How Veritas verifies a citation.
Verification means comparing your brief against the source record, not asking a model whether it looks right. Veritas runs deterministic checks first, so the result is a comparison you can reproduce, not a judgment you have to trust.
Three checks. One tamper-evident certificate. A boundary we name out loud.
Deterministic first.
A deterministic check returns the same answer every time, because it is a comparison rather than a judgment. Your quotation either appears in the opinion or it does not. The cited case either appears in the reporter or it does not. There is no sampling, no temperature, and no chance the check produces a citation of its own.
That is the entire design principle. A verification layer that could itself generate an unverifiable result would not be verification. So the work that decides whether a citation exists and whether a quote matches is done by comparison against the source, not by a model recalling law from memory.
Confirming that a case is real is the easy half. A reporter volume, a court, a case name: any tool can check those against a database without a model, and a growing number now do. The half that actually catches the error that sanctions a brief is the quotation. Confirming that the language in your brief appears in the opinion, word for word, means reading the opinion itself, not just looking up whether the case exists. That is why our quote check reads opinion text we hold in a corpus we own, rather than trusting a verdict returned by a database we query. The determinism that protects you is the determinism on the quote.
A model does appear in three narrow places, and the rule on all three is the same: a model can only ever make the result more cautious, never less. It cannot pass a citation, vouch for a quote, or manufacture a result the record does not show. The first place is extraction recall, where a model reads your brief to catch citations whose format our deterministic patterns missed — an unusual string cite, a reporter garbled by a scan — so they get checked too. It can only ever add a citation to the checks: every citation it surfaces has to appear word-for-word in your brief, and then runs the same deterministic verification as every other, so it widens what we examine without ever vouching for anything. The second place is the treatment check, where a small model acts as a fail-closed precision filter that can rule a candidate flag out as a false alarm but can never rule one in. The third is optional proposition analysis, which is grounded in the retrieved opinion text rather than asked to remember the law. Everything that could pass a citation as good is a deterministic comparison against the source.
The three checks.
Does the case exist?
We look each cited case up against an owned corpus of public court reporters and public records, and confirm case-law existence against CourtListener, the public opinion database. The check is a deterministic comparison against the record rather than a model's opinion about your brief. A cited case that does not appear in the record flags immediately. A citation we cannot confirm against the available record returns an inconclusive result that asks you to compare against the source. We do not guess, and we do not pass an unverifiable cite as if it were confirmed.
Does the quotation match the opinion?
We take the language you quoted and match it character-by-character against the actual text of the cited opinion. A verbatim match is reported as located in the cited opinion. A passage that does not appear in the opinion does not receive a verified verdict. It routes to a review warning that asks you to compare the quoted language against the source.
The existence check and the quote check run independently. A citation can point to a genuine case and still have its quotation routed for review. A real case does not carry an unverifiable quote through as if the case alone vouched for the words.
Is the case still good law?
Treatment detection is deterministic. We scan the record for explicit subsequent-history signals that a case was overruled, abrogated, or superseded, and we report what the record says about a case's later history. Behind that detector sits a small model serving one narrow role: a fail-closed precision filter. It can rule a candidate signal out as a false alarm. It can never create negative treatment the record does not show, and if it is unavailable or unsure the candidate simply does not count. It reads a single short passage and answers whether that passage explicitly states the case was treated negatively. It never reads two opinions and infers that one implicitly overruled the other. The verdict stays hedged regardless of what the filter returns.
Corpus as of.
A verification result is only as current as the record behind it, so we surface the as-of date of each layer rather than asking you to take its currency on faith. The dates below reflect the corpus this build was cut against.
Opinion full text and the good-law signal reflect a point-in-time snapshot dated March 31, 2026. We refresh that snapshot through incremental ingest, and until that ingest advances it, good-law currency is the snapshot date rather than today. The citation index and statute layers refresh on the cadence shown above.
Real editing, not needless warnings.
Lawyers edit quotations. Opinions get scanned. A careful verification tool has to tell the difference between legitimate editing and a real alteration, without crying wolf on the first or waving through the second. Here is how Veritas handles the cases that come up every day.
Editing conventions are honored, not exempted.
We strip the bracket characters and keep the inner text, then match. A standard editorial bracket like “[t]he” matches the source cleanly, so we do not raise a false alarm on normal Bluebook editing. At the same time, substantive text placed inside brackets still has to appear in the opinion, so the inner text is held to the source. Brackets are an editing convention, not an exemption.
What was cut is recovered and shown.
When a quotation uses an ellipsis, we recover the omitted opinion text and show it in a side-by-side redline. You see exactly what was cut out of the quoted passage. We do not simply confirm that the remaining words matched. We make the omission visible, so you can judge whether the cut changed the meaning, with the cut text in front of you.
Scan noise routes to review, never a silent pass.
Our matching is character-normalized, which absorbs routine formatting and scan noise. A residual OCR-level discrepancy is treated as a near-miss and routed to a compare-against-the-source warning rather than passing silently or reading as a failure. You confirm a scan artifact in seconds. The choice is deliberate: surface the borderline span for a quick human check, never wave it through and never read it as an accusation.
A single altered word belongs in human review.
A single altered token, for example a changed number, a swapped name, or an altered negation, is held below the verified threshold and routed to a review warning with the specific span shown. It is not reported as verified, and it is not reported as a failure. A single altered word in a quotation can be the most consequential change in a brief, so the right place for it is human review, not an automatic pass and not an automatic accusation. We do not raise false alarms on legitimate editing, and we do not let a real alteration slide.
The certificate is the evidence.
Every scan produces a Verification Certificate. It is a tamper-evident audit artifact. It carries an immutable timestamp and a SHA-256 hash of the checked content. It lives on an append-only per-firm audit trail. It is exportable, and it lives at a shareable URL you can hand to a partner, a client, or a malpractice carrier.
The certificate proves what was checked, what each verdict was, and when the check happened. If the underlying content changes, the hash no longer matches, so the record verifies itself. It is the durable evidence that a citation-verification step occurred before filing.
Your firm's data, isolated.
Each firm's data is isolated at the database layer using row-level security, so one firm's documents are not visible to another. The separation is enforced by the database itself, not only by application code. Data is encrypted with TLS in transit and AES-256 at rest, and it is hosted in US cloud regions. Access is single sign-on and organization-scoped, so only your authorized people reach your firm's records.
What Veritas does, and does not do.
Veritas verifies three things. It confirms that a citation exists in the record. It confirms that a quotation is accurate to the source. It confirms that a case is still good law per the record. It produces a tamper-evident certificate of those checks.
Veritas does not decide whether the cited case's holding actually supports the legal proposition it is offered for. That judgment belongs to the attorney who knows the argument and the governing jurisdiction. The certificate proves what was checked and when. It is not a transfer of the attorney's duty, and it is not a malpractice indemnity. The duty to read the case stays with the attorney of record.
We name this boundary on purpose. An audit layer that claimed to relieve a lawyer of the duty to read the case would be selling something no careful lawyer should buy. The limit is what makes Veritas a verification layer underneath the attorney's judgment, rather than a tool that pretends to assume it.
Is Veritas just another AI reading my brief?
No. The work that could pass a citation is deterministic comparison against the source, not a model's opinion about your brief. We look each cited case up against an owned corpus of public court reporters and public records, and we match your quoted text character-by-character against the actual opinion. A model appears in only three narrow places, and in all three it can only make the result more cautious, never less: extraction recall, where a model catches citations whose format our patterns missed so they get checked too, but which can only ever add a citation to the checks and never pass one; a fail-closed precision filter on treatment that can rule a flag out as a false alarm but can never vouch for a citation; and an optional proposition analysis grounded in the retrieved opinion text. Nothing a model does can pass a cite or a quote as good.
What does “deterministic” mean, and why does it matter?
A deterministic check returns the same result every time because it is a comparison, not a judgment. Your quotation either appears in the opinion or it does not. The case either appears in the reporter or it does not. There is no sampling and no chance the check produces a citation of its own. A verification layer that could itself generate an unverifiable result would not be a verification layer, so we built ours to compare rather than to guess.
Can Veritas catch a made-up quotation on a case that actually exists?
Yes. The existence check and the quote check run independently. A citation can point to a genuine case and still have its quoted language routed for review. We match the quoted passage against the actual opinion. A passage that does not appear in the opinion does not receive a verified verdict. It returns a warning that asks you to compare the quoted language against the source before filing.
What if a citation is in an unusual format your patterns don't recognize?
We still check it. Extraction begins with deterministic patterns that recognize standard citation forms across federal and state reporters. Because a citation those patterns miss would otherwise go unchecked, a model then reads the full brief to surface any additional citations — an unusual string cite, a short-form reference, a reporter garbled by a scan. Anything it surfaces has to appear word-for-word in your brief, and then runs the same existence, quotation, and treatment checks as every other citation. The model can only ever add a citation to the checks: it cannot pass one, and it cannot introduce a citation that is not already in your text. The effect is a more complete set of citations to verify, never a less cautious result.
I edit quotations with brackets and ellipses. Will Veritas false-alarm on normal editing?
No. We strip bracket characters and keep the inner text, so a standard editorial bracket like “[t]he” matches the source cleanly. For an ellipsis, we recover the omitted opinion text and show it in a side-by-side redline, so you can see exactly what was cut. Legitimate editing passes. At the same time, substantive text placed inside brackets still has to appear in the opinion, so the check holds the inner text to the source.
What happens if a quotation differs from the source by a single word?
A single near-miss token, for example a changed number, a swapped name, or an altered negation, is held below the verified threshold and routed to a review warning. It is not reported as verified, and it is not reported as a failure. You see the specific span and a prompt to compare it against the source. A single altered word in a quotation can be the most consequential change in a brief, so we surface it for your eyes rather than guessing what was intended.
Scanned opinions are messy. Will OCR noise trigger needless warnings?
Our matching is character-normalized, which absorbs routine formatting and scan noise. A residual OCR-level discrepancy is treated as a near-miss and routed to a “compare against the source” warning rather than passing silently or reading as a failure. You can confirm a scan artifact in seconds. The design choice is to surface the borderline span for a quick human check, not to wave it through.
How does Veritas decide whether a case is still good law?
Treatment detection is deterministic. We scan the record for explicit subsequent-history signals that a case was overruled, abrogated, or superseded. A small model sits behind that detector as a fail-closed precision filter: it can rule a candidate signal out as a false alarm, but it can never create negative treatment the record does not show, and if it is unavailable or unsure the candidate does not count. It reads a single short passage for an explicit statement. It never reads two opinions and infers that one implicitly overruled the other. Where the record contains an explicit signal, we report it. Where it does not, we say the treatment could not be confirmed and direct you to check.
What is the Verification Certificate, and what does it prove?
The certificate is a tamper-evident audit artifact. It carries an immutable timestamp and a SHA-256 hash of the checked content, lives on an append-only per-firm audit trail, and is exportable. It proves what was checked, what each verdict was, and when the check happened. If the underlying content changes, the hash no longer matches, so the record verifies itself. It is the durable evidence that a citation-verification step occurred before filing.
How is my firm's data stored and protected?
Each firm's data is isolated at the database layer using row-level security, so one firm's documents are not visible to another. Data is encrypted with TLS in transit and AES-256 at rest, hosted in US cloud regions. Access is single sign-on and organization-scoped, so only your authorized people reach your firm's records.
Does Veritas tell me my brief is safe to file?
Veritas tells you whether each citation exists, whether each quotation matches the source, and whether each case is still good law per the record, and it gives you a tamper-evident certificate of those checks. It does not tell you that your legal argument is sound or that a case supports the proposition you cite it for. That judgment is yours, and the duty to read the case stays with the attorney of record. Veritas is the audit layer underneath your judgment, not a substitute for it.
See it on your own brief.
Drop a PDF on the scanner. The three checks run, and you get a Verification Certificate of exactly what was examined. No demo, no sales call.