Coverage

The failure this project was built to answer is a map that showed nine castles in Saxony and let the reader assume that was Saxony. So coverage is claimed per (definition × jurisdiction) cell, and every cell states which of four things it is:

An unexamined cell naming its register is a work order. One naming no register is a research question.

The grid

Fortified residence

IE — CLOSED. All 4,552 records the Archaeological Survey of Ireland holds under MONUMENT_CLASS LIKE ‘Castle%’ have a record here, imported 2026-07-21 and reproducible from the committed snapshot.

READ IT NARROWLY. Complete against the SMR’s Castle class — not against the band’s criterion (4,551 of the 4,552 are assessment: register-derived at confidence: low), and not against any single idea of an Irish castle. The same register answers the same question with 129 if you accept only ‘Castle - Anglo-Norman masonry castle’. Each record keeps the register’s finer class on its register_entry, so the strict reading stays recoverable from the data without re-querying the service.

One record is assessed rather than imported: Trim Castle was hand-authored at founding, and the import found its SMR reference had been recorded wrongly. Correcting it removed a duplicate that had pushed this cell one over the register total.

GB-WLS — THE FIRST CLOSED CELL. Every one of the 104 scheduled monuments Cadw classes SiteType: Castle has a record here, imported from the register on 2026-07-21 and reproducible from the committed snapshot. Cadw is the only British register with a real castle-type filter, which is why Wales closed first.

READ THE CLAIM NARROWLY, because it is narrower than “Wales has 104 castles”. It means: complete against Cadw’s scheduled-monument castle typology. It does NOT mean complete against the band’s own criterion — all 104 verdicts are assessment: register-derived at confidence: low, transcribed from Cadw’s typology rather than assessed against this band’s rule, and counts report them separately for exactly that reason. Nor does it mean every Welsh castle: scheduling is a protection decision, so unscheduled castles are absent by construction, and Coflein — a record of what exists rather than what is protected — independently returns 158.

What closing this cell actually proves is that the pipeline can close a cell, and against exactly what. Upgrading the 104 to assessed verdicts is the next real work.

GB-ENG — Blocked, not merely unstarted, and the blocker is worth naming precisely: the National Heritage List’s bulk schema has NO TYPE FIELD, so castles cannot be filtered out of 379,680 listed buildings and 19,998 scheduled monuments. Keyword matching returns 1,190 and 3,237 respectively and is worthless in both directions at once. The Gatehouse Gazetteer is the scholarly alternative but is not openly licensed and not bulk-extractable. register_count stays null rather than borrowing a tourism figure.

GB-SCT — Blocked. Castle is not a category in the Scottish designation layers — the scheduled-monument CLASS field has exactly one distinct value across every record. Keyword matching gives 624 scheduled monuments and 2,214 listed buildings. trove.scot is WAF-blocked to automated access, so even the record side is unreachable without a manual step.

GB-NIR — register_count is deliberately null even though the register returns a number, because the number is 4 and it is wrong. Northern Ireland’s castles are classified as FORTIFICATION (353) and MOTTE (48); exactly one record in 18,506 carries the type TOWER HOUSE. Publishing 4 here would be the exact error this site exists to document.

FR — Closable with a caveat that must travel with the number: 6,286 is one of four defensible tokenisations of a free-text denomination field, and the others give 6,159, 6,285 and 5,703. The rule is published alongside the count.

DE-SN — Saxony — the jurisdiction whose coverage gap provoked this project — has no register that can settle this cell. The state Denkmalliste classifies by Denkmalart with no fortification category, splits archaeology to a separate agency, is licensed CC BY-NC-ND (not open data), and offers no bulk download. The circulating figures of 700, 723 and 821 all come from the literature, not a register, and count three different things. settled_by is empty because nothing currently could settle it: this is a research question, not a work order.

ES — Uncountable by law rather than by neglect. Spanish castles hold protected status automatically under the 1949 decree, so the protected class is open-ended and no enumerated list is derivable even in principle. settled_by is empty and will stay empty.

JP — Holds one exemplar, carried to test whether the band is Eurocentric in practice. Wikidata gives 2,266 castle-class items for Japan; no national heritage register has been assessed for enumerability here, so register_count stays null.

SY — One exemplar, carried for the military-order boundary case. No register assessed.

IN — One exemplar, carried to expose the band’s limits rather than to begin coverage. The European lord-versus-state axis does not transfer to Rajput polity, so systematic work here would need a band this dataset does not have.

Enclosure fortification

IE — The largest closable population in the dataset — 31,431 ringfort-class SMR records (rath 26,000, cashel 3,551, unclassified 1,880) — and DELIBERATELY NOT IMPORTED. This cell stays partial by decision, not by neglect.

The reason is measured rather than assumed. At 4,668 records CI takes 99s (data) and 71s (site); the Irish castle import that produced those records was itself free. Adding 31,431 more is 7.7x the corpus, which puts CI at roughly 15-40 minutes per pull request — the link checker walks every page against every other, so it grows faster than linearly. Every future change to this project would pay that, permanently, and the repository would carry about 72,000 files.

What that buys is one number. Total coverage of closable cells would go from 11% to roughly 85% — but composed entirely of register-derived transcriptions at confidence: low, of a register that is already open, already queryable, and already publishes the figure. This dataset would not learn anything it cannot state right here: Ireland holds 31,431 ringfort-class monuments, and that sentence is already true and already sourced on this page.

This is what diminishing returns actually looks like when you measure it. The remaining work with real value is the opposite shape: 4,653 of 4,668 records currently carry register-derived verdicts nobody has assessed, and closing that gap is worth more than any number of new rows. If the integer is wanted for its own sake, the import is a solved problem — the adapter pattern, the pagination and the adoption rules all exist — and it is one deliberate decision away.

GB-WLS — Cadw gives 287 scheduled hillforts; the Atlas of Hillforts covers Wales within its 4,147 entries.

GB-ENG — The Atlas of Hillforts is the bounding register and is enumerable, but its 4,147 entries span Britain and Ireland together and have not been split by jurisdiction here. Note the Atlas’s own concession that many of its entries “are not on hills and are not really forts” — the band’s criterion and the register’s criterion are close but not identical.

NZ — Recorded as an explicit gap rather than left silent. Māori pā are hilltop fortifications built in the thousands with timber palisades rather than stone. No record is held and no register has been assessed. A dataset that simply had no New Zealand cell would read as though there were nothing to find.

State fortification

FR — One exemplar from the UNESCO Vauban serial property. No register bounds French state fortification as a class; the Vauban inscription covers 12 component groups, which is a selection rather than a population.

GH — One exemplar. The UNESCO inscription covering the Ghanaian coastal forts and castles is a serial property rather than a register, and no national enumerable register has been assessed.

Palatial seat

WORLD — Open band — register_count is null by rule, not by omission, because no register enumerates “buildings called castle that are not fortified”. The best available evidence of scale is distributional rather than total: of 28,544 OpenStreetMap objects carrying castle_type, 23.7% are stately, 20.5% manor and 16.4% palace, against 19.5% defensive.

Revival and folly

WORLD — Open band with no register at all: “folly” is not a designation class anywhere. Individual examples are protected as listed buildings, so they are in registers but not extractable as a class. This band will never close and the site says so wherever it prints the number.

Popularly a castle

WORLD — Open by construction, and the band’s whole point. Popular usage has no boundary, so any global integer for it is a selection presented as a census. thecastlemap publishes 2,435 here; this dataset records that figure as a claim to be assessed rather than a population to be matched.

Why some cells will never close

Three of the six bands are open: no register bounds them and none could. A coverage cell for an open band carries no population figure at all — and that is enforced in CI, not merely intended. A cell that tried to publish one would fail the build.

The — in a population column is doing real work. It means no honest number exists, which is different from we have not looked yet.

The blocked cells are the interesting ones

Four cells are blocked rather than merely unstarted, and each is blocked for a different structural reason:

• England — the National Heritage List’s bulk schema has no type field. 379,680 listed buildings, and castles cannot be filtered out of them.
• Scotland — “castle” is not a category; the scheduled-monument class field has exactly one distinct value across every record. The record side is behind a WAF.
• Saxony — the state register classifies by Denkmalart with no fortification category, splits archaeology to a different agency, and is licensed CC BY-NC-ND. Not open data.
• Spain — uncountable by law: castles are protected automatically under a 1949 decree, so the protected class is open-ended by construction.

Northern Ireland is a fifth case and the sharpest. Its register does have a castle type, and querying it returns 4. Northern Ireland obviously has more than four castles — they are filed under FORTIFICATION (353) and MOTTE (48). Of 61 records mentioning a tower house, exactly one carries the type TOWER HOUSE. Anyone building a castle dataset by filtering national registers on a castle field would publish 4 and never notice.

That cell’s population is recorded as —, not 4.