Most compliance chatbots are a language model with a statute pasted into the prompt, and they fail in the way language models fail: confidently, on the details, where you are least likely to check. A commencement date is exactly that kind of detail — 2 August 2026 and 2 December 2027 are both plausible-looking answers, and only one of them is the one your programme has to be planned around.

So this is built the other way round. The model does the one thing it is genuinely good at — reading your description and extracting structured facts about the system — and every legal consequence of those facts is looked up in code. Then a second deterministic stage re-reads the model's own prose and flags any article or date that did not come out of that lookup.

0/900

What just happened

Four stages, and the trace panel labels each one, because the split is the entire point:

Read — model
Your description goes to a language model, which returns structured facts about the system: does it interact with people, does it generate content, does it fall in one of the enumerated Annex III areas, are you the provider or the deployer. It answers from a closed set — an invented ninth category fails schema validation rather than reaching the next stage.
Apply — rulebase
Those facts go into a pure function over a table held in the repository. It decides the tier, selects the duties, and resolves each one's commencement date. No model, no network, no randomness. Run it twice with the same facts and you get the same answer.
Explain — model
A model writes the prose, and is given the decided tier, the selected duties, and an explicit list of the only article references and dates it is permitted to mention.
Check — rulebase
The prose is then scanned for every article, annex and date it contains, and each is compared against what the rulebase actually produced. Anything else is flagged in place. A guardrail you can watch fire is worth more than one you are asked to believe in.
The honest limits

This is a triage tool and a demonstration, not legal advice and not a conformity assessment. It reads a paragraph you wrote, not your system.

The classification is only as good as the description — a system that sounds like recruitment software will be treated as recruitment software. Annex III turns on the actual use case, the Act has definitions this table compresses, and the real answer to a borderline case involves a lawyer looking at what the thing does. What this can do reliably is tell you which questions are the ones that matter, and give you the dates without inventing them.

The same rulebase is what the dates on the omnibus breakdown are drawn from, so the page and the agent cannot drift apart. What this site itself has to comply with is on the conformity page— including the row that still says “Partial”.

The other half of thisThis site's own obligation ledgerThe same law, applied to the AI running here — what is met, what is partial, and what is deliberately not claimed.