What this tool does not do

It is not legal advice

This is a preliminary assessment. It is not legal advice, not a conformity assessment, not a certification, and not a registration. It does not create a defence and it will not satisfy a regulator on its own. A result that points anywhere serious should go to qualified EU regulatory counsel.

No rule has been reviewed by a named lawyer

Every rule in the current set carries the status "requires qualified review" with no reviewer and no date. The rules were read from the primary text carefully, but careful is not the same as reviewed. Nothing in the build can promote a rule to reviewed.

The legal text is the version as published, and amendments are unverified

The rules were read from the Official Journal text of Regulation (EU) 2024/1689 published on 12 July 2024. That document cannot show later amendments.

Secondary sources report that an amending regulation deferred the high-risk application dates, moving the Annex III regime to 2 December 2027 and the Annex I regime to 2 August 2028, and gave a short grace period for the Article 50(2) marking duty. Those reports could not be checked against primary text from the environment this tool was built in, because the network policy blocked access to EUR-Lex.

So the tool shows the date read from the Regulation as published, and flags the reported change beside it. It does not use an unverified later date to tell you an obligation has not yet arrived. Before relying on any date here, check the consolidated text on EUR-Lex.

The two existing checkers were not inspected

The network policy in the build environment blocked both the European Commission checker and the Trail checker, along with EUR-Lex and other Commission hosts. No screen, question, branch or result of either tool is described anywhere in this project, and no comparison is made to behaviour that was not observed. The design here is argued from the Regulation and from first principles, not from theirs.

Commission guidelines are not incorporated

Article 6(5) required the Commission to publish guidelines on the practical implementation of Article 6, with examples of use cases that are and are not high-risk, no later than 2 February 2026. Those guidelines, and any others issued under Article 96, could not be retrieved. Where a classification turns on a line those guidelines would draw, the tool says so on the finding.

Areas deliberately left out

• Which conformity assessment procedure applies under Article 43, and which notified body to use. Articles 28 to 39, and Annexes VI and VII.
• The content of technical documentation under Annex IV, and the quality management system under Article 17.
• Registration mechanics in the EU database, Articles 49 and 71.
• Penalties and their calculation, Articles 99 to 101.
• Regulatory sandboxes and the procedure for testing in real-world conditions, Articles 57 to 63.
• Member State national law, which Article 5(5) and parts of Annex III expressly defer to, and which differs between countries.
• Every jurisdiction outside the European Union.

Interpretations the tool had to make

Article 6(3) is read conservatively

The first subparagraph requires that the system not pose a significant risk of harm, including by not materially influencing the outcome of decision making. The second subparagraph says that applies where any of four conditions is fulfilled. Those two sentences can be read as a single test or as a gateway plus a test.

This tool takes the stricter reading. It requires both a condition and an absence of material influence, so a system whose output drives a decision stays a high-risk candidate even where a condition is arguably met. If the looser reading is right, the tool will sometimes say high-risk where a court would not. That is the direction the error was chosen to run in, and a reviewer should decide whether it fits your facts.

The Annex III mapping is an interpretation

The tool asks what a system does in ordinary words and maps that onto Annex III entries. That mapping is a reading of the entries, not a quotation of them. The reasoning trace shows which entry was matched and from which answer, so a reviewer can disagree with a specific step rather than the whole result.

Open source is kept in two separate boxes

Article 2(12) covers AI systems released under free and open-source licences, and switches off for high-risk systems and for anything under Article 5 or Article 50. Article 53(2) and Article 54(6) cover general-purpose AI models, are narrower, and disappear entirely for a model with systemic risk. The tool never generalises one into the other.

The result is only as good as the answers

Every conclusion follows from what you entered. An answer that is wrong, or that describes how a system is meant to work rather than how it does work, produces a conclusion that is wrong in the same way. The tool has no way to check your answers against reality, and does not pretend to.

If something here looks wrong

Each finding names the rule, the article, the page and the facts it used. That is enough to disagree with it precisely. The sources page lists every rule in the set against its place in the Official Journal.