AI that never invents the law.
In legal work, a hallucination isn’t a typo — it’s a fabricated citation in front of a judge. So we made a deliberate architectural choice: the language model never touches the facts. Every case name, date, figure and citation comes from real data through deterministic code.
An LLM is not a calculator
Large language models are probabilistic by design. They answer by predicting the most probable next words — pattern-matching against everything they’ve read. Ask “what is 2 + 2?” and the model computes nothing; it returns “4” because it has seen it a million times. Fine for prose. Dangerous the moment the answer is a date, a figure or a citation.
The model has no idea what the actual hearing date is. It returns the pattern that looks most likely. Confident, fluent — and, some of the time, wrong.
Computes the result. Right every time, by construction.
Predicts the most likely next token from patterns it has seen. Usually right — but guessing, not calculating.
Two ways to answer the same question
Give a generic AI tool and LexEdge the same query. The difference is where the facts come from.
Built on millions of judgments
Accuracy isn’t luck — it’s engineering. Every judgment is processed through a data pipeline, then organised with high-level indexing and database optimisation, so a search returns the right passage in milliseconds.
Keyword index
Exact-term matching over structured fields — party, statute, court, citation number.
Vector store
Semantic similarity finds the right passage by meaning, not just matching words.
Court registry
The official authority VerdictVerify cross-checks every citation against.
And when it doesn’t know, it says so
An LLM is built to always produce an answer — so when it doesn’t know, it invents one. Our retrieval is built the opposite way: if there is no grounded match in the corpus, it tells you plainly, rather than guessing.
Where we draw the line
The rule is simple and absolute: the LLM handles language; code handles truth. Anything factual is removed from the model entirely.
Facts from the database
Case names, dates, figures and holdings come from your real records and the judgment corpus — not the model’s memory.
Deterministic search & vectors
Retrieval runs on keyword queries, database indexes and vector similarity — repeatable, inspectable, same answer every time.
Every citation verified
VerdictVerify cross-checks each citation against the official court registry, so a fabricated case is flagged before it ships.
Confident answers, because they’re grounded.
The AI you can put in front of a judge — because the facts came from the record, not a guess.