An answer is a single event. A register is a state with a date on it. Asking a model — or building an agent that asks one for you — is an excellent way to understand the EU AI Act; it is not a way to show someone, long after the fact, what you decided and what you decided it against.
Both columns are artefacts, not tools: a piece of text on one side, a record on the other, and what each one carries.
| What a reviewer asks about | An answer | A Complipath record |
|---|---|---|
| Repeatability | One question, one session, one piece of text. Asked again in different words, it is a second piece of text. | The engine is deterministic: the same answers produce the same outcome, every time. |
| Which version of the law it was decided against | Whatever you wrote down beside it. | Stamped on the record: the base act and every amending act, by name. |
| Which version of the tool decided it | Whatever you wrote down beside it. | Stamped on the record and carried in the export: the engine version. |
| What happens when the law changes | A saved answer is a document. Nothing re-reads it, and this is as true of ours as of anyone's. | The register lists which assessments cited a provision that moved — and which did not. |
| Who decided, and when | Whoever kept the transcript. | A named person and a timestamp, recorded against the assessment. |
| Where the reasoning stops | Where the answer stops. | The engine declines where the Act requires judgement rather than guessing, and the provisions it declined are on the record. |
| What you hand a reviewer | The transcript, and your account of it. | A JSON file per assessment and a CSV of the whole register, each stamped with when it was produced. |
Anything about what a language model can or cannot do. We are not going to describe a capability we have not measured, in a product that is not ours, in order to sell one that is. The left column describes a piece of text — any piece of text, ours included — and the right column describes what a record carries.
Ask a language model the same question twice in different words and you get two answers. For research that is a feature — it is how you find the argument you had not thought of. It stops being a feature when somebody asks what you based a legal classification on, because then the second answer is not a second opinion. It is the absence of a first one.
Complipath's classification engine is deterministic. The same answers produce the same outcome every time, and the outcome is stamped with the engine version that produced it and the version of the law it was decided against. That is not a claim about accuracy. It is a claim about repeatability, and it is the one property a record needs before anything else about it matters.
Regulation (EU) 2026/1744 amended the 2024 text in force, and it will not be the last. Nothing you were told before that happened updated itself when it did. A register that records which version of the law each assessment answered can list exactly which entries cited a provision that moved — and, just as usefully, which did not. That is a question you can ask a register. There is nobody to ask a conversation.
For one system, once, a language model is excellent — better than a form, and faster than reading the Annexes yourself. If you have three AI systems and a founder who remembers every decision, you do not need us yet. The difference shows up at forty systems, when the person who made the calls has left and somebody asks you to show them.
The neighbouring case — a spreadsheet and one person's time — is argued at what in-house compliance work does well, and where it stops. What our own engine will and will not decide is listed as a limit rather than a feature.
Complipath is EU AI Act compliance software for AI-heavy software companies without a compliance team — an AI system register, deterministic risk classification, the obligations that follow, and the evidence behind every decision.
Complipath is built by Yobel Tzegai in Gothenburg, Sweden.
Complipath provides legal information, not legal advice. Every guide cites its source on EUR-Lex — Regulation (EU) 2024/1689, and Regulation (EU) 2026/1744 where that has amended it; where the law is still settling, the guide says so.
We measure page views with Vercel Web Analytics. It uses no third-party cookies. Visitors are identified by a hash derived from the incoming request, which is discarded after 24 hours, and no identifier is stored that could follow a visitor to another site. What is collected: the time of the visit, the URL, the referring page, filtered query parameters, city-level location, operating system, browser and device type.