First, an honest word
Language models don’t reason. They remember.
A language model does not think. It has read more of human thought than any person ever will, and it is remarkably good at producing the answer that looks right — the argument shaped like the arguments experts make. That is fluency, not reasoning. Most of the time it is enough. Sometimes it isn’t, and the model cannot tell you which time this is.
Checking it yourself runs into an awkward paradox. If you already know the answer, you didn’t need the machine. If you don’t, you are now researching the very question you asked it — verifying fluent work in a field where you can’t see the joins.
The numbers are genuinely impressive. Depending on whose benchmark you read, today’s best models are somewhere between 80% and 95% correct, somewhere between 80% and 95% of the time. Now hold that against the standard you apply to people. You would eventually part company with an adviser who was right eight times out of ten by luck. The model offers similar odds, shows you no working — and is rather harder to dismiss.
This is not hypothetical
The citation was real. The law was not.
The famous failures were invented cases — Mata v Avianca, the headlines, the sanctions. Any competent legal AI now checks that a cited case exists. The failure that survives that check is quieter: a real authority, cited for something it never said.
Thirty defective citations — most of them real cases
In Coomer v Lindell, the court found some thirty defective citations in a single brief: misquotations, misstated holdings and real cases cited for propositions they do not stand for. A handful were invented; most existed. Existence was never the problem. Counsel and co-counsel were fined $3,000 each.
The sanctions order ↗The same lawyer. A new brief. The same failure.
Ten months later the same counsel was sanctioned again — $5,000 this time — for filings that, in the court’s words, “misrepresent what courts have said”. He told the court he had cite-checked the brief himself. The citations were real. The propositions were not. An adviser on a second warning; a tool that does not take warnings.
The second order ↗Six real cases. Six quotations none of them contain.
In Gauthier v Goodyear, a summary-judgment response attributed fabricated quotations to six real cases, alongside two invented ones. Every one of the six would have sailed through a does-this-case-exist check. Counsel was sanctioned and ordered onto a course about generative AI in legal practice.
The order ↗A citation checker verifies that the case exists. Nothing in that pipeline verifies the argument. Whether an authority actually supports the proposition it is cited for is a question of reasoning — and that is the gap Altien Ratio closes for supported areas of law. A Ratio conclusion is not written first and checked afterwards. It can only exist when the requirements of an approved, versioned rule are satisfied by the stated facts, with the provisions linked. The connection between proposition and authority is not asserted by a model; it is the structure of the answer itself.
And this is no longer only about reputation. In London, the Divisional Court has already referred lawyers to their regulators over unverified authority, and a public database logs new court decisions on the failure roughly five times a day. It is your licence.
Compliance, one enquiry at a time — verifiable by design. Checking stops being research and becomes review.
Why “Ratio”
Ratio, as in ratio decidendi — the reason for the decision.
Every lawyer learned the word in their first week: the ratio is the reasoning a decision stands on. That is exactly what this engine produces — not just an answer, but the rule applied, the facts that satisfied it and the questions still open.
Under the hood this is what researchers call neuro-symbolic AI: language models for the everyday work, symbolic legal rules for the assessment. We prefer the name lawyers already know.
The oldest algorithm in law
If the rule applies, and the facts hold, the conclusion follows.
Lawyers have run this reasoning since Rome. Ratio runs it against maintained rule sets — each requirement tested against the stated circumstances, each conclusion traceable to its rule and its facts.
Ratio decidendi: the reason for the decision. The rule applied, the facts that satisfied it, and the questions still open.
What the reviewer receives
Not a confidence score. Grounds for judgment.
The experience
Ask normally. Review professionally.
The additional assurance appears in the result. The user does not need to learn a rules language or choose an internal reasoning system.
Where it fits
Powerful because it is applied deliberately.
Ratio assessments are available for supported regulatory domains with approved automated rule sets. They strengthen suitable scenario-based questions while keeping professional legal judgment in control.
Encode your domain
Your expertise, encoded once.
Applied right, every time.
AI agents have the potential to deliver legal work at a scale far beyond human oversight. But at that scale, a single mistake can be repeated across hundreds or thousands of matters. The EU AI Act and similar regulations won't absolve firms of responsibility for the decisions their AI makes. Yet AI providers are increasingly keeping their models' reasoning out of sight, and even if they made it available, no human team could possibly review it all.
Human oversight alone is no longer a scalable answer. Firms need AI that can show its working: what evidence it relied on, which rules it applied, and why it reached a particular conclusion. Without that, they risk being unable to defend their AI's decisions when challenged by clients, courts or regulators.
Ratio shows its working. It turns a domain your team knows best into maintained, deterministic rules. Your agents apply them on every matter: the same facts give the same conclusion, with the authority attached. Your people check the facts, not the reasoning.