Citation check

A brief with one wrong page is a brief the other side can discredit in a minute. This skill is the last step before any cited document leaves the team: it opens every citation, reads the page, and says whether the sentence that cites it is supported. It never corrects the document silently; it reports, so the author decides.

The skill checks three things for every citation: that the cited source exists and the page is in range (resolves); that the statement in the document is on that page (located); and that the document says no more than the source does (supported, overstated, contradicted, not found).

1. Inputs

• The document to check: a Markdown, DOCX, or PDF brief, timeline, memo, or report. Convert to text, keeping paragraph boundaries.
• The archive the document cites: the case-brief archive (index.csv; one text file per document with === PAGE n === markers; deposition transcripts with page:line text). If the archive is missing, build it first with the method in case-brief; if the user supplies only the PDFs, extract them the same way. A check against a different copy of the archive than the one the brief was built from is still valid, but say so in the report.
• The citation forms in use. Default: [D-0042 p.7], [BATES-000123 p.2], [D-0042 pp.7–9], [Depo Lastname 112:4–113:2], [D-0042 p.7, D-0051 p.2]. Read the document's own Sources section for its conventions and add any pattern you find there.

2. Parse

Run a script, not your eyes, for the parse: a long brief has hundreds of citations and a reader skips some.

import re, csv, sys, pathlib
text = pathlib.Path(sys.argv[1]).read_text()
pat = re.compile(r'\[(?P<body>(?:D-\d{4}|[A-Z]{2,}-\d{3,})\s+pp?\.\s*[\d–\-, ]+(?:,\s*(?:D-\d{4}|[A-Z]{2,}-\d{3,})\s+pp?\.\s*[\d–\-, ]+)*|Depo\s+[A-Z][\w\-]+\s+\d+:\d+(?:[–\-]\d+:\d+)?)\]')
rows = []
paras = re.split(r'\n\s*\n', text)
for pi, para in enumerate(paras, 1):
    for m in pat.finditer(para):
        # the claim is the sentence that ends at the citation
        start = max(para.rfind('. ', 0, m.start()), para.rfind('\n', 0, m.start())) + 1
        claim = para[start:m.start()].strip()
        rows.append({'n': len(rows)+1, 'para': pi, 'citation': m.group('body'), 'claim': claim})
with open('citations.csv', 'w', newline='') as fh:
    w = csv.DictWriter(fh, fieldnames=['n','para','citation','claim']); w.writeheader(); w.writerows(rows)
print(len(rows), 'citations')

Then list every paragraph that states a fact (a date, a name, a quotation, a number, "the record shows") and carries no citation; those go into the report as uncited claims. Headings, the method note, and the author's own open questions are exempt.

Split multi-source citations into one row per source, keeping the same claim.

3. Resolve and locate

For each row:

1. Resolve: the ID is in index.csv; the page (or every page in the range) is within the document's page count; a deposition page:line is within the transcript. Failures: no such document, page out of range, line out of range.
2. Locate: read the cited page(s) from the text file. Find the words that support the claim. Prefer exact numbers, dates, names, and quoted phrases as anchors; for a paraphrase, find the sentence that carries the same meaning. For a deposition, read the cited lines plus three lines on each side, so a cut-off answer is seen in full.
3. Judge with one word:
4. supported — the page says what the claim says, at the same strength.
5. overstated — the page supports a weaker form (the claim says "confirmed", the page says "may indicate"; the claim says "the manufacturer", the page says one engineer; the claim gives a date the page gives only as a month).
6. incomplete — the page supports the claim but the next sentence or the next answer qualifies it in a way the brief omits.
7. contradicted — the page says the opposite or gives a different value.
8. not found — the claim is not on the cited page. Then search the same document for it, and the archive if that fails; record where it actually is (found at D-0042 p.9) or not found anywhere.
9. quotation error — the brief quotes the page and the words differ (give both).
10. cannot check — scanned page without text, redacted, or OCR too poor; say so, never guess.

Record for each row: status · anchor (the words on the page that support or contradict, verbatim, short) · found at (when different from the citation) · note.

Do not stop at the first failure in a paragraph; check every citation, because the report is also a measure of how reliable the document is as a whole.

4. Check the apparatus

Beyond the individual citations:

• Sources section: every document ID the brief cites appears in its Sources list and in index.csv; every Sources entry is cited at least once (or is listed as read but not cited).
• Internal consistency: a date, a name, a serial number, or a version given in two places in the brief is given the same way both times; the timeline tables agree with the prose.
• Deposition names and pages: the witness name matches the transcript caption; the page:line format is page:line, not page:page.
• Numbers: every count in the brief ("fourteen complaints") is recomputed from the table it comes from.
• Dates: every interval stated ("filed 41 days after") is recomputed from the two dates in the sources.

5. Write the report

citation-report.md:

1. Summary: number of citations checked; count per status; number of uncited factual claims; the three to five findings the author must fix before the document is used (contradicted and not-found rows first, then overstated rows on load-bearing claims).
2. Findings table: one row per non-supported citation: n · paragraph · claim (short) · citation · status · page says (anchor, verbatim) · found at · suggested fix (a rewording that the page does support, or "remove", or "cite X instead" — as a suggestion, never applied).
3. Uncited claims: paragraph number and the sentence, each with a suggested source if the archive search found one.
4. Apparatus: Sources-list mismatches, internal inconsistencies, recomputed numbers and dates that differ.
5. Cannot check: pages that need OCR or a better copy, so the team can request them.
6. Method: archive used (path, index date, document count), citation patterns parsed, time of the check.

Also write citations.csv with the status columns filled, so the author can sort it.

6. Rules

• Report; do not edit the document being checked. The author applies the fixes and reruns the check.
• Read the page, not the index and not the brief's own quotation. The archive text is the only authority; if the archive is wrong, the PDF page is.
• Every status except supported carries a verbatim anchor from the page, so the author can see the difference without opening the file.
• overstated is a finding, not a style note. In litigation the strength of a claim is the claim.
• Three lines of context on each side of a deposition cite; an answer read without its question is not read.
• Say what you could not check. A clean report with silent gaps is worse than a report with a cannot check section.
• No opinion on the merits of the document. The report is about whether the sources support what is written, nothing else.