Uk Citation Verification Matei Clej

UK citation verification — against the official register

1. Establish what is being checked: a single citation, a quotation, or a whole document's authorities.
2. Run the check (scripts/ukcite.py), or work the register by hand where no shell is available.
3. Report the graded verdict for each citation — never collapse the grades into "real" and "fake".
4. For anything not VERIFIED, say exactly what the register could and could not answer, and what the user must still do.
5. Never treat a VERIFIED existence check as verification of the proposition the case is cited for. That requires reading the judgment.

What this is for

In Ayinde, R (on the application of) v The London Borough of Haringey [2025] EWHC 1383 (Admin), the Divisional Court dealt with lawyers who had put AI-generated material before courts containing "false information (typically a fake citation or quotation)", and said plainly that the duty to check is the lawyer's and is not discharged by trusting the tool. Regulators in several jurisdictions have since said the same.

This skill makes the checking mechanical where it can be mechanical. It verifies citations against the official public register of judgments for England and Wales — The National Archives' Find Case Law — and statutory references against legislation.gov.uk. Both sources are operated by The National Archives; nothing else is consulted, so every answer is traceable to the official record.

It is not for finding authorities on a topic (use a research service), and it is never a substitute for reading the judgment cited.

The verdicts — what each one proves, and what it does not

There are two ways to fail at citation checking. The obvious one is letting a fabricated case through. The less obvious one — and the one careless tooling commits constantly — is declaring a real case fabricated because one database does not hold it. The graded verdicts exist to keep those apart:

✔ VERIFIED — the citation resolves on the register; where a case name, pinpoint or quotation was supplied, it matched. This proves the case exists and says what was quoted. It does not prove the case is authority for the proposition it is cited for, that it is still good law, or that it has not been distinguished out of relevance. That is reading, not checking.

✖ MISMATCH — the citation resolves, but to a different case, a wrong paragraph, or a quotation that is not verbatim. This is the most dangerous failure mode in AI-assisted drafting: a real citation stitched to the wrong case name, a tidied "quotation" that no longer matches the source. The verdict prints what the register actually holds so the error can be repaired rather than merely deleted.

✖ NOT ON REGISTER — no record, though the court and year fall inside the register's stated coverage. Treat the citation as unsafe to rely on until the judgment is located somewhere — but do not report it as fabricated on this evidence alone. The register's own caveat is that coverage within a range is incomplete: not every judgment is sent for publication. Cross-check by party name (find), then on a subscription service, before any conclusion is stated.

◌ OUTSIDE COVERAGE — the register cannot answer. House of Lords judgments are not on Find Case Law at all; most courts' holdings begin in 2001–2003 and several tribunals much later. For these citations the check is not run and no inference arises. A tool that returns "not found" for Kakis v Cyprus has not found a hallucination; it has found its own boundary.

◌ UNCHECKABLE — a law-report citation ([1978] 1 WLR 779). The register resolves neutral citations only. Find the neutral citation by name where one exists (post-2001), or verify in the printed series. Says nothing about whether the case is real.

Running it

The checker ships in this package; Python 3.9+, standard library only, nothing to install:

# One citation, with everything checkable checked
python3 scripts/ukcite.py check "[2015] EWHC 1274 (Admin)" \
    --name "Polish Judicial Authorities v Celinski" \
    --para 13 \
    --quote "the public interest in ensuring that extradition arrangements are honoured is very high"

# Every citation in a draft — .docx is read natively; plain text and
# markdown too. A PDF is refused with instructions, never scanned as noise:
# a verification tool must not return a false clean bill because it could
# not read the file.
python3 scripts/ukcite.py scan skeleton.docx

# Find a neutral citation from a case name (party-name search first)
python3 scripts/ukcite.py find "Celinski"

# A statutory provision — existence, current version, unapplied amendments
python3 scripts/ukcite.py statute "Extradition Act 2003" --section 2

Exit code 0 means everything checked came back VERIFIED; 1 means something needs attention. Every run writes a dated report to ~/.cache/uk-citation-verification/ so that months later the user can say what was checked, when, and what the register answered.

Reading the output for the user. Report MISMATCH and NOT ON REGISTER prominently, with the register's actual holding alongside the claimed one. Report OUTSIDE COVERAGE and UNCHECKABLE as unverified, with the reason — never in the same breath as the failures. A quotation flagged NEAR MISS is corrected from the source text the tool prints, not from memory.

Without a shell — in an assistant that cannot run Python — do not pretend to check. A neutral citation maps to a register URL directly: [2024] UKSC 12 → https://caselaw.nationalarchives.gov.uk/uksc/2024/12; [2019] EWHC 100 (Admin) → .../ewhc/admin/2019/100. Open it, confirm the case name and date, and read the passage cited. For statutes, open https://www.legislation.gov.uk/ukpga/<year>/<chapter>/section/<n> and check the version notes. Consult reference/coverage.md before drawing anything from a miss, and say plainly which citations were checked by hand and which were not checked at all.