How we verify
The reason to trust a legal AI is not that it sounds confident — it is what it does when it cannot prove something. Here is the Lexera pipeline end to end: where the law comes from, how citations earn their place, and where we draw the line. The mechanism, not a marketing number.
The verification pipeline
Retrieve, don't recall
Every answer starts with judgments fetched live from Indian Kanoon for that specific query, plus provisions from our statute corpus — not the model's memory of the law. The model reasons over retrieved text, so every citation traces to a source that was actually on the table.
The verification gate
After the answer is drafted, each cited case is matched against the judgments actually retrieved. A citation the gate cannot back is removed or flagged “verify before citing” — it is never dressed up as confirmed. Where a general-purpose AI shows you a citation that merely looks right, Lexera shows you what it can stand behind, and marks the rest.
A verified statute corpus
Statutory text comes from a corpus of 16,700+ sections across 214 Acts, built from official India Code text — including the transition-era codes (BNS, BNSS, BSA) alongside the IPC-era statutes they replaced. Repealed and superseded provisions carry honest repeal notations: the corpus tells you a section is gone rather than quietly pretending it still operates.
Never regenerate to fix a gap
If an answer is missing an authority, we do not re-prompt the model to “find one” — that is exactly how hallucinated citations are born. The unbacked reference is stripped and the answer says so. A shorter honest answer beats a fuller fabricated one, every single time.
Judged from the outside
The pipeline is benchmarked by an external AI judge scoring real legal questions against graded rubrics — correct matter type, mandatory statute sections, landmark authorities that must appear and verify, and hallucination traps that must not. In local certification runs, all five practice areas in our evaluation set scored 9/10 or higher on a median-of-three judging protocol, and live answers average around 9. Individual answers vary — which is exactly why the gate in step 02 exists.
Retrieval → gate → corpus →
no regeneration → external judge.
Every answer walks this line.
The numbers, stated plainly
From official India Code text, repeals marked honestly.
Including BNS, BNSS and BSA alongside the codes they replaced.
Local certification, median of three external-judge passes.
Measured on real traffic — individual answers vary, so the gate stays on.
The honest lines
That citations are verified against retrieved sources, that unbacked ones are stripped or flagged rather than hidden, that statutory text traces to official India Code publications with repeals marked, and that we benchmark accuracy with an external judge and publish the method.
We do not claim a perfect score on every possible question. No system can honestly promise that, and overclaiming is the precise failure this pipeline exists to prevent. Indian law is vast; our certification covers defined practice areas, our scores are measured not marketed, and the verify-before-citing flag tells you where to look twice.
Lexera is a research tool for advocates. Every output is a verification aid — the advocate of record reads it, checks it, and exercises professional judgment before anything is filed. The product is built to make that verification fast, not to replace it.
Built inside the rules
Lexera operates within the Bar Council of India’s professional framework, including Rule 36’s restraints on solicitation. We are a research and drafting tool for licensed advocates and chambers: we do not provide legal advice, do not represent clients, and do not solicit professional engagement. No output creates an advocate–client relationship — the advocate of record does, after applying professional judgment to everything this system produces.
See it strip a bad citation.
Ask a hard question and watch the answer flag what it cannot confirm. That flag is the product.
Try Research