A horse crossing a border is a regulatory object before it is cargo.

Destinations do not all work the same way, and the system has to know which world it is in before it can judge anything.

Inside the EU and the wider European area, a movement travels on the intra-Union certificate and goes through the European trade control system. For a third country with a bilateral agreement in place, there is a published national protocol and the requirements are known in advance. Those two cases are the comfortable ones.

Then there are the countries with no state-to-state agreement at all. There the movement proceeds on a private arrangement between the company and the authorities on each side: an import permit obtained from the destination, and a one-off certificate issued for that shipment rather than drawn from a standing protocol. There is no published rulebook to compile, so those routes are maintained deliberately, country by country, and kept under recurring review because the terms change and an out-of-date assumption there is worse than no assumption at all.

Encoding that tiering is the difference between a system that answers easy questions and one an operations team can actually rely on. Most compliance tools quietly assume every destination has a rulebook. Most of the world does not.

Underneath it, the horse is understood before the journey is planned. Identity, vaccination history and lab results are resolved into a regulatory picture first, so the question is never 'can we get this animal on a plane' but 'is this animal permitted to enter, and what is still missing'.

Reading a passport is a job for a machine. Judging whether a horse may travel is not. The whole system is built on that distinction.

wAI, our operation engine, reads scanned passports and lab reports and returns structured records: identity, microchip, every printed vaccination row, every lab result, as fields rather than prose. It never writes a record directly. Each extraction is a proposal a person confirms field by field, and every record carries whether it came from the engine or a human hand.

The compliance judgement itself is deterministic code, not a model. Blood results and vaccination timelines are evaluated in plain server-side logic precisely because a pass or fail must be identical for every member of staff and impossible to influence from a browser. Anything ambiguous returns 'verify' rather than a false pass.

There is one adversarial check on top, and its design is the part worth stealing. A model is shown only the items the deterministic engine has already passed, and is allowed exactly one kind of answer: do you trust this, or not. It cannot mark anything as compliant. So a hallucinating model can add caution, never remove it. Caution only moves upward.

One more safeguard: once a horse's export preparation begins, the ruleset for that order is frozen. If a rule file is edited afterwards, an animal that was compliant yesterday cannot silently become non-compliant halfway through its preparation.