Describe a technology. Osso asks only for the figures the EU dual-use list turns on, classifies it, and finds the export authorisation that applies.
Interview
Download the memo first if you need it: a new case clears this one.
This assistant is an AI (Claude, by Anthropic). Its answers can be wrong
and must be reviewed by a qualified professional before they are relied on, used or shared.AI (Claude, by Anthropic): answers can be wrong. A qualified professional must review them before any use.
Annex I, by category
What it does
Annex I of Regulation (EU) 2021/821 is the European Union's list of dual-use items: goods, software and
technology that have civilian uses but could also serve military ends, and that need an export authorisation
to leave the EU. The Osso Export Classifier tells you where a product stands against that list.
You describe your product in plain words. The assistant asks the questions the regulation turns on, such as
a frequency, a resolution, a depth rating or an output power, and you answer them from the datasheet. You get one
card with one of three outcomes: listed, not listed, or needs expert review. Every provision the answer rests on
is quoted word for word from the consolidated text of the regulation, and the card says which export authorisation
applies. A memo of the case can be downloaded.
Common questions
What is a dual-use item? An item that can be used for both civilian and military purposes.
Annex I of Regulation (EU) 2021/821 lists them by category, from nuclear materials to information security.
What does "needs expert review" mean? The description did not settle the question, for example
because a figure is missing. The tool does not guess: it says so and shows what to check.
Is it free? Yes. There are 5 free chats a day, and browsing the whole of Annex I needs no chat.
Is it legal advice? No. Every result must be reviewed by a qualified professional before it is
relied on. How often the tool is right, and where it is not, is published in
the benchmark.
Your data and security
We do not train any AI model on what you type.
By default we keep no text of your conversation. When a case ends on a classification we keep only anonymous
outcome data (the entries found, the status, the destination and licensing outcome, the number of turns and the
timings), with no IP address, for 12 months.
Saving a case is your choice: a box on the memo, unticked unless you tick it.
No accounts, no cookies, no analytics and no third-party scripts. The fonts are served from this site.
Anthropic, which provides the AI model, processes what you type to write the replies. Its
Commercial Terms say that it "may not train models
on Customer Content from Services". It deletes API inputs and outputs within 30 days, save exceptions it
describes, and keeps them for up to 2 years if its safety systems flag them.
The page and the assistant answer only over HTTPS, and the page runs under a strict Content-Security-Policy.
The accuracy benchmark, with every case and every known failure, is public. The
assistant is refused from countries where Anthropic does not offer its service or sanctions forbid providing it.
The details, and your rights, are in the Privacy notice.
How accurate is it?
92 out of 100 real products classified correctly in the published benchmark, 30 September 2026: 37 of 39 deliberately hard edge cases and 55 of 61 everyday products.
Correct means the status matches (listed, not listed, or needs expert review), every expected entry is on the
card, and the licensing outcome matches where one is expected. Each expected answer was drafted from the product's
public datasheet by one AI model and checked against the text of the regulation by a second; the calls they
disagreed on were decided by the operator, a lawyer. The products are described without their names, so the
assistant cannot lean on a remembered answer.
Every case, expected answer and run is published in the benchmark.
Known gaps
All eight misses were cautious ones: the tool said needs expert review where the expected answer was a definite listed or not listed. None gave a wrong listed or not-listed result.
Where no threshold table exists yet (15 of the 100 cases), the AI interview decides, as before, and is less predictable than the tables.
The same product does not always get the same answer: between two runs of the same cases, some flipped each way. Read the score as about 92, not exactly 92.
The score is on the benchmark's own 100 cases, and its earlier misses were used to build the tables. Every card must be reviewed by a qualified professional before use.
