How this site was made

Every fact is graded and cited

There are 1,457 graded facts across the 46 destination pages, drawn from 325 declared sources. Not "Bahamas requires a microchip", but "Bahamas requires a microchip — inferred, from this Department of Agriculture page, read on this date."

That is the whole design. A reference on this subject is only as good as its weakest unlabelled claim, because you cannot tell which claims are weak. So nothing is stated without saying where it came from.

The marks are graded like a nautical chart's zones of confidence — each one tells you how much weight the fact can bear. The gauge beside Last checked answers a different question: how long ago anyone re-read the sources. A recent check is not a review — nothing here has been certified by a human.

◆Confirmedconfidence: Confirmed. Stated explicitly by an official government source.none

An official source states it, and a person has checked it against that source.

▣AI-vettedconfidence: AI-vetted. An AI read the cited document and quoted the text it found; where the wording differed from ours, an AI judged the two equivalent. Not yet checked by a human — read the quote.79 facts

An AI read the cited document and quoted the text it found; where its wording differed from ours, an AI judged the two to mean the same thing. No person has checked it yet.

◈Inferredconfidence: Inferred. Backed by an official source but not yet certified by a human reviewer.1,007 facts

Read off an official source, carefully, but not yet checked by a second pair of eyes.

▲Conflictingconfidence: Conflicting. Sources disagree — verify before you rely on it.30 facts

Sources disagree. Both are cited so you can see the disagreement yourself.

◇Secondaryconfidence: Secondary. From a non-government source; treat as a lead, not gospel.131 facts

The only thing supporting it is a non-government source.

○Unknownconfidence: Unknown. Not located in any source yet.210 facts

Not documented. The value records what could not be established.

▲No person has checked any of this, and that is the most useful thing on this page.Nothing is confirmed, because that would mean a person had sat down with the primary document and checked the fact against it, and nobody has. A handful are AI-vetted: the machine pass below re-read the cited document and quoted the sentence it found, which is better than one reading and still not a person. Treat the whole site as an unreviewed draft that cites its sources.

Where the information came from

• Government documents first, and named. Each page lists its sources with the issuing authority, the page title and the date it was read. Agriculture ministries, veterinary authorities, customs notices and statutes count as official. A commercial pet-travel site does not — it can support a fact, but the fact is then marked secondary and you can see that it is.
• Nothing was written from a search result. A search snippet is a summary of a page, and summarising is exactly where a requirement loses the scope that makes it true — whether it covers arrival by sea, whether it covers cats, whether it applies to a visitor or a resident. Facts here come from the document text.
• Sources are archived. Government sites reorganise constantly. The British Virgin Islands moved domain mid-project and took its veterinary PDFs offline; where a document is gone, the citation points at an Internet Archive snapshot and says so rather than sending you to a dead link.
• A copy of the evidence is kept. 271 cited documents are stored as text alongside the data, so a claimed quote can be checked against the document that is supposed to contain it — without depending on the original site still being up.
• A rule is written down once. US states do not restate federal rules; they inherit them and add only what they impose themselves. The same for the EU regime and the Dutch Caribbean. Restating a rule is how you get it subtly wrong.

"Not published" is a finding, not a gap

Where a page says a fee is not published, or that no minimum age is specified, that is a result: somebody looked for it and recorded that the authority does not state it. That is different from a subject nobody has researched yet, and the two are marked differently. It matters because "no published fee" is something you can plan around, while "we have not checked" is not.

How it is checked

Bad data is meant to fail loudly rather than ship quietly. Before anything reaches this site, 29 automated rules run over the dataset and will stop a release: a fact citing a source that does not exist, a strong grade without an official source behind it, a "sources disagree" mark with only one source, a stored quote that is no longer in the document it came from.

Alongside those, an automated test suite that has to pass before a release. Parts of it exist because something went wrong once: the check that every piece of data has somewhere to appear on a page is there because eight fields were silently invisible, including Sint Maarten's ban on animals from South and Central America and the British Virgin Islands' power to refuse entry outright.

A second pass that argues with the first

A fact read off a document can be read the way you expected it to read. So a second, adversarial pass tries to disprove the data: it gathers the cited documents for a section, asks an independent reader what those documents actually say — showing it the field names but never the recorded answers — and then compares the two mechanically.

It never edits anything. Any agreement it finds is a proposal for a person to apply. That restraint is deliberate: an early version of this pass reported Costa Rica's fee exoneration figure as unsupported, and the figure is quoted verbatim in the cited PDF. A pass trusted to act on its own findings would have deleted a true, sourced fact.

It was run seven times against known defects deliberately put back into the data, and it failed its own standard the first six times — false alarms fell from 81 to 1, and 1 is not 0. Each round turned up a different flaw in the method. On the seventh it reported seven real defects and no false ones, and it has since graded the first few facts on this site. Every round is written up, failures included.

It has still been worth building. While checking one of those fixes, an independent reader mentioned in passing that a US document route looked out of date. Chasing that found this site telling readers the Dominican Republic was not high-risk for dog rabies. It is on the US list. Anyone cruising there and returning to the States needs a rabies blood test, possibly an import permit, and a form that has to be endorsed before the dog leaves the country — which you cannot go back and get. That page had been saying the blood test does not normally apply.

What it still does not do

• No fact has been human-verified against its primary document. See the empty grades above.
• Two destinations still have an ungraded arrival section. Their arrival procedures are written up and cited but not broken into graded facts, so the machine checks skip them rather than guess. Every other written section on every other destination is graded.
• Departure advice is dog-only. The pre-departure steps are scoped to dogs because the US federal sources are. If you are sailing with a cat, treat that page as silent and work from the destination pages.
• It cannot tell you what will happen at the dock. It records what documents say. An officer on a particular morning is not a document.

Known gaps are not hidden. Each destination page carries its own list of what is still unsettled, in plain words, because a gap you can see beats a confident page that leaves it out.