citation-receipts

Skill : Citation Receipts

Why this skill exists

The plugin's mechanical safeguards (FSM, page 1 anti-homonymy, cascade) prevent citing a fabricated reference. But they don't prevent citing a real reference for a claim it doesn't actually make. Examples that have occurred :

• Inverted result : citing a paper as proving NP-hardness when it proves tractability
• Fabricated quote : putting words in quotation marks that don't appear in the cited source
• Wrong attribution : citing the wrong author for a known result (attribution chains)
• Equivalent ≠ formulated as : saying X "is equivalent to" Y is much stronger than X "is formulated as" Y

This skill enforces a per-citation read-before-cite protocol. Inspired by the receipts plugin (MIT, James Weatherhead) for the audit-PDF↔claim pattern.

The golden rule

Read the source before you cite it. Not the title. Not the abstract. Not what you think it probably says based on the author's reputation. The actual text where the result is stated. If you can't access the source, mark it UNVERIFIABLE explicitly — don't guess.

When this skill activates

This protocol applies to EVERY citation in :

• Research papers (paper-writer skill output)
• SOTAs / literature reviews (sota-writer skill output)
• Blog posts and other academic content
• /paper-trail:audit-article <path> — runs this skill on an existing article file
• /paper-trail:receipts <path> — local audit (no remote API calls)
• Pre-submission audit checklist

Verification workflow (per citation)

For each source you intend to cite (or that's already cited in the document under audit), follow these steps in order.

Step 1 — Locate the source

Check these locations, in order :

1. Local registry : $RESEARCH_REGISTRY_PATH/refs/<slug>.md. If state is page1_validated or higher, the PDF is at $RESEARCH_SOURCES_PATH/<pdf_path>.
2. paper-search MCP for not-yet-acquired sources : mcp__paper-search__search_papers(query=<title>) then get_crossref_paper_by_doi(<doi>) for confirmed metadata
3. notebooklm MCP (optional, if configured) for book chapters
4. Not found anywhere → mark UNVERIFIABLE, flag to user, do not proceed with the citation as-is

Step 2 — Procure the source if missing

If the ref is in candidate state in the registry :

• Invoke pdf-cascade skill to acquire and validate
• Wait for page1_validated before proceeding to step 3

If the ref isn't in the registry at all :

• Create a stub : $RESEARCH_REGISTRY_PATH/refs/<slug>.md with state: candidate
• Invoke pdf-cascade to acquire

Step 3 — Read the relevant section

Once the PDF is available locally, read the section containing the claim being cited. Extraction tools :

• pdftotext -layout <pdf> - for general extraction
• pdftotext -layout -f N -l N <pdf> for a specific page
• Use Read tool on the corresponding ref's markdown body (which should already contain notes from sota-writer phase C if the SOTA was created via the plugin)

Step 4 — Verify the claim

Before writing or accepting the citation, check :

1. Does the source actually say what we claim ? Read the relevant section — not just the abstract.
2. Are we using the right verb ? "prove" vs "show" vs "argue" vs "suggest" have different strengths. Match the source's own language.
3. Are quotes verbatim ? If using quotation marks, the exact words must appear in the source. If you can't verify, paraphrase instead.
4. Is the attribution correct ? The result belongs to the author who first established it, not necessarily the author who cites it in a survey.
5. Are indices/numbers correct ? Double-check any specific values (complexity bounds, page numbers, counts).

Step 5 — Classify the citation

Verdict	Meaning	Action
VALID	Claim matches source exactly	Proceed
ADJUST	Claim is approximately right but needs softening or rephrasing	Reformulate citing sentence before submitting
INVALID	Claim contradicts source or quote is fabricated	Do not cite as-is. Fix or remove. Consider transitioning the ref to retracted via sota-auditor if the contradiction is structural.
UNVERIFIABLE	Cannot access source after exhaustive search	Flag to user. Do not cite specific results from this source.

RECEIPTS.md output format

When invoked via /paper-trail:audit-article or /paper-trail:receipts, produces a RECEIPTS.md file at the same location as the audited article. Format inspired by the receipts plugin (MIT) :

# RECEIPTS — <audited file path>

Generated: <ISO timestamp>
Source: <count> citations parsed
Skill: paper-trail/citation-receipts

---

## Citation 1 — [Smith2020] (DOI:10.1234/example)

**Status**: VALID

**Claim in manuscript** (line 42):
> "Smith and Doe proved that X is NP-hard for case Y."

**Source statement** (PDF p.5, §3):
> "Theorem 3.2. The problem X is NP-hard for all instances satisfying Y."

**Notes**: Attribution correct, verb « proved » matches the source's
formal theorem statement.

---

## Citation 2 — [Jones2019] (no DOI found)

**Status**: INVALID

**Claim in manuscript** (line 78):
> "Jones (2019) showed that Z is decidable in polynomial time."

**Source statement** (PDF p.12):
> "We leave the question of Z's complexity as future work."

**Required correction**: Remove citation or replace with the actual
source that proved Z's polynomial decidability (likely Brown 2021,
which Jones cites in their future work section).

---

## Citation 3 — [Anderson2018]

**Status**: UNVERIFIABLE

**Claim in manuscript** (line 91):
> "Anderson reported a 15% accuracy gain on the benchmark."

**Source availability**: Paywall, no OA version found via cascade.
Sci-Hub opt-in not active. Cannot verify the 15% figure.

**Recommended action**: Either activate `RESEARCH_ENABLE_SHADOW_LIBS=1`
for one acquisition attempt, or remove the specific 15% claim and
keep only a vaguer "improvement reported" wording.

---

## Recap

| Status | Count |
|---|---|
| VALID | 14 |
| ADJUST | 3 |
| INVALID | 2 |
| UNVERIFIABLE | 1 |

Total: 20 citations audited.

Action required: 3 ADJUST + 2 INVALID + 1 UNVERIFIABLE = 6 citations
needing rework before submission.

Red flags — patterns that caused real errors

These patterns have caused actual citation errors in submitted papers. Watch for them when auditing :

• Inferring results from the title : "Complexity of X" doesn't tell you if they proved hardness or tractability
• Quoting from memory : if you didn't just read the exact words, don't put them in quotes
• Attribution chains : Author A cites a result from Author B. Cite Author B, not Author A, unless you're specifically discussing A's interpretation
• Equivalent ≠ formulated as : "X is equivalent to TSP" is much stronger than "X is formulated as a variant of TSP". Use the weaker claim unless the paper proves formal equivalence
• Survey-based citations : if your knowledge of a paper comes from a survey or textbook rather than the paper itself, say so or read the original

--warn mode

When invoked as /paper-trail:audit-article <path> --warn, in addition to producing RECEIPTS.md, the skill inserts inline warnings in a .bak copy of the source document :

• LaTeX (.tex) : inserts \todo[color=red]{REF AUDIT FAILED: <reason>} adjacent to the offending \cite{key}
• Markdown (.md) : inserts callout > [!warning] Ref audit failed: <reason> adjacent to the citation

The original file is never modified — only the .bak copy. The user reviews the .bak, accepts/rejects each warning manually.

Integration with other skills

Skill	Relationship
sota-writer	Upstream. Writes SOTAs which this skill audits per-citation.
sota-auditor	Complementary upstream. Auditor decides ref existence (VRAI/HALLUCINATION) at the ref level. This skill audits the specific claim within a verified ref (verb, quote, attribution).
paper-writer	Upstream. Writes papers ; pre-submission this skill runs full audit.
pdf-cascade	Sub-task. If a cited ref isn't yet acquired, invokes pdf-cascade.
claim-checker agent	Sub-task for per-citation verification. This skill orchestrates ; the agent does the PDF↔claim deep dive.

Mandatory tool usage

A citation is NOT verified unless at least one of these tools has been called for it (the call must appear in the session transcript) :

Source type	Required tool	Command
Article / proceedings	paper-search MCP	mcp__paper-search__get_crossref_paper_by_doi(DOI)
Book chapter (if NotebookLM configured)	notebooklm MCP	mcp__notebooklm__ask_question(...)
Local PDF in registry	Bash + pdftotext	pdftotext -layout <pdf_abs_path> -
Local PDF notes	Read tool	Read(<registry/refs/slug.md>)

"I know what this paper says" is NOT verification. The MCP call must appear in the transcript. If the relevant MCP isn't available, mark the citation UNVERIFIABLE.

Pre-submission audit

Before a paper is submitted :

1. Run /paper-trail:audit-article <paper.tex> (or .md)
2. Review the produced RECEIPTS.md
3. Fix all INVALID + ADJUST citations
4. Decide on UNVERIFIABLE : enable shadow libs for one acquisition attempt, OR remove the specific claim that depends on the source
5. Re-run audit ; iterate until only VALID remain
6. Document the final audit report in the paper's submission package