Common methodology

Principles

The project turns three national registers of built heritage into research infrastructure, using a single site engine and the same rules in all three countries. The decisions below apply to all of them.

• Nothing invented. Every value comes from the source, from an explicit rule or from a verified link. When no rule decides, the field stays empty: an honest gap is better than an error.
• Provenance everywhere. Translations, alignments and derived values carry a [C] (confirmed in a source) or [I] (inferred by rule) mark, and the rule files are published with the data.
• Generated numbers, never typed by hand. Page totals are read from the data at build time, including the ones on this page.
• Quality is measured. Wherever reading is automated (LLM agents, heuristics), there is an audit by sample or over the whole population, and its results are published.
• Each register keeps its own semantics. The three registers are not forced into a single model: their differences are documented, and comparisons only use what is comparable.

Portugal: SIPA

SIPA (Sistema de Informação para o Património Arquitectónico, DGPC) has no API and no export. Its 26,380 records were collected page by page, with limited concurrency so as not to overload the server, and stored in an SQLite database with about 47 fields per record.

• Construction chronology: a free-text parser reads years, ranges, centuries in Arabic and Roman numerals, BCE dates and decades ("the 1940s"). A century expands to its full range (16th century = 1500 to 1599).
• Persons: the authorship field uses upper-case role labels as structural markers. The parser (5 versions) checks roles against a vocabulary of more than 170 terms.
• Materials and elements: controlled vocabularies, compound terms first (longest match wins), validated on a stratified sample of 100 records (F1 above 0.96 for element and material pairs).
• Protection dates: SIPA has no date field. Dates were read by agents from the texts of the legal acts (4,803 dated listings), using the publication date in the official gazette (DG/DR). A decree number is never treated as a date.
• Typology: the SIPA classification has five levels. For the common ontology, each record uses the most specific level that has an AAT alignment (t3, then t2, t1, c2, c1), one value per record.

SIPA includes records outside Portugal. Country comparisons only use present-day Portugal (24,939 records). The detailed methodology is at monumentospt.xyzbim.eu/sipa-methodology.html.

France: Mérimée

The Mérimée database of the Ministère de la Culture is published as an open CSV (Licence Ouverte 2.0): 46,760 protected buildings and 78 columns. Starting from an official file means everything can be recounted from the raw source.

• Normalisation: centuries read from expressions such as "4e quart 19e siècle", prehistoric eras kept apart from numeric centuries, and coordinates validated for mainland France and the overseas departments.
• Vocabularies: statuses decomposed deterministically, conditions mapped one to one and uses derived from the denominations, with translations marked [C]/[I].
• Persons: authorship strings were read by agents in batches and merged into a register of entities (people and firms).
• Wikidata: linked through property P380 (Mérimée identifier), which covers 98% of records. Images from Wikimedia Commons (P18).
• Protection: 51,628 protection events read from the protection dates and types field.

On 2026-09-02 a panel of 8 agents audited the fidelity between site and source over the whole population: identity, dating, protection, use, coordinates, text, persons and aggregates. The reports are at monumentsfr.xyzbim.eu/methodologie/.

Spain: BIC and the Andalusian inventory

There is no open national register of Spain's protected heritage. The Spanish dataset combines 20 open regional sources, each with its own licence, for a total of 55,154 records.

• Two populations, flagged in the data: 25,132 Bienes de Interés Cultural (BIC) from 18 autonomous communities, and 30,022 items from the Andalusian inventory (IAPH), which mixes BIC with merely inventoried items and has no field to tell them apart. These are flagged as "inventoried", and the protection filter separates them.
• One parser per source, each with a count test against the real figures and a geographic check per community. Source defects (swapped UTM columns, systematic duplication, invalid geometries) were fixed and documented. In the Basque Country the unique key is the Ondarea record number, and old links are redirected through an unambiguous alias table.
• Coordinates: 98% of BIC are georeferenced. From the Andalusian inventory, 4,598 assets have an IAPH locator point and 971 from Wikidata; 17,579 remain with a municipal circle only, no coordinate.
• Derived typologies: no open Spanish source provides a building type. The 39 types (with AAT alignment) are assigned to 95% of records, in this order: exact source values, rules on the name (the head noun decides, landform place names do not count, false friends are resolved), Wikidata P31 and, last, fallbacks by archaeological category. Precision was measured by agents in two audit rounds (825 and 218 records, about 90% and 92%), and the failing rules were fixed.
• Construction chronology: only a small share of records (about 4%) has a construction date. By priority, the date comes from source fields, from Wikidata P571 (year, decade or century precision only) or from unambiguous clues in the name ("románico"). BCE dates are kept as archaeological eras. Construction dates are never mixed with protection dates.
• Wikidata: linked through P808, through geographic proximity (within 150 m, with a similar name and a compatible municipality) or through the name when there is a single candidate. Ambiguous cases are left out. A sample of 100 audited links showed no errors.

Details per source and licence at monumentoses.xyzbim.eu/metodologia/.

Common layer: the AAT ontology

The three registers use different vocabularies (the five-level SIPA classification, Mérimée terms, derived typologies in Spain). Rather than invent its own vocabulary, the project uses the Getty Art & Architecture Thesaurus (AAT) as a pivot: each national term points to an AAT concept, and two terms from different countries are "the same" when they point to the same concept.

• Current size: 757 AAT concepts. Aligned terms: 721 in Portugal, 740 in France and 36 typologies in Spain. 199 concepts are shared by France and Portugal, and 27 by all three countries.
• Linked records: 26,379 in Portugal, 43,451 in France and 51,952 in Spain. Each record counts once per concept.
• Labels in five languages, by priority: the project's curated vocabulary, the plain national source value (no compound terms) and, for missing languages, Getty's own labels (read via SPARQL and cached). Each label records its origin. All 757 concepts have an English label, but only 354 have a German one, because Getty has fewer translations.
• Publication: SKOS in Turtle, with project URIs linked to Getty URIs through skos:exactMatch, plus JSON tables per country at /comum/ontologia/. The build is deterministic: the same data gives the same files.
• Use: all three explorers accept the ?aat= filter with a concept identifier (for example, castles: 300006891), so the same search can be run in the three countries.

Comparing countries

The comparison page applies the same charts and the same rules to each register. Because the registers are not equivalent:

• Shares are compared, never absolute values. Per-capita normalisation uses official population figures (INSEE, INE).
• The instruments differ: Mérimée lists protected buildings, SIPA combines listed and inventoried items, and BIC is a regional protection category. That is why only the 25,132 Spanish BIC are included: the Andalusian inventory is left out because it does not separate BIC from merely inventoried items.
• Construction and protection are never mixed. The century and sacred versus secular charts use construction dates (France and Portugal). Spain only appears in the overview table and in protections per decade, because its dates are mostly protection dates.

Known limits

• Spanish coverage is partial: some communities have unclear licences, the gaps are published in the Spanish methodology, and 17,579 Andalusian assets still have only a municipal circle.
• Spanish typologies and construction chronology are derived, not taken from the sources. Their precision was measured by sampling, not checked record by record.
• Spanish protection dates come from heterogeneous texts, and some sources have no dates at all.
• Not every national term has an AAT alignment, and Getty does not have labels in every language.
• Descriptive texts for the Spanish records and the fidelity audit of the Spanish dataset are still to come.

Country methodologies and data

• Portugal: SIPA methodology
• France: Mérimée methodology
• Spain: BIC and IAPH methodology
• Common ontology: SKOS (Turtle) and JSON
• Register comparison: comparison data