Deterministic vs. Probabilistic Citation Checks, Explained Simply
When a brief cites a case, someone has to answer a simple question. Is that case real, and does it say what the brief claims it says? There are two very different ways a computer can try to answer that. One method looks the case up in a real record. The other method makes an educated guess. The difference between them decides whether a fake case gets caught before it reaches a judge.
This post explains both methods in plain terms. It covers where each one fails. Then it shows how Veritas puts them together so the weak spot of one is covered by the strength of the other.
The deterministic check: look it up
A deterministic check gives the same answer every time you ask the same question. Think of looking up a word in a printed dictionary. The word is on the page, or it is not. Your mood does not change the answer. The time of day does not change it. Ask ten times and you get the same result ten times.
For a citation, a deterministic check opens the real record. It goes to the court reporter or the public docket and looks at the named volume and page. If the case is printed there, the check confirms it. If nothing is printed there, the check reports that too. It is a lookup, not an opinion.
Smith v. Jones, 100 F.3d 100, 105 (2d Cir. 2000)
→ Located. An opinion is published at 100 F.3d 100 that matches the parties and the court.
The probabilistic check: guess from patterns
A probabilistic check predicts the most likely answer. This is how the autocomplete on your phone works. It has read a huge amount of text, and it guesses what usually comes next. It does not open a record. It estimates.
When you ask an AI model whether a case is real, you are asking for a guess like that. The model reads the citation and predicts whether it looks like a real one. Sometimes the guess is right. The problem is that the guess can sound just as sure when it is wrong. And because it is a prediction, you can ask the exact same question twice and get two different answers.
Smith v. Jones, 100 F.3d 100, 105 (2d Cir. 2000)
→ "This looks like a valid federal appellate citation." (A prediction. No record was opened.)
The risk in the probabilistic check
The biggest risk is the confident wrong answer. The same kind of model that can invent a case can also turn around and vouch for it. If you write a brief with an AI tool and then ask an AI tool whether the cases are real, you may be asking the guesser to grade its own guess. That is how lawyers ended up sanctioned in Mata v. Avianca. The tool returned sure answers about cases that were never published, and the brief was filed anyway.
A guess also leaves no paper trail. There is no record you can hand a court that shows what was checked and what was found. And because the answer can change from run to run, you cannot fully trust that a clean result today means a clean result tomorrow.
The risk in the deterministic check
A deterministic check is honest, but it is only as good as the records it can reach. If a real case sits in a library the tool cannot open, a plain lookup may report that it found nothing. The case is real. The tool just could not see it. A tool that treats "I did not find it" as "it is fake" will flag good citations by mistake.
A lookup can also be too strict about small things. A stray comma or an odd abbreviation can trip up a rigid match, even when the case is right there. And there is one thing a plain lookup simply cannot do. It can confirm that a case exists, but it cannot read the case and tell you whether it actually supports your point. That is a question about meaning, and a lookup does not judge meaning.
Why Veritas uses both, with each in its place
Neither method wins on its own, so Veritas does not pick just one. It uses both, and it lets each one lead where it is strong. The key idea is that the probabilistic layer is added on top of the deterministic one. It never replaces it. A guess is only brought in where a guess is actually the right tool, and a guess never gets to overrule a real record.
Even the first step, pulling every citation out of a long, messy brief, uses both. Deterministic code finds the text that follows the shape of a citation. An AI layer sits on top to catch the ones that are worded in an odd way and might otherwise slip through. The two work together so fewer citations get missed.
The questions with a hard, checkable answer stay deterministic. Does this case exist? Does this quotation actually appear in the opinion? Those are lookups, not opinions, so Veritas answers them against real reporters and public records. A prediction is never the last word on whether a case or a quote is real. The record is.
The probabilistic layer is saved for the softer questions, the ones that call for reading and judgment. Does this real case actually support the point the brief cites it for? That is not a simple lookup, so an AI layer weighs in. Its answer comes as a recommendation in hedged language, not a final ruling. Where the answer is unclear, Veritas flags the ambiguity and points the lawyer to review it. The human makes the call.
Veritas also does the honest thing when it cannot reach a record. Instead of guessing, it says so. A citation the tool cannot confirm comes back as not located in the reporter, which puts the check back in the lawyer's hands rather than hiding behind a false yes. Every result is written into a Verification Certificate that shows what was checked and what was found, so there is a real record to keep in the file.
Doe v. Roe Industries, 300 F.4th 222 (5th Cir. 2021)
→ Not located in reporter. No opinion is published at 300 F.4th 222 that matches the cited parties and circuit. Confirm against an independent source before filing.
The rule: a guess can only lower risk, never raise it
Here is the design rule that ties all of this together. Veritas brings in a probabilistic check only where it can make the filing safer, never where it would add a new way to be wrong. The test is simple. A guess is allowed to raise a flag or add a catch. A guess is never allowed to clear a citation or overrule the record.
Look at the extraction step again. Adding an AI layer there can only find more citations to check. It cannot mark a real citation clean, because whether the case exists is still settled by the deterministic lookup. So the AI only lowers the chance that a citation slips through without being checked. It cannot create a false pass.
The proposition check works the same way. The AI view on whether a case supports a point is offered as a hedged flag for a person to weigh, not a green light. At worst it points a lawyer at something to double-check. It never signs off on a filing. A layer that can only flag, and never approve, adds safety without adding a new way to get it wrong.
That is the whole philosophy. Deterministic checks make every call where a wrong answer would end up in the filing. Probabilistic checks are allowed only where a wrong answer costs nothing worse than a second look.
What the other tools do, and what the numbers show
Most legal AI tools take the other path. They use probabilistic methods to both find and check the citations. The model reads, the model answers, and the model is trusted to be right. Some tools try to make that safer by having two or three AI models argue with each other and vote on the answer. That helps a little. It lowers the chance that one model's bad guess slips through on its own. But a vote among guessers is still a guess. If none of them opened the record, the group can agree on an answer that sounds sure and is wrong. Having the models debate can lower the risk. It can never remove it.
The size of that risk is not a guess either. Researchers at Stanford and Yale tested the biggest legal research platforms and had legal experts hand-check the answers. Westlaw's AI research tool gave a made-up answer on more than one in three queries, over 34 percent. Lexis did better and still missed on more than one in six, over 17 percent. These are the most trusted names in legal research, and the wrong answers still came out sounding confident.
That is the whole reason Veritas does not put the existence question on a guess. No amount of models arguing can turn a prediction into a record. So Veritas checks the record.
The short version
A probabilistic check guesses whether a case is real. A deterministic check looks it up. A guess can be confidently wrong and leaves no record. A plain lookup is honest but can miss a real case it cannot reach, and it cannot judge meaning on its own. Veritas uses both. Deterministic lookups decide the hard facts, whether a case exists and whether a quotation is really in the opinion. A probabilistic layer is added on top for the subjective calls, with hedged language and a nudge to human review wherever the answer is unclear. The guess never gets the final word on whether a case or a quote is real.