Figures
The shape of the dataset — what it holds, how well evidenced it is, and how much of the knowable population that represents.
There is no headline number on this page, and its absence is the point.
A dashboard usually opens with one big figure. This dataset’s whole argument is that the figure everybody wants — how many castles are there — does not exist, so putting one at the top would contradict every other page on the site. What follows is the shape of the data instead: what it holds, how well evidenced it is, and how small a fraction of the knowable population that represents.
Every figure prints its numbers. The bars are there to let you compare along a straight edge; the text is what carries the meaning.
Six answers, not one
Records held per definition band. One band dwarfs the other five — not because castles cluster there, but because it is the only band a register has been imported into.
- records held
Bars share one scale, so the five short ones are honestly short. The three enumerable bands can in principle be counted to completion against a named register; the three open bands cannot, and never will — popular usage has no register and no edges. Adding records to an open band increases its size and never its completeness.
What the holdings actually rest on
The same records, split by whether anyone applied the band criterion to the structure — or whether a register's own typology was transcribed in bulk and nobody checked it.
- assessed — a person applied the criterion
- register-derived — transcribed, unchecked
Read this figure against the one above it, not on its own. Five bands are almost entirely assessed — but they hold two to ten records each, so the proportion is flattering and the magnitude is not. The sixth holds 4,659 records of which 4,650 are unchecked transcriptions. Proportion and magnitude tell opposite stories here, which is exactly why both figures exist.
How much of the world this represents
Each track is the population a register says exists. The filled portion is what this dataset holds. Only cells with a bounding register appear — an open band has no denominator, so it cannot appear here at all.
- held
- the register's population
Two cells are full because they were closed deliberately: every monument Cadw classes SiteType: Castle, and every record the Irish survey holds under MONUMENT_CLASS LIKE 'Castle%'. The Irish enclosure cell is the extreme case — one record against 31,431 — and it is deliberately not being filled. A bar too small to see is the correct rendering of that number.
Coverage, honestly stated
Twenty cells, one per definition band and jurisdiction. An unexamined cell is a published admission; three of them can never be closed by anyone.
Of the eight unexamined cells, three name no register that could settle them: Spain is uncountable by law (castles are protected automatically, so the class is open-ended by construction), Saxony's register carries no fortification category and is not open data, and England's national list has no type field at all. Those are research questions, not work orders — and they are stated rather than left as gaps.
Where two registers disagree
680 monuments classified differently by two registers joined on a shared identifier. The kind of disagreement matters more than the count.
- granularity — same band, different precision
- band-conflict — a real dispute
- category-error — someone is wrong
Most apparent disagreement is granularity — one register distinguishing a motte from a ringwork where the other says “castle”. That is two registers agreeing about the world and describing it at different resolutions, and reporting it as conflict would overstate the discord badly. Only 15 of 680 are cases where somebody is simply wrong. Read the ledger.
How well sourced the numbers people repeat are
Castle counts in circulation, graded by method rather than by plausibility. A folkloric number can be roughly right; the grade is about whether anyone can check it.
Ordered best-attested first. The reliable and the unreliable have opposite signatures: every citable figure here has a named author, a datable publication and a disclosed database, while every folkloric one is a round number with no author circulating in several magnitudes at once. “France has 45,000 châteaux” appears as 20,000, 30,000, 40,000 and 45,000 simultaneously. Read the ledger.
The numbers, as a table
Every figure above renders from these values. If a bar and a number ever
disagree, the number is right — and CI should have caught it, because
scripts/counts.py regenerates them from the records and the build fails on drift.
| measure | value |
|---|---|
| records | 4668 |
| signals | 4725 |
| definition bands | 6 — 3 enumerable, 3 open |
| registers catalogued | 13 |
| coverage cells | 20 — 2 complete, 10 partial, 8 unexamined |
| cross-register disagreements | 680 |
| count claims graded | 17 |
A note on how these are drawn
There is no chart library on this page, no JavaScript, and no colour that is not already a theme token. Every mark is a rectangle sized by a percentage, and the whole page works with scripting disabled — the numbers are in the markup, not in a tooltip.
The near-monochrome palette is a design decision twice over. The studio theme’s law is that colour is wayfinding only, so a chart here does not get a categorical palette to play with. It also happens that encoding classes by lightness rather than hue survives every form of colour blindness by construction — the two readable steps used here measure ΔE 31 apart, and the third appears only in figures where every segment is labelled in text.
The forms follow Stephen Few: bullet graphs for a measure against a reference, one shared scale per figure so bars compare along a straight edge, no gridlines, no gradients, no chart borders, and no number printed on a mark where the value column already carries it.