How lawyerbase proves its answers

Independent, peer-reviewed research (Magesh, Surani, Dahl, Suzgun, Manning, and Ho — Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, Journal of Empirical Legal Studies, 2025) tested the leading commercial legal AI systems and found they fabricate or misground citations 17–33% of the time — including systems marketed as “hallucination-free.” One answer in every three to six carries an authority that does not say what the answer claims it says.

For a lawyer, a fabricated citation is not an inconvenience — it is a sanctionable event. Courts in multiple jurisdictions have disciplined counsel for filing AI-invented authorities. Any legal AI that asks for your trust without showing its proof is asking you to carry that risk personally.

lawyerbase is built on one architectural decision: no claim reaches you unchecked. Every answer passes through a pipeline in which the reasoning and the verification are separate, independent steps:

Grounded retrieval. Your question is answered from a corpus of real statutes, codes, and treaties — retrieved by a hybrid engine that combines lexical search, dense semantic search, and a cross-encoder re-ranker. The sources are real documents with a verifiable chain of custody, not model memory.

Citations to the passage, not the vicinity. Each statement in an answer carries a numbered citation to the exact source passage it rests on. One click opens the passage itself — not a paraphrase, not a landing page.

The grounding audit. Before the answer is shown, an independent review pass compares each cited statement against the passage it cites and asks one question: does this passage actually support this claim? Statements that fail are repaired or re-attributed; what cannot be supported is not presented as supported. You see the result of this audit on the answer itself, statement by statement.

Sourced and unsourced, kept apart. Where an answer draws on general legal knowledge beyond the retrieved sources, that content is set apart in its own clearly marked section — so cited and uncited reasoning are never blended.

Uncertainty, named. Where doctrine divides or jurisdictions disagree, the answer says so and shows the competing lines rather than resolving the split with confident tone.

The corpus spans statutes, codes, and treaties across the United States, the European Union, the United Kingdom, Germany, France, Türkiye, Japan, India, Canada, Australia, Spain, and Ireland — and is expanding continuously. Every document is content-addressed with a cryptographic fingerprint and a recorded chain of custody from the official source to the answer, so a citation can always be traced back to the authority it came from.

We hold ourselves to numbers, not adjectives. Before changes reach you, answers are evaluated against a faithfulness harness that scores whether each cited statement is supported by its source — our internal gate requires at least 90% faithful statements with zero fabricated citations on the evaluation set. Sector coverage is probed with expert-grade retrieval questions across maritime, tax, competition, employment, IP, family and other fields. When a gate fails, the release waits.

lawyerbase is a support mechanism. It strengthens a lawyer's judgment — it never replaces it, and it never decides. The final read, the strategy, and the signature remain where they belong: with the human. What lawyerbase changes is the cost of being careful — verification that took hours now takes seconds, because the proof arrives with the answer.

Create a free account and put an answer to the test — everything is included, free for a limited time.