Platform

Research that shows its working

DALI is built on one constraint: it may only tell you what it can cite. Everything else follows from that.

How it works

From question to cited note

  1. 01

    Understand the question

    DALI resolves what is actually being asked: the activity, the instrument and the jurisdictions in play. Where the question is genuinely ambiguous it confirms its reading with you, rather than answering a different question confidently.

  2. 02

    Retrieve from primary sources

    A hybrid search over the ingested corpus pulls the governing instruments, the regulator guidance that interprets them, and the case law that has tested them. Retrieval is weighted toward material that is actually in force.

  3. 03

    Synthesise, with the citation attached

    The note is written from what was retrieved. Each proposition carries the source it came from, so the reasoning can be checked rather than trusted, and where the corpus does not support a proposition, DALI does not make one.

  4. 04

    Separate the settled from the contested

    The residual uncertainty is stated explicitly: which factors are determinative, which turn on an untested reading, and where a regulator has not yet spoken. This is the part a general-purpose model reliably papers over.

Design decisions

Six choices that shape every answer

Corpus, not recollection

DALI answers from documents it has ingested and can point to. It does not answer from a model's memory of the law. This is the difference between a citation you can open and a citation that merely looks right.

In force is a first-class fact

Instruments carry a regulatory status and, where known, a commencement date. Material that is made but not yet in force is held apart from law currently in effect and is not allowed to answer a question about today.

Uncertainty is output, not noise

Where the law is genuinely unsettled, that is the finding. DALI marks which factors are determinative, which are merely material, and where the answer turns on a reading no regulator or court has yet confirmed.

Citations are built before the model writes

The references behind an answer are assembled from database records before the language model is called, each already carrying a real source and a precise pinpoint into the document. The model works from citations it was handed rather than composing its own, so there is no step at which it could produce one that does not exist.

Output clears a gate before you see it

Generated text is checked before it is returned: enough distinct sources were actually cited, the answer references material it was given rather than general knowledge, and every inline marker resolves to a pinpoint that was genuinely supplied. Output that fails is discarded and you get an explicit insufficient-corpus response instead of a fluent guess.

Claims it cannot ground are flagged, not buried

After an answer is written, substantive claims are checked back against the sources actually retrieved. Anything the sources do not support is marked as ungrounded, and shown with the material it most likely leans on. You can see which parts of an answer rest on cited authority and which do not.

Against a legal AI tool, not a chatbot

Retrieval fixes the citation. It does not fix the characterisation.

A general-purpose legal AI already retrieves and cites. Four things separate that from what this field requires, and none of them are solved by better retrieval.

A corpus built for mature practice areas

General legal AI is built for high-volume, settled fields where the authorities are stable and plentiful. The instruments, regulator guidance and case law of digital asset, AI and fintech regulation are none of those things. Coverage here is a curation problem, not an indexing one, and a corpus assembled for conveyancing does not acquire MiCA by scale.

No taxonomy to resolve a characterisation against

This is the structural difference. Without a jurisdiction-aware map of what a concept legally is in each regime, a tool has nothing to resolve a term against, so it collapses divergent regimes into whichever meaning is most common in its training data. Retrieval cannot supply what the tool has no structure to represent.

See the taxonomy map

Tuned to draft confidently, not to stop

General legal AI is optimised to produce usable drafting, which means resolving ambiguity rather than reporting it. In a field where a large share of questions have no settled answer, that is the wrong default. An evidence-bounded posture that degrades to an explicit insufficient answer is a design choice, and it has to be built in rather than prompted for.

No view on whether your own work is still current

Retrieval tells you what the law says today. It does not tell you that the memorandum your team relied on last year rests on guidance since replaced. Tracking the currency of a firm's own precedents against a moving body of law is a different capability, and it catches a class of error that is easy to miss.

See precedent currency

See DALI on your own matters

We run a walkthrough against a question your team is actually working on, so you can judge the citations rather than a demo script.

Request access