One engine across ten markets
An advertisement changes status at a border. A risk warning that satisfies one regulator can be too short, too late on the page, or in the wrong words for the next. The claim that returns are guaranteed breaches both the Dubai and the EU rules, and the provision to cite differs in each.
Licentium Atlas runs a single review engine across every market rather than a separate product for each. One engine keeps the reasoning, the severity scale and the output shape consistent, so a compliance lead comparing a Dubai review with an EU one reads like for like. It also raises a problem a single-market tool never meets: the engine has to bind its choice of source material to the market the user picked, and hold that binding at every step of the review.
The output to design against is a finding that quotes a real provision from the wrong jurisdiction. That is the most dangerous result a multi-market screener can produce, because it has the shape of a correct answer: a named instrument, an article number, a sentence that reads like law. A Dubai campaign reviewed against MiCA Article 7 returns something a busy reader can accept without pausing, and nothing in the finding signals that the instrument has no application in the Emirate. The same risk runs inside a single country, where Dubai, the DIFC and the federal regime each have their own rules for the same advertisement.
Two rules follow from that, and we hold both. Source material resolves from the market the user selected, at the point of retrieval, and no other path may feed text into the review. A market whose law we have not yet structured returns no citations at all rather than the nearest available body of law.
Silence is the safer failure. An officer who receives a finding with no citation knows to do the research. One who receives a confident citation from the wrong country may never check it. We also made the coverage signal load-bearing: a market cannot be marked ready until its material exists and its evaluation passes, which stops the claim of coverage running ahead of the corpus behind it.
What the Dubai rules require
The Virtual Assets Regulatory Authority issued the Regulations on the Marketing of Virtual Assets and Related Activities 2024, in force from 1 October 2024. They bind every entity marketing in or targeting the UAE, domestic or foreign, whether or not VARA licenses it.
Marketing Regulation I.C.2 requires marketing to be fair, clear and not misleading in both substance and presentation. I.C.2.e prohibits stating or implying that returns are guaranteed. I.C.2.h prohibits implying urgency or creating a fear of missing out on future appreciation. I.B.3 permits marketing of a VA Activity only by a VASP licensed by VARA for that activity, or on its behalf and with its approval. Schedule 1 sets a fine of up to AED 10,000,000 for a breach of I.C.2, doubled for a repeat within one year. (VARA Rulebooks, rulebooks.vara.ae, retrieved 13 August 2026.)
Jurisdiction inside the UAE divides. VARA regulates Dubai including its free zones and excluding the Dubai International Financial Centre. Copy aimed at Abu Dhabi, the DIFC or the wider UAE can engage ADGM FSRA, the DFSA or the federal SCA. A tool that answers "the UAE" without naming the regulator hands the officer something they cannot act on.
What a grounded finding looks like
We keep one test advertisement for regression work. It reads:
Stake $TOKEN today. Guaranteed 40% APY with zero risk for early supporters. Limited spots, don't miss out! Your funds are fully insured and we are regulated in Dubai.
Reviewed against UAE rules, Atlas returns five findings and blocks publication. The guaranteed-return claim draws I.C.2 together with the Schedule 1 fine. The claim to be "regulated in Dubai" draws I.B.1 and I.B.3, which govern who may market a VA Activity at all. The scarcity language draws I.C.2. Twenty source links resolve to VARA's own pages.
Links point to the section containing the provision rather than to the paragraph. VARA publishes no per-paragraph anchors. An anchor we invented would look precise and land the reader at the top of the page, so we did not invent one. We confirmed the section map two ways before shipping it: against VARA's published navigation, and against VARA's own internal cross-references inside Schedule 1, which point at the same sections.
How the accuracy was measured
We wrote 164 evaluation questions before tuning anything, in three categories: questions the rulebook answers, questions it cannot answer (custody, tax, employment, another emirate's regulator, activity outside the UAE), and questions phrased to tempt the system into quoting a different jurisdiction's law.
Results from internal runs between 11 and 13 August 2026. Twenty-seven of 29 answerable questions cited an expected provision, or 93 per cent. Across roughly 465 questions in three full runs, no answer cited an instrument outside the one in scope. We report proportions with Wilson intervals, because a percentage from 25 trials carries an interval wide enough to mean very little on its own.
For questions the rulebook cannot answer we publish a range rather than a single figure. Two behaviours are correct: the system can stay silent, or it can decline and quote the boundary provision that puts the question outside these rules. The second is more useful to a reader, and it requires citations, which is why an early version of our own scoring counted it as a failure. A program can confirm silence, and can confirm a decline when the answer uses boundary language. It cannot separate a decline worded in a way we did not anticipate from an answer that should have been refused. The harness prints a floor, a ceiling, and every unresolved case in full for a person to read.
Six problems, and what we changed
1. A retrieval rule starved a single-instrument corpus. Our result diversifier capped how many passages could come from any one document, which stops a single long text crowding out everything else. Every provision of the VARA marketing rules sits inside one document. Three provisions of roughly twelve reached the reviewer and the rest were discarded before it saw them. The rule now diversifies by provision reference when a market's material is a single instrument, and keeps the per-document cap for markets with many.
2. Paraphrase defeated citation. Early UAE reviews produced findings with no source attached. The reviewer paraphrased less familiar text, and a verbatim check then discarded the citation as unmatched. We moved the source of truth from extracted text fragments to the structured provisions themselves, with every quotation verified against the stored wording. Findings carrying a source went from one in five to five in five on the test advertisement.
3. A deadline set below the cost of the operation. Our first live run grounded one finding out of five. The four failures were not the service failing: its own log showed every request answered, in 10 to 22 seconds. Our client gave up at 12 seconds and discarded work already done. We raised the deadline and issued the requests concurrently, which costs about one request's time instead of five, and that phase now completes in about 22 seconds. Every component had reported success; only a comparison of two logs showed the loss.
4. A penalty figure taken from the wrong instrument. One finding stated exposure to fines of up to AED 50 million while displaying, in the same card, the schedule capping that breach at AED 10,000,000. AED 50 million is a real VARA figure. It sits in Schedule 3 of the 2023 Regulations, for corporate breaches of the conduct rulebooks, and it is not the marketing penalty. The reviewer had produced the number from prior knowledge rather than from the retrieved text. Any figure stating a legal consequence, a fine, a threshold, a deadline, a retention period, must now come from the source in front of it. Figures quoted from the advertisement under review are exempt, because an advertised yield is the evidence for the finding. Reading the output against the primary source found this. No test and no evaluation did.
5. Two of our own instruments were wrong. One printed a clean result after zero cases had executed, which is a green light for an experiment that never ran. The other scored a refusal as a failure whenever the system declined by explaining, and that understated our result by 26 points: 73 per cent reported against about 99 per cent measured after correction. Both are fixed, and the harness now stops with an explicit failure when nothing was measured.
6. A pass that meant "we did not check". Internal testing found a path where a failed engine call returned no findings, and no findings scored as a clean pass. We separated two states that an empty result cannot distinguish: the engine ran and found nothing, and the engine did not run. The second now returns an error the reader can see. A screener that reports a pass while its engine is down leaves a reader unable to tell a clean result from an absent one, and only one of those is safe to publish against.
