Authorisation lag
When asked whether a named crypto firm is authorised under MiCA, do AI answer engines give the answer ESMA's register gives, and how does accuracy change with the time since that status became visible in the register?
Results
None yet. The protocol below was written before any data was collected; the figures go here when the analysis is done, whatever they show.
What the study asks
MiCA's transitional arrangements for existing crypto-asset service providers ran out on 1 July 2026, the final EU-wide backstop. ESMA's interim register listed 361 firms at the 2 October 2026 snapshot: 149 authorised before 2026, 141 in the first half of 2026, and 67 since July. A separate list names 173 entities as non-compliant.
The study asks four AI model-and-retrieval configurations, twice each, whether a named firm is authorised, and codes the answer against the register as it stood when the answer was given. The primary code answers one question only, whether the present status stated is right; dates, authorities and service scope are scored separately, so a right licence with a wrong month is not counted as a wrong licence.
How it is measured
- Two clocks. The register records the legal date of each authorisation, but the register itself is republished weekly and a firm can appear in it weeks after that date. Archived copies of the register give the date each firm was first observed in it. Accuracy is analysed against time since that first observed appearance, with the legal date as the secondary clock.
- Cohort. 90 firms, 30 from each period, drawn from the register with a recorded seed: 20 crypto-native firms and 10 regulated financial institutions in each, since half of the post-July entries are German cooperative banks that an engine may never have seen described as crypto firms. The type contrast is reported before anything else is interpreted. A separate strand of ten named entities from the non-compliant list, chosen for recognisability and declared as a hand-picked sample, asks whether the engine warns.
- Prompts. One branded question to every firm, a consumer phrasing to a 30-firm subset, and two unbranded questions for eight countries, scored against the register's passporting field.
- Configurations. The same four as the first study, reached through one API gateway, including the matched memory-only control. Results describe those configurations, not the consumer products that share their model names.
- Two waves, five weeks apart, to see whether answers change.
- Claims. Descriptive accuracy and its relation to information age only. Nothing about why an engine is wrong, and no attribution of change between waves to search indexing rather than model updates.
Where it stands
Protocol v1.0 was frozen on 7 October 2026 after an external design review and a 48-call pilot. Wave 1 is scheduled for 12 October; wave 2 five weeks later.
Disclosure
The author works at Tesseract Investment Oy, which is on the register. It is not excluded: it sits in the sampling frame like any other firm and is flagged in the data. It was not drawn for the current cohort; if a later draw includes it, its result is reported wherever it lands.