Why this page exists
Tripozi publishes accessibility and dietary guidance. Someone reads a page here to decide whether a city is navigable in a wheelchair, or whether a venue is worth trying with coeliac disease. Those are questions where being wrong has a cost, and where the reader is entitled to know who is accountable.
My name has been on the privacy policy, terms and contact page since the site launched — the legal entity is Usługi Informatyczne Szymon Łysek. This page and the bylines say the same thing on the pages themselves.
How the site is built
I am not a travel expert, and this site does not need me to be one. It is not built on opinions about destinations — it is built on other people's data and on official sources:
- OpenStreetMap venue lists. Every venue on the accessibility and dietary pages comes from OSM contributors' tags and is published with the tag it came from. A failed lookup is shown as unavailable, never as “this city has none”.
- The visa matrix. The visa pages are built from fixed records, not written by a language model when you load them. Each entry links to the destination country's own ministry, immigration authority or official portal, with the date it was last checked.
- The open dataset. The wheelchair-accessibility extract on /data is published under ODbL with the OSM attribution it requires, so anyone can check the pages against the data.
Responsibility for all of it — including the mistakes — is mine.
Where AI sits, and where it does not
Itineraries, destination prose, and the overview and neighbourhood ratings on the accessibility pages are generated by a language model, and labelled where they appear. The dietary pages carry no model-written text, and the venue lists and the visa requirements are not generated either — those come from OSM and from government sources. The methodology page sets out which layer is which.
What went wrong, and what was done about it
In August 2026 our own Copenhagen accessibility page was read the way an outside expert would read it, and it turned out to recommend an accessibility symbol that does not exist — the word it named is simply the Danish adjective for “accessible”. So the named things across both specialist guides were audited properly.
Of 566 named entities, 19 were invented and 37 misnamed. A hospital with a street that does not exist. Taxi operators no register knows. A transit assistance service under the wrong name. A Hebrew phrase for asking about kosher supervision that actually said something else. An Estonian phrase about cross-contamination that asked about cumin instead.
The cause was in our prompts, not in the model. They said never to invent numbers and never to fabricate an address, and nothing at all about names — and a model reads that contrast as a statement about which categories tolerate invention. Worse, one prompt offered an invented app as an example of good practice, so the pattern it was shown was fabrication.
Those sections were taken down rather than patched, because fixing only the errors that had been found would have left the ones that had not. They come back per city, after that city is regenerated and re-audited — not before. The full record, including what the audit did not cover, is published with the site.
It is on this page deliberately. It is the least flattering thing about this project and the most useful thing you can know about it: when the content here was wrong, it was removed rather than left up.
Corrections
If something here is wrong — a venue that has closed, an address that does not exist, a visa rule that has changed — please say so. Send the URL and what is wrong to [email protected]. Every message is read, and a correction that can be verified gets fixed. Corrections that must survive regeneration are recorded in the codebase rather than patched into a cache, because re-running the generator on a bad claim tends to produce the same claim in a more convincing form.