Methodology

Describes the verification engine as shipped. Live measured error rates: /quality · weekly public scorecard: /scorecard.

The three layers

1. Layer 1 — Existence. A deterministic lookup: does this citation exist in the authoritative public source for its type? No LLM is involved in producing this verdict.
2. Layer 2 — Quote match. If your document quotes the source, we check whether that quoted language actually appears in the source’s text. Also deterministic.
3. Layer 3 — Proposition support. An LLM-assisted, advisory assessment of whether the cited authority appears to support the proposition it is attached to, reported with a confidence figure. We never phrase this as “verified correct” — it is evidence for your judgment, not a substitute for it.

Verdict semantics — the language contract

Result copy is generated from a single module in our codebase (a “language contract”), so the same verdict always means the same thing everywhere:

• VERIFIED — the citation exists in the named source as of the named timestamp. Where the citation type supports it, this includes a name cross-check: the case name your document claims must match the case actually sitting at that citation. A real reporter slot occupied by a different case does not produce VERIFIED.
• RETRACTED — the citation exists but the source flags it as a retracted publication. Never folded into plain VERIFIED.
• NOT FOUND — the authoritative index for this citation type was reachable and affirmatively has no record. Confident non-existence.
• UNCONFIRMED — we could not confirm the citation as cited, but the signal is not confident non-existence (for example, a claimed-name mismatch, or only a weak fuzzy search was possible).
• UNCONFIRMED — OUTSIDE COVERAGE — the citation type or source is outside our coverage map; not a negative finding.
• SOURCE UNAVAILABLE — the source could not be reached at check time. We report the outage instead of guessing.

Quote-match results use the same discipline: QUOTE CONFIRMED, QUOTE UNCONFIRMED (“we could not find this language” — not “you misquoted”), SOURCE UNAVAILABLE, or NO QUOTE ATTACHED.

What is checked, per vertical

Legal (Citation Safe)

Case citations are checked against CourtListener’s database of published opinions (Free Law Project), with the claimed-case-name cross-check described above. Quotes are matched against the opinion’s actual text. Coverage is strongest for federal courts and state appellate opinions; statutes, regulations, and most trial-level state filings are outside coverage and labeled as such. The live coverage map is on /quality.

Tax (Tax Cite Safe)

• IRC sections (26 U.S.C.) — checked against Cornell LII.
• Treasury Regulations (26 CFR) — checked against the eCFR.
• Tax Court opinions — checked against CourtListener, scoped to the Tax Court. This is a weaker signal than case-law checking (the citation field is often empty for Tax Court memo opinions, so this is a caption/snippet search), and results carry that lower confidence rather than hiding it.
• Outside coverage (labeled, not faked): Revenue Rulings, Revenue Procedures, IRS Notices, and Private Letter Rulings.

Medical (Med Cite Safe)

• PubMed IDs (PMIDs) — checked against NCBI E-utilities.
• DOIs — checked against Crossref.
• Clinical trial registrations (NCT numbers) — checked against ClinicalTrials.gov.
• Retraction check: the retraction signal that matches the identifier you cite is checked — PubMed’s “Retracted Publication” flag for PMIDs, or Crossref’s retraction metadata (with a “RETRACTED:” title-prefix fallback) for DOIs — so that a paper the source flags as retracted returns RETRACTED rather than plain VERIFIED.

Government contracts (FAR Check)

• FAR and DFARS clauses — checked against acquisition.gov, including a currency signal from the published Federal Acquisition Circular data where available.
• GAO bid-protest decisions — we recognize the citation format and construct the canonical gao.gov link, but GAO blocks some automated retrieval; when we cannot fetch the source we return SOURCE UNAVAILABLE with the link, never a fabricated verdict in either direction.

Disputes and the automatic refund

Legal, tax, and medical citation results each carry a “Dispute this result” control. Filing a dispute immediately re-runs the check through the same source used for the original verdict (and, where the citation type supports it, the same name cross-check). FAR Check (government contracts) disputes are not yet wired to an in-product control — file one by emailing support@citationsafe.com and a human reviews it; the same refund guarantee applies. Two asymmetric rules apply to the automated re-check:

• Fault confirmed = automatic refund. If the original result was VERIFIED on a paid check and the re-check affirmatively finds no record, the refund is issued automatically — no human in the loop on the money — and the failure is logged to our public error rate. Details at /refund.
• Upgrades are never automatic. A re-check that suggests a citation should be upgraded to VERIFIED is always escalated for human confirmation, because a wrong upgrade is the dangerous direction. A lawyer double-checking a real citation is an inconvenience; a lawyer relying on a wrong VERIFIED is not.

The published-error-rate commitment

We publish our False-Verify Rate — the rate at which we stamp VERIFIED on something that is actually wrong — per layer, from a benchmark of confirmed-fake citations (drawn from published AI-sanction orders) paired with confirmed-real ones. Live numbers and eval-set sizes are at /quality; the weekly snapshot and industry comparison are at /scorecard. Where a layer has no persisted measurement yet, those pages say “measurement pending” — we publish measured numbers only, never estimates.

Citation Safe is a verification workflow tool, not legal, tax, or medical advice — see /disclaimer. Digital Empire LLC · 30 N Gould St Ste N, Sheridan, WY 82801.