You have the corpora. One per country, one per publisher deal, each one authoritative on its own. What you don’t have is the layer that connects them: the one that says this provision here is the same requirement as that provision there, and here are the four differences that matter.
That layer is what we build, and we build it on your content.
They match context. We match function.
Every new jurisdiction is a separate negotiation with a separate rights holder, a separate ingest, a separate index. The corpus grows. The connective tissue does not.
So a user in your product can search Italian law and search Danish law, and get a good answer in each. Ask what the Danish equivalent of the Italian provision is and the product has no mechanism at all. Not a weak one, not a slow one. None. The relation was never in the data.
That gap gets wider with every deal you sign. Ten jurisdictions is ten corpora and forty-five possible pairs, none of them modelled. The coverage number on your website goes up; the thing your users actually asked for stays out of reach.
Meanwhile the market has already priced access. Nobody has priced the link.
You could build it. The question is what it takes and what it costs you elsewhere.
Time. A rule-level layer is not an ingest job. It is a per-jurisdiction extraction into a shared functional structure, an anchor ontology that survives contact with civil law and common law at the same time, a derivation step that produces the actual differences, and an adversarial validation pass that catches what the extraction invented. Each of those is a research problem before it is an engineering problem. Getting the first anchor family right takes longer than the whole of your last integration.
People you don’t have. Not annotators. The hard part isn’t volume. It’s the small number of people who can tell, in the source language, whether two provisions do the same work in two legal systems and where they stop doing it. That capacity is the constraint, and hiring for it competes directly with hiring for your product.
Attention. This work sits one layer below everything your roadmap is about. It never demos well on its own, it never ships as a feature, and it will occupy your best content people for quarters. Every team that has built adjacent infrastructure in this market has said the same thing afterwards: it pulled focus from the core product.
And then it decays. A layer of derived differences is only correct on the day it was derived. A ruling lands, a guidance note is revised, a transposition deadline passes, and every assessment that leaned on the old text is now quietly wrong, with no signal that it changed. Building the layer once is the smaller half of the problem. Keeping it true is the permanent half.
The alternative: point us at your content and your domains. We run the process end to end. You host the finished layer and build product on it. The machine stays with us; the layer lives with you, and we keep it alive.
Operated, not licensed.
A partnership has three moving parts and no ambiguity about who owns what.
Your corpora, your publisher rights, your jurisdictions, your priority order. Your upstream licences stay yours; nothing about them changes.
The full process from source text to a validated, source-chained functional layer: extraction, ontology, linking, derivation, validation, and the propagation that keeps it current. This runs on our side. It is not a tool we hand over and not a methodology we document for you.
The finished layer, in your infrastructure, under your brand, queried by your product. Delivered as a query interface against structured relations, not a file drop. Your users never leave your platform, and every answer they see carries links back to the primary sources they already trust.
The relationship is continuous because the law is. As the layer’s dependencies change, the changes propagate and the layer updates. This is a running service, not a delivery date.
Scoped exclusivity is available: by vertical, by jurisdiction pair, or as a first-to-market window. Not general exclusivity: the layer is infrastructure, and infrastructure that only one platform can reach is worth less to everyone including its holder.
At the functional level, this is what the layer contains and what your product can ask it.
Nodes: rules, not documents. The unit is the normative provision as it does its work: what it requires, of whom, on what trigger, above what threshold, with what consequence. Each node sits in its own jurisdiction, in its own language, in its own legal tradition. Nothing is flattened into an interlingua and nothing is translated into a lead jurisdiction’s assumptions.
Anchors. Shared reference points that jurisdictional graphs attach to: a regulation, a directive, a model rule, a standard. Anchors are what make a link between two national provisions defensible rather than asserted: both hang from the same instrument, and their distance from it is measurable.
Relations: provision to provision. Not document similarity. A directed, typed link between one rule and its functional counterpart elsewhere, including the cases where the counterpart is partial and the cases where there is none. Absence of a counterpart is data, and the layer stores it as such.
Differences: derived, not described. Every relation carries the concrete divergences between the two provisions: thresholds, conditions, who is covered, what procedure applies, what actually happens when the rule bites. These are computed from the structured attributes of both sides, which means they are reproducible, comparable across pairs, and countable. Your product can query them, sort them, and display them as a table rather than a paragraph.
Source chains. Every statement in the layer resolves to a citable primary source: the provision, the ruling, the guidance. One statement, one chain. This is what lets your product show its work when a user asks why, and it is what makes an answer defensible when it is challenged.
Change flow. Every node knows what depends on it. When a source changes, the layer determines which assessments in which jurisdictions are affected and how strongly, with a reason attached to each. This runs deterministically and returns in microseconds, so it can be a live surface in your product rather than a monthly report.
Coverage and freshness are stated per jurisdiction and per anchor family, and the layer reports its own currency.
(How the layer is built is not on this page. What it holds and what it answers is.)
Your data is never used to train models. The layer is a retrieval and reference surface: structured relations and verifiable links back to primary sources, priced by use.
That is not a concession made under pressure. It is the same position most of this market has already taken publicly, and for the same reason. What a lawyer needs from a legal answer is not plausible text. It is an answer they can check against the actual source, in the actual jurisdiction, and stand behind afterwards. A layer built for training would be built wrong for that job.
Upstream provenance is treated the same way. Where the layer builds on your rights, those rights stay bounded by your agreements, and the chain of authority behind every source is part of what the layer records.
Four numbers, all measured end to end on live corpora, all reproducible.
98 % vs 25 %. The question your users actually ask: what is the equivalent of this provision in that country? Answered through the layer at 98 % precision at rank 1, and at 100 % where the citation is unambiguous. The same question, same corpus, answered by cross-lingual embedding search, which is what every platform in this market has today: 24.9 %, against a chance floor of 14.3 %. Near random. This is not a gap a better embedding model closes; the relation isn’t in the text to be found.
96–99 % on blind reconstruction. Rebuilding rule correspondences across 24 languages with the answer key withheld, on pairs that are almost all cross-lingual: 96 % from raw text, 99.35 % with the layer’s structure, errors down to a sixth. Language is not the failure mode. Function is.
0 fabricated statements out of 2,079. Independently and adversarially reviewed across two bodies of law. Every statement carries its source chain to primary source.
Recall 1.0 on change propagation. Measured against real events, including gaps human reviewers had missed. One CJEU ruling propagated across twelve national regimes: three high exposure, four to verify, four unaffected, with the reasoning stated per country, deterministically, in microseconds.
The layer is not a dataset with a domain baked in. The same engine has been run on two bodies of law that share almost nothing:
Pillar Two. 66 jurisdictions, 24 languages, computational, threshold-driven, built on OECD model rules.
GDPR opening clauses. 12 jurisdictions, 10 languages, non-computational, discretion-driven, built on a directly applicable regulation where divergence exists only where the text permits it.
The second was built, validated and measured in under an hour, and scored higher than the first. What changed between them: a reader for the new source format and two domain fields. What did not change: the graph, the matching, the derivation, the ontology.
That is the relevant fact for a platform. Your first anchor family is a project. Your fifth is a configuration.
Blind reconstruction
Measured across 66 jurisdictions in 24 languages.
Blind reconstruction
Measured across 12 jurisdictions in 10 languages.
Nobody’s first agreement here is comprehensive, and it shouldn’t be.
Phase one: one anchor family. Pick the instrument where your users are already losing time across borders, and the jurisdictions where you have the strongest rights. We build the layer for that family, on your content, end to end. It is narrow enough to finish and complete enough to ship.
Phase two: measured against your stack. Before anything goes into your product, the layer is scored on real questions from your domain against what your platform answers today. You see the error matrix. If it doesn’t beat what you have, you have lost a phase, not a roadmap.
Phase three: it goes live and stays live. The layer moves into your infrastructure, propagation switches on, and the next anchor families follow on your priority order, not ours.
On your existing suppliers: this layer sits above access and indexing. If you already buy crawling, extraction or search infrastructure, none of that is displaced. The layer consumes what those deliver and adds the dimension they don’t produce. It does not replace your publisher agreements either. It makes them add up.