Grounding

Every answer carries its source, or there is no answer.

The citation guard is the product. This page sets out how it works, what we have measured, and the things juris will not do.

The citation guard

Every claim is checked against the retrieved source before you see it, and the result is shown rather than hidden.

2025 IND 78 Verified

The cited passage was retrieved and supports the claim. This is the only state that should carry weight.

2023 SCJ 112 Cross-reference

The source mentions the citation but is not itself the authority for it. Useful for following a chain, not for relying on.

2019 INT 44 Unverified

The citation could not be matched to a retrieved source. It is marked, never quietly removed, so you can see the gap.

Colour is never the only signal: each state also carries its own glyph, so the distinction survives a monochrome print and a colour vision deficiency.

Refusing is a feature, not a failure

When the corpus does not support a precise answer, juris says so. A confident-sounding answer with nothing under it is the specific failure mode that makes legal AI unusable, and refusing is the only honest response to a question the sources cannot carry.

A refusal notice in a research thread, explaining that the question falls outside the corpus.

Citation health is visible to the firm

Your administrators can see the split between verified, cross-referenced and unverified citations over time, across the whole firm. If grounding quality moves, you find out from your own console rather than from us.

The Citation health panel in the admin console, charting verified, cross-referenced and unverified citation counts over time.

Screens throughout this site show a demo firm with illustrative matters. The legislation and judgments cited are real.

What we have measured

A gold set of 43 real questions with known correct sources, run through the actual production retrieval path. Every stratum, not only the flattering ones.

0.77
Case-law NDCG@10
18 queries. Recall@20 of 1.00: the right judgment was always in the top 20.
1.000
French Codes NDCG@10
6 queries on the Code Civil, Code de Commerce and Code Penal.
0.639
Overall NDCG@10
All 43 queries, including the layman and Kreol strata that pull the average down.

Case law and the French Codes are what a practitioner asks for, and they are where the system is strongest. The overall figure is lower because it includes plain-language and Kreol Morisien questions from the citizen surface, which are harder and which we are still improving. We publish both because a page that showed only the case-law number would be selecting its own evidence, and this product's whole claim is that it does not do that.

Scored with NDCG@10 and recall@k per stratum, against known-correct documents, on the live retrieval stack (voyage-4-large embeddings with Voyage rerank-2.5). The method and the raw results live in the evaluation set behind the product, not in a slide.

What it is not

It does not predict outcomes

juris never expresses a probability of winning or losing, and never tells you how a court is likely to decide. The corpus carries no outcome labels, so any such number would be invented. A separate check strips prediction language from every answer before it is shown, on every surface, for every audience.

It does not give legal advice

Under the Law Practitioners Act 1984, only a person on the Roll may give legal advice or act in a matter. juris does neither. It retrieves, answers from what it retrieved, cites it, and refuses where the sources do not support a precise answer. The instructing practitioner remains of record for every matter run through it.

Questions practitioners ask

What is actually in the corpus?
Published judgments of the Supreme Court, the Master's Court, the Intermediate Court, the Industrial Court, the District Courts and the other lower courts, together with Mauritian Privy Council appeals, the Codes (Code Civil, Code de Commerce, Code de Procedure Civile, Code Penal), the Constitution, and several hundred Acts from the Attorney-General's Office revised laws, plus National Assembly material. Supreme Court coverage is substantial but not complete: the corpus still cites judgments it does not hold, and the live composition is published on the corpus page rather than summarised here.
What happens when a citation cannot be verified?
It is shown to you marked as unverified. It is never silently deleted and never quietly presented as if it had checked out. Your admin console tracks the rate over time, so an increase is visible to your firm.
Is our matter data used to train anything?
Not without your explicit, separate consent. Your firm's documents and questions belong to your firm. They sit in your own instance and are not read across into another firm's deployment. If we ever want to use them to improve a model, we have to come and ask you, as a distinct decision from using the product.
Can it draft?
It extracts, summarises and quotes from sources it can cite, including verbatim clause extraction across a bundle in the Review Grid. It is not a drafting tool that invents provisions, and it will not produce text it cannot ground.

Talk to us about your firm

There is no self-serve signup. Every engagement starts with a conversation about your firm's matters and how the back office should be set up for you.