Compliance & regulatory intelligence

One rule. Every market. Every change.

They match context. We match function.

Your platform already tells a customer which regulations apply to their product in 195 countries. It does not tell them that the requirement they just met in France is the same requirement waiting for them in Poland: three differences, each one named, each one sourced.

Homolog builds that layer. On top of your content. Under your name.

The problem

Your customers don’t have a feed problem. They have a sameness problem.

Ask a compliance leader why they consolidated from a handful of monitoring subscriptions down to one, and the answer comes back in some version of the same sentence: too many services were sending the same update.

That complaint is usually read as a procurement problem: too many vendors, too many logins, too much noise. It isn’t. It’s a diagnosis of what’s missing from the category.

If several services deliver “the same update” about several different countries, someone has already done the hard work in their head: recognised that a French labelling obligation, a German one and a Polish one are one requirement wearing three national costumes. The human made the link. No system did. And because no system made it, nothing knows what to do when one of those three moves.

The arithmetic behind that is not survivable by staffing. A single product entering a single market can carry roughly a thousand applicable regulations. Global regulatory movement for a typical product runs in the low hundreds of changes per month. And most compliance functions do this work with five people or fewer, still primarily out of spreadsheets. Meanwhile only a small minority of multinational groups will claim full compliance across all their entities. Most know they aren’t there and say so out loud in surveys.

More feeds do not close that gap. More feeds widen it. What closes it is knowing which of the incoming changes are the same thing they already handled somewhere else, and exactly where the sameness stops.

Age of consent for information society services (member states may lower 16 to not below 13)

The gap

What nobody sells: the link between the rules

The intelligence market has, correctly, taken applicability: matching a customer’s product, claims, ingredients and markets against the rules that touch them. That’s a real product and a hard one, and the leading platforms do it well across a lot of countries at once.

But “at once” is the tell. Coverage is parallel, not relational. Every market is assessed against the customer’s product independently. Nothing in the stack holds the statement:

This is the same requirement you already satisfy in France. Here are the three ways Poland differs: the threshold, the documentation you must retain, and who you notify. Here is the source for each.

Two consequences follow, and your customers feel both.

Evidence gets re-produced instead of re-used. Work already done, already documented, already defensible in one market gets rebuilt from scratch in the next, because nothing certifies that the two obligations are functionally the same, or specifies precisely where they part company.

Change propagates in one direction only. Continuous reassessment across the category is triggered by the product changing: a new ingredient, a new claim, a new supplier, a new market. When the rule changes, there is no mechanism that walks the change outward to every equivalent obligation in every other country and grades the exposure.

That second one is where the damage lands, because regulatory change doesn’t only arrive as a new national statute. In the EU it arrives as a single ruling that lands on twelve national rulebooks simultaneously.

A real event

One ruling. Twelve countries. Graded, with reasons.

CJEU · C-34/21 · 30 MAR 2023 A judgment on national employment-data rules. Art. 88 — employment data
DK PT SE ES FI IT PL AT FR IE NL DE
  • 3 High exposure
  • 4 Check
  • 4 Not affected

Here is what the layer does, on a real event.

The Court of Justice of the European Union rules on how far national rules made under a GDPR opening clause may go: specifically, on national provisions that do little more than restate the Regulation itself. One judgment. It names one member state.

A conventional monitoring feed does what it is built to do: it reports a judgment about that one member state. The other eleven national provisions built on the same opening clause sit in eleven separate country libraries, and nothing connects them to the event.

Homolog holds those twelve national provisions as one anchor: the same function, twelve forms. The ruling lands on the anchor, and the anchor lights up the whole set:

  • 3 countries: high exposure. In one, the national provision lacks the safeguards the ruling requires. In another, the national supervisory authority had already struck down a provision on this exact ground. In a third, a whole parallel track for collective agreements is now in question.
  • 4 countries: check. Detailed national regimes that engage the same clause and need a human read against the new standard.
  • 4 countries: unaffected. Countries that never used the opening clause in the first place. Not “no result found”: affirmatively unaffected, with the reason stated.
The layer, kept alive · real changes, propagated Last change propagated 21 JUL 2026
2023202420252026 ART. 88 ART. 3.2.3 ART. 4.1.3 ART. 9.1
ART. 8830 MAR 2023 · Court judgment

CJEU ruling on national employment-data rules (GDPR Art. 88)

One judgment; eleven other member states' provisions flagged and graded: three high exposure, four to check, four affirmatively unaffected.

ART. 3.2.321 JUL 2026 · Commentary revision

OECD Commentary review finding — arm's length adjustment (Art. 3.2.3)

A Commentary-driven revision propagated to all 59 jurisdiction assessments on the anchor; the human-curated update set was reproduced in full, plus six cells the curators had missed.

ART. 4.1.321 JUL 2026 · Inclusive Framework development

Inclusive Framework development — qualified refundable tax credits (Art. 4.1.3)

An Inclusive Framework-level development: every dependent assessment flagged the day the note landed.

ART. 9.121 JUL 2026 · Safe-harbour extension

Transitional CbCR Safe Harbour extension (Art. 9.1) — Inclusive Framework level

A date-parameter change on the widest anchor in the corpus: 59 core assessments flagged, 774 dependent cells marked check: parameter changes propagate as verification, not as re-assessment.

  • 21 JUL 2026 Transitional CbCR Safe Harbour extension (Art. 9.1) — Inclusive Framework level
    Safe-harbour extension · Art. 9.1 — transitional CbCR safe harbour

    A date-parameter change on the widest anchor in the corpus: 59 core assessments flagged, 774 dependent cells marked check: parameter changes propagate as verification, not as re-assessment.

  • 21 JUL 2026 Inclusive Framework development — qualified refundable tax credits (Art. 4.1.3)
    Inclusive Framework development · Art. 4.1.3 — qualified refundable tax credits

    An Inclusive Framework-level development: every dependent assessment flagged the day the note landed.

  • 21 JUL 2026 OECD Commentary review finding — arm's length adjustment (Art. 3.2.3)
    Commentary revision · Art. 3.2.3 — arm's length adjustment

    A Commentary-driven revision propagated to all 59 jurisdiction assessments on the anchor; the human-curated update set was reproduced in full, plus six cells the curators had missed.

  • 30 MAR 2023 CJEU ruling on national employment-data rules (GDPR Art. 88)
    Court judgment · Art. 88 — employment data

    One judgment; eleven other member states' provisions flagged and graded: three high exposure, four to check, four affirmatively unaffected.

Twelve flags, each with a reason a lawyer can argue with. Not twelve alerts. Not one alert about one country.

And the same run exposed something else: of the twelve national provisions, only one had any coverage of the ruling at all. Eleven gaps, surfaced by the machine, in material that already existed.

The propagation itself is deterministic and takes microseconds. It is not the model deciding what’s related: it’s the structure of the layer. The judgment about how much each country should worry is where your experts spend their time, on eleven pre-framed questions instead of a blank page.

The layer

What the layer holds

Each jurisdiction keeps its own knowledge graph, in its own language, on its own sources. The graphs are linked through shared anchors: an EU regulation or directive, an OECD model rule, a standard. The links carry the derived differences.

For any two jurisdictions on the same anchor, the layer answers:

  • Is this the same requirement? Functional equivalence at rule level: provision to provision, not term to term, not topic to topic.
  • Where exactly does it differ? Thresholds, conditions, legal consequence, procedure. Derived, not narrated.
  • On what authority? A source chain per statement, down to the provision. Every sentence in the layer is traceable to the primary source it came from, in the original language.
  • What just changed, and who does it reach? Change enters at the anchor and propagates to every dependent assessment, with per-country reasoning attached.

The layer is living. It is not a snapshot delivered once and stale by the next quarter. The machine keeps running, and the layer keeps moving with the law.

Your customers keep asking for verifiability before they’ll trust any of this: they want to check the answer against the source before they rely on it. That requirement is met at the sentence level, by construction.

The offer

We operate the machine. You keep the content.

Homolog is a technology, not a dataset. We are not selling you a competing corpus, and we are not asking you to swap out the regulatory content your customers already buy you for.

You point at your content and your domains. We run the whole process: extraction, ontology, linking, validation, propagation. We hand back a living layer that sits on top of what you already own. You host it. You brand it. Your product gains the difference-and-propagation dimension without your team building it.

What that means in practice:

  • Your corpus becomes the moat it already is, only relational. Depth built over decades is exactly what a functional layer needs underneath it. We add the dimension your library doesn’t have; we don’t replace the library.
  • No new analyst headcount per jurisdiction pair. The expensive part of relational coverage is combinatorial. The machine does the combinatorics; your experts validate outcomes, not raw pairs.
  • No training on your data. Partner and client content is never used to train models. The layer is retrieval and reference, with verifiable sources.
  • The process never leaves the house. We operate it: it isn’t licensed out, and it isn’t a black box handed over with a manual.
  • Scoped exclusivity is available where it makes sense: a vertical, a jurisdiction family, a time window.

The commercial shape follows the work, and every engagement is negotiated on its own terms.

Evidence

Measured, not asserted

Everything below is end-to-end measurement of the working system, not a benchmark on a component.

96–99 % Blind reconstruction of rule correspondences, across 24 languages. Almost every test pair is cross-lingual. The system never sees the answer key.
80–98 % Correct answers to “what is the equivalent of § X in country B?”, against 17–25 % for embedding search, the industry default. That lower number is close to chance.
0 of 2 079 Fabricated statements, under independent adversarial review, across two areas of law. Every sentence carries a source chain to primary source.
Recall 1.0 Change propagation against real events. It found gaps humans had missed. Deterministic, microseconds.
One engine, many domains Pillar Two across 66 jurisdictions and 24 languages; GDPR opening clauses across 12 jurisdictions and 10 languages. The second one was built, validated and measured in under an hour.

That last row is the one that matters for a partner conversation. The cost of adding a domain is the question you’ll want answered before you commit anything, and it has been measured rather than promised.

First engagement

Where it starts

Phased is normal. Nobody’s first engagement is 195 countries.

  1. One anchor family. Pick a regulation or standard your customers actually get asked about across borders, and a handful of jurisdictions where you have strong content.
  2. We build the layer on it. Extraction, ontology, linking and validation, run on your material, in the original languages.
  3. You test it against your own experts. Not our claims: your people, your cases, your judgment on whether the derived differences hold.
  4. We replay real change through it. A ruling, a guidance note, an amendment from the last twelve months: you see what it flags, what it grades, and what it says is unaffected.

If it doesn’t hold up at step 3, there is nothing to talk about. If it does, the layer extends anchor by anchor, and the machine keeps it alive.