Hotel records a traveller's own words can search, with evidence for each value.

The goal was hotel records detailed enough that a traveller's own phrasing can match them. A request like "walkable to the old town, quiet at night" depends on facts that sit in different places: some in the supplier's payload, some in the room-by-room inventory, some only visible in photographs, and some only in map data around the hotel. A flat amenity list drops most of that, and a filter with no field for the question answers no when the honest answer is that nobody knows.

The pipeline draws on six sources. The supplier payload gives 19 attributes and the room inventory 8 more. Photographs are tagged against a fixed vocabulary of 21 tags, OpenStreetMap supplies 10 signals such as walkable dining and nightlife density, a licensed location index places each hotel in its real area, and climate records give the weather in the month someone is travelling. Review text is planned but deliberately not built yet.

Every value carries the evidence it came from, the source that produced it and a confidence. Where two independent sources answer the same question, both are kept: one beach measured 807 m and 754 m from two different lineages, and the record shows both rather than picking a winner. A reviewer can ask the pipeline why any single value is what it is, and get the derivation back.

The one design rule is that unknown is never turned into a silent false. On the sample hotels, a facility code proves a pool bar exists but cannot say whether it is a swim-up bar, and none of the 142 photographs shows one, so the attribute stays unknown. That is the correct answer, and the traveller can tell it apart from a confirmed no.

The pipeline is held to that rule by checks that run offline: five structural gates, golden records and CI. Runs replay from a cache by default, so a three-hotel run makes no calls, costs nothing and finishes in about a second and a half. Live calls are an explicit opt-in.

This is a proof of concept. The original demo covered three hotels and 36 attributes and was frozen in September 2026; the newer reader, with every record, the search and the supplier document behind each value, is password-protected. The estate feed of 41,040 properties is processed a slice at a time. The enrichment module, its field registry and its evidence model are built to be extended attribute by attribute, and no traveller-facing result has been measured yet.