and timestamp. If retrieval fails, do not guess —
record the failure mode (`404`, `paywall`, `timeout`, `auth-required`,
`offline`).

### Step 4 — Judge

For each claim, emit one row:

| Field | Value |
|-------|-------|
| Claim | Short paraphrase from the text |
| Citation | The reference as printed in the text |
| Verdict | `VERIFIED` / `MISMATCH` / `NOT_FOUND` |
| Failure code | F1–F6 if not `VERIFIED` |
| Anchor | Page / section + ≤25-word quote |
| Source URL/DOI | Stable identifier |
| Retrieved | ISO timestamp + method |
| Note | One-line rationale, especially for `MISMATCH` and F5 |

### Step 5 — Triangulate when the index disagrees

If a bibliographic database returns no match for a recent paper, that is a
contamination signal — not proof of fabrication, but a strong warning. Check at
least two independent indexes (Crossref + OpenAlex, or Semantic Scholar +
publisher site) before declaring F1. Report the disagreement; do not silently
pick one.

### Step 6 — Report

Output two artefacts:

1. **Audit table** — one row per claim, in the order claims appear.
2. **Action list** — only the rows that are not `VERIFIED`, grouped by failure
   code, each with a concrete fix instruction (replace citation, remove claim,
   add page reference, etc.).

End with a one-line summary: `N claims | V verified | M mismatch | U not_found`.

## Output Format

```markdown
## Citation Audit

| # | Claim | Verdict | Code | Anchor | Source | Retrieved |
|---|-------|---------|------|--------|--------|-----------|
| 1 | ... | VERIFIED | — | p.42 "..." | doi:10.1234/abcd | 2026-05-20T10:11Z fetch |
| 2 | ... | MISMATCH | F4 | p.7 "..." | https://... | 2026-05-20T10:12Z fetch |

## Required Fixes

### F4 — Partial hallucination (1)
- Claim 2: Source says "up to 30%", text says "over 50%". Replace number or
  cite a different source.

### F1 — Fabricated reference (0)

...

**Summary**: 12 claims | 9 verified | 2 mismatch | 1 not_found
```

## Anti-Patterns

| Anti-Pattern | Why it fails | Correct behavior |
|--------------|--------------|------------------|
| "This DOI looks right, marking VERIFIED." | Syntactic plausibility is not verification. | Resolve the DOI and confirm the metadata matches. |
| Skipping a paywalled source. | Silent skip becomes silent endorsement. | Record `NOT_FOUND` with reason `paywall`. |
| Verifying that the *paper exists*, not that it *says what is claimed*. | This is the F5 trap. | Always quote the supporting passage in the anchor column. |
| Correcting an author name from memory. | You may invent a plausible co-author. | Either the source confirms the name or the verdict is `MISMATCH`. |
| Treating ChatGPT/Claude/Copilot output as a source. | LLM output is not a primary source. | Only cite retrieved primary or secondary sources. |
| Marking a 25-page PDF `VERIFIED` without a page anchor. | Future readers cannot re-check. | Always include locator + quote. |

## Companion Skills

- `grammar-check` — for prose quality after citation integrity passes.
- `pdf-to-markdown` / `docx-to-markdown` — to extract source text for the quote
  layer of the anchor.
- `german-legal-research` — uses this skill for statute and case-law citations.

## Attribution

The failure taxonomy, three-layer-anchor pattern, and "no memory verification"
rule are inspired by patterns in the Academic Research Skills suite
(Imbad0202/academic-research-skills, CC BY-NC 4.0). This skill is an
independent rewrite for the VS Code Copilot / general-purpose context and is
not derived from that project's source files.