Tap the map to guess — and keep playing past midnight

Changelog

A small atlas, kept honestly: every user-visible change lands here, newest first. It is run by one person building in the open — AI does much of the heavy lifting, a human decides what ships. Corrections are welcome.

Follow via feed: /changelog/feed.xml (Atom)

Spotted an error? ceoflightmussy@gmail.com

16 August 2026 — The translation ran one way for a month — now every landmark tells its story in English

• 1,292 landmarks whose only Wikipedia article is Hungarian, Russian, Chinese or one of nineteen other languages now tell their story in English. Bobruisk Castle used to open with a paragraph of Russian on its English page, and a reader who could not read it got nothing else. The reason was structural: translation only ever ran outward, from the English lead into the six other editions, so a landmark with no English article had nothing to be translated from and stayed in its own language for good. It now runs both ways. Every one of these stories names the Wikipedia it came from and says plainly that a machine translated it — the condition the licence attaches to a translation, not a courtesy.
• And 156 landmarks that had no story at all now have one — among them Hadrian’s Wall, which ranks ninth in the entire atlas, the Ishtar Gate, the Antonine Wall, the Walls of Constantinople, Mystras, Monteriggioni and the Porta Nigra. They arrived with August’s expansion into walls and gates, and the script that fetches English leads had not been run in full since — every run after it only deepened the stories already there, and never noticed the new pages had none. English coverage across the atlas goes from 79% to 99.6%.
• The six other editions gained the same safety net. 847 pages on /de/, /es/, /fr/, /ja/, /nl/ and /pl/ showed no story at all where the landmark’s own Wikipedia had a good one, because the atlas looked in the reader’s language, then in English, and then stopped. They now fall back to the landmark’s home wiki, in its own language, marked and credited — a paragraph written by a human editor beats a blank space, whatever language it is in.
• 1,857 map cards stopped describing themselves and started describing the castle. Hovering Osaka Castle used to produce the words “Japanese castle”; the Aljafería “Cultural property”; Kenilworth “castle ruin” — a Wikidata label repeating the chip printed beside it. Worse, 380 cards said nothing at all, Gjirokastër among them, on pages whose own text ran for paragraphs. A card now takes the first sentence of the landmark’s own article — including the ones only translated into English today — so 5,642 carry a real description and 61 are left with nothing to say.
• Smaller things. The map’s four controls fit on one row again — the arrows are gone, and the type is now sized per language, because “Tout parcourir / Surprenez-moi” needs half again the room of “Browse / Surprise me” and one size for everybody meant everybody got French’s — instead of leaving About stranded on a second row, and the panel keeps its width so the map keeps its room. On a phone those four now sit above the panel, always visible: reaching the castle list or the game used to mean opening something labelled “Filters”. CastleGuessr drops the instruction bar under the photo on phones — the map is for tapping, and nobody needed telling. The sister-atlas links now read World Train Map, which is what TrainRouter has been called since it moved. And one castle left the map: قصر مارد turned out to be a second Wikidata item for Marid Castle, 117 metres away and already here, so the atlas holds 7,044. The panel also lists Detourmap at last — every other footer on the site has since July. And twenty stories that opened with extraction debris rather than a sentence — a stray }}, a bare line of coordinates, or dewiki’s i3i6 — now open with their first words; sixteen of those were German pages, Montezuma Castle and Fort Mackinac among them.

11 August 2026 — Every short story becomes the whole story — in all seven languages at once

• 2,338 landmarks whose page told a single line now tell their whole article. Wherever the English Wikipedia lead ran under 400 characters, the atlas now reads the entire article — the history sections, not just the definition — and the page's story budget doubles to fit it. Karlštejn, which introduced itself in one sentence, now carries five paragraphs of Charles IV, the Imperial Regalia and the Hussite siege.
• The new text was read for dates the way this atlas always reads them: 327 more landmarks now state a sourced founding, bringing dated coverage to 75% — and the rejects show the guards working. An architect's lifespan in parentheses is not a founding; "for example Glenveagh Castle was constructed in 1870" dates Glenveagh, not the page it appears on; a belltower added in 1905 does not date a 14th-century castle; and "built before 1400" or "tradition holds" now honestly wear the "c." they deserve. Písac is a case study: its article offers Pachacuti's reign years in parentheses, and the honest reading — "no earlier than 1440" — is what the atlas now states.
• The six other languages caught up with English in one night: more than 22,600 stories were translated anew, replacing the single-sentence translations that stood where a reader's own Wikipedia has no article. And where a reader's Wikipedia has an article that says less than the English one, the page now keeps the native lead first — real prose by a human editor — and continues below it with the English story, clearly marked as a machine translation with its own credit. Each language edition now tells, as nearly as the sources allow, everything the English page tells.
• Two changes for the search engines: every castle page's search snippet now ends with what only this atlas knows — the landmark's rank in its country, its century, and that it sits on an interactive map — instead of duplicating Wikipedia's first sentence; and every hero photograph now declares its creator, its licence and its Wikimedia Commons source page in machine-readable form.

10 August 2026 — The atlas reads eleven more languages for its dates, Zvečan comes home, and every page now promises only what it holds

• The year-reading pass learned Italian, Portuguese, Russian, Hungarian, Czech, Slovak, Turkish, Greek, Japanese, Chinese and Korean — and 333 more landmarks now state their founding, read from their own articles, never guessed (dated coverage: 66% → 71%). The chain finds dates where you would not look: Sacsayhuamán's Killke origin around 1100 arrived via the Czech article, Stirling's 14th-century core via the Korean one, the Vilnius Castle Complex's 10th-century beginnings via the Chinese, and Van Fortress now reaches back to the 9th century BC with the Urartians. Persian was left out on purpose: a bare year on the Persian Wikipedia is routinely Solar Hijri — 1250 there is 1871 here — and a silent 600-year error is worse than honest silence.
• Every candidate date was read against its source sentence before it was believed, and the rejects tell you why: Gloucester Castle's lead offers 1787 — the year it was demolished and a prison built on the site; the Vilnius articles offer 1390 — the year the Teutonic Knights burned the Crooked Castle; one Slovak lead dates a mining academy's relocation; an Italian one, the year researchers discovered a moat. None of those are foundings, and none of them are in the data.
• Zvečan Fortress now stands in Kosovo — where it physically stands, in Zvečan municipality north of Mitrovica. A reader caught it filed under Serbia, which is what Wikidata's country statement says; the atlas files every landmark by where it stands on the ground, and now this one too.
• And the atlas has a new landmark, by request: Dún Dealgan Motte — Cú Chulainn's Castle in Dundalk, the Norman motte on the hill the Táin Bó Cúailnge makes the birthplace of Cú Chulainn, today the centre of a local preservation campaign. Jaki McCarrick wrote in to report it missing; its four Wikipedia editions sit under the atlas's usual notability floor, and that is exactly what the reader's door is for. Ireland now counts 171.
• The country pages stop promising what a third of the atlas could not deliver. "Each with its photo, founding century and story" was simply false for undated landmarks — so every country page in all seven languages now claims the founding date only where every landmark has one, and says honestly how many do: "315 landmarks, every one mapped with a photo and its story — 232 with sourced founding dates." Search-result titles carry the same per-entry guarantee where it fits, and a new image sitemap pairs each castle page with its Commons photograph, so the photos behind every pin are no longer invisible to search engines.

9 August 2026 — The castles that were really palaces, the palaces that were really ruins, and 431 years the atlas could finally state

• A "castle" built after 1800 is almost never a castle — the age of castles was over — and this release acts on it. Every castle-category landmark dated after 1800 was re-read against its own encyclopaedia lead, and forty moved to the palace category: Schwerin, whose German name has said Schloss all along; Miramare, a Habsburg seaside residence; Peckforton, "a Victorian country house built in the style of a medieval castle" — the styling no longer outranks the noun; Glenveagh's "castellated mansion"; the Schlösser of Luxembourg and Sweden. The landmarks the world genuinely calls castles keep the name: Casa Loma, Larnach, Hearst Castle and Penrhyn stay castles by identity, as their communities asked, and Saint Michael's Castle in St Petersburg stays a castle because Paul I really did build one — moat, drawbridges and all — in 1800.
• Ten "castles" turned out to be forts wearing the word: Castle Williams and Castle Clinton are circular batteries in New York Harbor, Goryōkaku is a star fort, the Eternal Golden Castle is a coastal battery in Tainan, and three castellated American armories are armories. All are filed under fortress now. In the other direction, the Reduta in Bratislava — a neoclassical concert hall that carried the castle label — is a palace, and Kosava Castle is recorded as what its own name says it is: the Puslowski Palace.
• The ruin detection learned Spanish and French. It had read English, German, Polish and Dutch leads — so a landmark whose only articles are Spanish or French could say "en ruinas" or "en ruine" in its first sentence and never be heard. Thirty-two landmarks are now filed as the ruins their own articles describe, Dryslwyn and Wiston in Wales, Tiffauges, Haut-Barr and Pflixbourg in France, a dozen Spanish castles among them. Four that the French Wikipedia calls ruins were kept whole on the evidence of the ground: Medvedgrad above Zagreb (restored, a visitor centre), Oshi Castle (its turret rebuilt in 1988), Puivert (the keep stands roofed) and Moret (the donjon stands) — a stale foreign lead is not a demolition order.
• And the years. Twenty-one landmarks carried a rebuild, replica or restoration date as their founding — Peel Castle said 1860 when its own lead says the Norwegians built it in the 11th century; Dankwarderode said 1963; Inverness said 1836 when castles have stood on its rock since 1057; De Haar said 1892, the year of its celebrated rebuild, not 1391, its first mention. All corrected against their own leads. Then the year-reading pass learned French, Spanish, Polish and Dutch on top of English and German, and 410 previously undated landmarks now state their founding — read from their articles, never guessed: Kızıl Kule's exact 1226, Bled Castle's 11th century, Chittorgarh's 8th. The atlas now dates 4,644 of its 7,044 landmarks; the remaining third say nothing datable in any of six languages, and honest silence beats a guess.

8 August 2026 — Twenty-two things that were never castles leave the atlas, found by the atlas auditing itself

• Nobody reported any of this. Every fix below came out of an audit the atlas ran on itself — and the method is the part worth writing down, because the old one had quietly stopped working. Previous sweeps read each landmark's NAME and its one-line description, which is how a wall, a bunker and a city of a quarter of a million people all reached the top hundred wearing the word castle. This sweep asked Wikidata what each of the 7,044 landmarks is, by type, in one pass over the whole atlas. It found twenty-two things that were never castles and eleven more filed under the wrong one of the four categories. Three new automated checks ship with it, so each class of error stays fixed rather than being found again in a month: a release now refuses to build if the map cards have gone stale, if the machine-readable summary this site publishes contradicts the data, or if any page still links a sister atlas at a domain that has moved.
• The most-read castle page in the United States was NORAD. Not a castle, not a fortress, not a building: Wikidata calls it a multinational military command, which is to say an organisation. It had ranked 97th in the world, inside the band this atlas labels world-famous, and it was the first landmark listed on the page for the United States. That page now opens with Fort Sumter.
• It had company. The Israeli West Bank barrier and the Mexico–United States border wall were both filed as castles, ranked 161st and 217th. The release earlier today removed three other twentieth-century barriers on the stated grounds that a separation wall is a fortification in the dictionary sense and in no other; it simply missed these two, because that pass searched names and these are only identifiable from their type. The Führerbunker, the Churchill War Rooms and the Battle of Britain Bunker have gone the same way, alongside the twentieth-century gun batteries that left in August.
• Groningen was in the atlas. The city, ranked 22nd in the world, illustrated by a map of itself drawn in 1575. Culemborg, Weesp and Doesburg were in too. All four are typed "fortified town" by Wikidata, which is true of their history and not of what they are — places where a quarter of a million people between them live now. The line is habitation, not age: Mystras, Numantia, Ani and Carteia are ruins and they stay, and so do fortress towns like Terezín and Neuf-Brisach whose whole identity is the fortress. A disused Tube station, an urban park eighteen metres from Osaka Castle, Valletta's 2014 Renzo Piano gate and a former malthouse in Elbląg also left.
• Groningen keeps a landmark, and it is the right one. Fort Bourtange was in the atlas all along, at 1,204th, unrecognisable because Wikidata labels it "Bourtange Fortress" — an inversion no other source uses. It is the five-bastion earthwork thrown up in 1593 on the one road across the Bourtanger Moor, built to cut the Spanish supply line to the city and now restored to how it looked in the 1740s. It carries its English Wikipedia name from today, and it is pinned, so a star fort with eleven Wikipedia editions can never quietly fall below a threshold.
• Then the same question was asked a second time, deliberately: if the Brandenburg Gate could sit at ninth for weeks, what else is here on a technicality? Six more. Three are defence lines — the Hindenburg Line, a hundred and fifty kilometres of First World War trench from Arras to Laffaux, which the atlas had filed as a castle at 190th; Israel's Bar Lev Line of 1968, a sand rampart along the Suez Canal; and the Toblerone Line, a row of concrete anti-tank blocks poured in Vaud in 1938. Two are working United States Army posts that inherited the word fort and nothing else about it, Fort Gordon and Fort Polk. The last is Bishop Castle in Colorado, begun by one man in 1969 and still unfinished, which English Wikipedia describes as a roadside attraction while putting the word castle in quotation marks — when the encyclopaedia will not commit to the noun, this atlas should not either.
• Knowing what to keep is the same work as knowing what to cut, and the sweep is only worth trusting if it can tell them apart. Hadrian's Wall, the Antonine Wall and the Limes Germanicus stay: Roman frontier works with forts strung along them, and a heritage class of their own. The Arco de Santa María in Burgos is also typed a triumphal arch and also stays, because it is one of the twelve gates the city had in the Middle Ages, merely refaced for Charles V — a defensive gate in ceremonial dress, which is the exact inverse of the Brandenburg Gate. Twenty-three historic forts carry the type "military base" only because they still stand on military land — Kangra, Attock, the citadel of Ghazni — and every one of them is a fortification and stays. So do the follies that are also castles, Swallow's Nest and Gwrych among them.
• What all of these had in common is how they were found. Every previous sweep read the name and the description; this one read what Wikidata says each thing is. "Castle in Bavaria" was never the tell. The type field named all fifteen in a single query, and it also settled eleven that were in the right atlas but the wrong category: Daikaku-ji and Shōren-in are recorded as imperial residences and are now palaces rather than castles, two Transylvanian fortified churches are fortresses as Biertan already was, Mystras is a ruin, and Bir Hakeim, a fort in the Libyan desert, is no longer a palace.
• And the cards learned to speak again. A card should tell you something the chips beside it do not, and three thousand six hundred and eighty of them were saying things like "Castle in Bavaria, Germany, Europe" — that was Neuschwanstein — or "Palace in Vienna, Austria", or "Japanese castle complex". The pass that writes a readable opening line from Wikipedia had not been re-run since the atlas grew from about 2,400 landmarks to over 7,000, so it was writing for three of them. It now writes for all 3,700, and a release cannot ship again until it has: the check is part of the pre-flight, with the count that must be zero.
• A related twenty-four were saying something worse than too little. A Wikidata description is a noun phrase — "Castle in Bavaria, Germany" — and these opened with a verb, so they read as sentences begun in the middle: "Was built under orders of William the Silent and completed in 1593", "Has a long history. The southern tower used to be…", "Is the medieval ruin." They had survived every check because they are long and specific, and length was the only question being asked. Grammar is the right question, and twenty-two of them now open with a proper sentence instead. The last two have no English article to draw one from.
• A hundred and ninety-one photographs were being used under licences that require naming the photographer, and named nobody. Their Commons pages predate the machine-readable author field, so the credit line read "Photo · Wikimedia" — for Chenonceau, Stirling Castle, Qasr Amra, Livadia. Every one of them now names a person, and the published CSV and GeoJSON carry the photographer and the licence as columns, along with a link back to the landmark's page, which the CSV never had.
• Smaller repairs. The three castles of the Dahner Burgengruppe — Altdahn, Grafendahn and Tanstein — stood at one identical point, so the map drew three castles as one dot and only ever opened one of them; they are 50 to 110 metres apart and now sit where they stand. Six landmarks were founded in this century according to the atlas, which is a good way to notice a restoration date wearing a founding date's clothes: Namdaemun was not built in 2013, the year it reopened after the fire, but in 1398. The Venetian palace listed as "Hotel Danieli" is now Palazzo Dandolo, the building rather than its tenant. Lions' Gate in Jerusalem was the one landmark with a country and no country code. Six more maps and plans have become photographs.
• The file this site publishes for AI assistants was contradicting itself. It said the atlas reads 26 Wikipedia language editions when it reads 22, and in one sentence gave a dated and an undated count that added up to the atlas as it was two revisions ago. Every number in it is now checked against the data on every build.
• The atlas now holds 7,044 landmarks across 141 countries.

8 August 2026 — The map learns where the sea ends, the game borrows the site's own controls, and Crimea stops depending on luck

• The night chart had a real problem and a reviewer was right to say so. Keir Clarke of Maps Mania wrote that the map was barely legible, and it measured: the land was separated from the sea by a contrast ratio of 1.21 to 1, where three to one is the floor for a graphic that carries meaning. Finding the Adriatic or the Gulf of Bothnia meant already knowing where they were. The land is lighter now, and — the part that actually does the work — the coastline has stopped being a faint wash and become a drawn line. A map separates land from water with a coast, not by hoping two fills differ enough. The sea itself never moved, so it is still a night chart.
• The panel could not be read either. Its faintest labels sat at a contrast of 1.98 to 1 against their own background, at roughly nine pixels; the standard for text that size is 4.5. Thirty-nine of those tiers have been raised across the panel, the castle card and the game, with the ordering between them kept intact.
• CastleGuessr aims properly. The cursor was the plain system crosshair, a thin dark cross that all but vanished against dark land — in a game whose only move is clicking an exact spot. It is now the same high-contrast reticle the atlas has always used. The game also carries the site's own language and theme controls at last: every edition is one click away without leaving your round, and switching between light and dark now actually repaints the board, which it never did before.
• A reader pointed out, kindly, that the atlas could not decide what it contained: the name says castles, the title added fortresses, the description added palaces, and the legend then offered four categories. Ruins are about a fifth of everything here. Every place the atlas lists what it holds now names all four, in all seven languages.
• Crimea. A reader repaired the Wikidata records for the peninsula's castles, fortresses and palaces so that this atlas could show them, and in checking that work we found something worse than a gap: Wikidata records these landmarks with two countries at once, often at equal weight, so which one the atlas printed came down to the order two rows arrived in. Every Crimean landmark had read "Ukraine" by luck rather than by rule. Coordinates decide it now. Crimea is Ukraine, and it no longer depends on chance. Six landmarks joined at the same time, including Swallow's Nest and the Khan's Palace at Bakhchysarai — two of the most photographed places on the peninsula, absent until now.
• Northern Ireland had 15 landmarks against Ireland's 170 and Scotland's 144, and not because it is short of castles. The atlas sorts the United Kingdom into its four countries, but it was doing so after deciding what to include — so during selection a Belfast castle's country was simply "United Kingdom", and no UK country could ever be topped up. It now sorts them first. Northern Ireland has 32, among them Killyleagh, Monea, Antrim and Harry Avery's.
• Eight landmarks were showing a map, a plan, an engraving or a coat of arms instead of a photograph. All 6,903 images were audited; 31 were not photographs, and Wikimedia Commons had a genuine replacement for eight of them — Domus Augustana, Southampton Castle, Fort d'Évegnée, Fort Knox in Maine, Okopy, Inzlingen, the Puerta de Santiago in Melilla and the Citadel of Homs. The rest keep their drawings on purpose: Worcester Castle, Fort Coligny, Shalfak and the others are demolished, drowned or vanished, and a period plan is the only surviving picture there is.
• The atlas now holds 7,066 landmarks across 141 countries.

7 August 2026 — The atlas stops giving up on a tab you opened in the background, and Japanese stops meeting English

• Open one of these maps in a background tab, leave it a few seconds and then switch to it, and the atlas had quietly settled for its plainest basemap — country outlines, no coastlines, no relief, no place names — and stayed there for the rest of the visit. A browser does not draw a tab nobody is looking at, and the map library assembles every style inside a drawing step, so in the waiting tab nothing could load. The map read that silence as a tile server that was down, swapped in the offline chart, and spent both of its recovery attempts while nobody was watching. That clock now starts only once the page is genuinely on screen — and a tile server that really is unreachable still gives way to the standby map as quickly as it ever did.
• CastleGuessr could lose the world it is played on. The board opens as bare land with a grid and slides the real map underneath the moment the tile server answers — but if that answer arrived before the board was ready, the swap was abandoned for good and the whole session was played on an empty stage. That is exactly what a background tab does, and on a repeat visit an answer served from the cache can do it too. The tiles now wait for the board rather than the other way round. That same swap could also wipe your pin, the castle and the line between them off the map if it landed after you had already tapped, while the result card still told you how far out you were; the board now redraws the round it is holding.
• The game page is the game now: the description that sat one screen below the board is gone.
• And a round whose photograph refuses to load offers the next castle instead of leaving you to guess at nothing.
• Japanese readers were still meeting English on 141 landmark pages — every fortress added to the atlas this week, from Mont-Saint-Michel Abbey to the Citadel of Cairo and the Walls of Jerusalem. All 141 open in Japanese now. Of the pages that carry a description at all, 92.6 per cent read in Japanese; what is left are landmarks documented only in a third language, which are shown in that language and credited to it.

6 August 2026 — CastleGuessr gets its map back, and on a phone the photo steps out of its way

• The game had no board. Yesterday gave the atlas real relief, and the game builds its board out of the atlas’s own style with every label stripped away — so it inherited the new hillshade layer and left the elevation source behind. MapLibre discards an entire style over a single layer pointing at a source that is not there, and by then the working one has been taken down: the result was a game whose only means of answering, the world map, was simply absent, on desktop and phone alike. The board carries the elevation now, ignores any layer whose source it does not have, and keeps its chart grid instead of losing it the moment the basemap arrives. Every map style the site draws is put through MapLibre’s own validator on each build, so this kind of fault cannot ship quietly again.
• On a phone the photo had become the question and the answer’s hiding place: it filled the panel, and the map you are asked to tap lay underneath. It is now a thumbnail in the top corner — under a tenth of the screen — and a tap opens it up when you want a closer look. The instruction moved to its own line below it, the running session score appears on the bar, which phones never had, and once the round is played the photo stretches into a strip across the result card.
• Polish readers were being offered castles “w Macedonia Północna”. North Macedonia was missing from the table of locative forms, so the name stood undeclined in every heading, description and question on that page; it reads w Macedonii Północnej now. French gained the same country’s article forms.
• Two different castles were both called Janowiec Castle — one the ruin above the Vistula in Lublin province, the other a border castle 450 km west — and both, unhelpfully, stand in a village named Janowiec. The second is Janowiec Castle (Lubusz) from now on.

5 August 2026 — The fame ranking learns to read every language, Poland trades ninety entries for its dignity, and the land gets its shape back

• Fame was being judged in English. The ranking blends how widely a landmark is documented with how widely it is read, and the reading half counted English Wikipedia only — so a castle whose article exists in Polish, German or Spanish and nowhere else scored zero readers, and sank. That was 1,408 landmarks, one in five on the atlas. Readership is now summed across every language edition a landmark has: Neuschwanstein is read 850,000 times a year in English and 290,000 more in German, and Moszna Palace turns out to have eight times the readers its English article suggested. Not one landmark is unranked now. The old caveat on the ranking page — that the numbers leaned English — is gone, because it is no longer true.
• Poland has lost ninety entries and is better for it. Answering a Polish reader’s complaint that the map was empty where he lived, the country’s target was raised to Germany’s that same morning — but Poland’s pool of well-documented castles runs out around 360, so the last ninety admissions were the scrapings: places carried by two Wikipedias, several not even named, only described (“Castle in Dzięgielów”, “Manor house II, so-called castle in Stary Sielec”). They are removed, the target is set where the evidence actually ends, and a floor can no longer buy quality that is not there. Poland now stands at 361 landmarks, none below three languages.
• Seven Polish palaces had been filed as castles, because the classifier read English and German and never Polish — so a landmark whose only article opens “Pałac w…” was never asked what it was. It reads Polish now, and eighteen landmarks moved to the right category.
• And a gun battery has left the castle map. Battery Dietl is a German coastal artillery position of 1942, and it had been filed as a castle; nine more 20th-century emplacements went with it, most carried by two Wikipedias. The 19th-century coastal forts stay — Tung Chung of 1817, Murray of 1841, Konstantin of 1840 — because a fort you can visit is a landmark and a wartime gun position is an installation.
• The map itself now shows the ground. Real elevation shades every hillside, ridge and valley, drawn from open government terrain data, so it is finally visible why a castle sits where it does. It costs the opening view nothing: the relief arrives as you come down towards a country, which is also the first point at which a hillside means anything.

5 August 2026 — The Dutch edition learns Dutch, Poland stops looking empty, and the citadels come home

• The Dutch atlas now reads Dutch. 3,961 landmark stories that had been standing in English on /nl/ were translated, so Dutch was the last edition where a reader could meet a castle in a language they had not asked for. With it goes the last machine-translation notice on the site: no edition carries one now, because none of them needs to — each translated story says so in its own credit line, which is the honest place to say it.
• Poland was thin and a reader said so plainly. On the Polish link site wykop, jaroslaw-lazarczyk called the map useless where he lives: Lubusz Voivodeship showed two castles when several stand within twenty kilometres, and arturo1983 and bialy100k said the same of their own regions. They were right. Poland has gone from 279 landmarks to 451, Lubusz from fourteen to twenty-two, and the same pass repaired five Polish entries whose Wikidata labels had shipped a street address, a genitive ending and two buildings named after whatever later stood on the site.
• And a whole class of fortress was missing. Berat Castle — Albania’s UNESCO citadel, with a village still living inside its walls — was absent because Wikidata files it as a district of Berat, so no query the atlas asks could ever have found it. Checking whether others were hidden the same way turned up fifty-six, among them the Cairo Citadel, read in thirty-five languages. Twenty-nine are now on the map — Saladin’s citadel above Cairo, Mont-Saint-Michel, the Alamo, the walls of Jerusalem, Anadoluhisarı on the Bosphorus, Skopje’s Kale, Kotor’s walls, the Amman Citadel, the Kasbah of the Udayas, Rhodes, Kangra, Cổ Loa’s spiral ramparts and the Warsaw Barbican — and the query that missed them has been fixed at the root, so the rest arrive on their own. The atlas now maps 7,003 landmarks across 141 countries, Bahrain among them for the first time.
• CastleGuessr keeps only the photograph. The castle’s name is gone from the board, and so are the faint dots that used to scatter every other castle across the map — between them they answered half the question, since the name could be searched and the dots showed which neighbourhoods were even possible. One photograph, the whole Earth, one tap. The name arrives with the reveal, where it belongs.

5 August 2026 — CastleGuessr finds its final form, every photo names its photographer, and 364 silent pages find their voice

• CastleGuessr is now the one game its name promised — the shape proven on our sister atlas FilmMap. The daily puzzle and the endless typing rounds are gone, autocomplete and all. Every round asks the same clean question: here is the castle, name and photo — where does it stand? One tap on the world map answers, scored out of 5,000 by how close you land, and the reveal draws the line from your tap to the truth. The board also stopped naming cities — a map that prints Osaka would answer for you. Play as long as you like; rounds, average and best stay on your device.
• Every photograph now names its photographer. 6,469 of the atlas’s 6,653 photos carry their author’s name from Wikimedia Commons — on the map card, on each castle’s page and in the game — beside the license each photographer chose. The rest record no machine-readable author and stay credited to Commons itself.
• 364 landmarks that had no story in any of the atlas’s seven languages now tell one in their own. Their only Wikipedia article lives in Hungarian, Russian, Chinese, Romanian, Portuguese, Czech, Arabic, Greek… — so that is the story their pages now open with, marked with its language and credited to its edition, in place of a bare data sentence.
• And the splitter that had been quietly beheading sentences at abbreviations — an extract that ended “located in Strelna in St.” and called itself complete — is fixed at every place it lived, including the freshly caught hyphen case (“Feldbrunnen-St. Niklaus”). Fifty-one pages whose broken sources had refused translation now read whole in Japanese, Polish, Spanish, French and German.

4 August 2026 — A reader caught the rankings broken — and a caravan of fixes rode in behind

• On Menéame, yarkyark looked at the fame ranking and said the quiet part: castles nobody has heard of ranked high, “and the Alhambra doesn’t show up”. He was right, and it was worse than it looked: a refresh two days earlier had silently dropped the readership numbers for 5,807 of the atlas’s landmarks, so fame was being judged on Wikipedia coverage alone — Spain’s page ranked Numantia first and the Alhambra #107. The leak is fixed at its source, every landmark’s readership was refetched, and the Alhambra sits where it belongs. Earlier doubts from runxel and prmph on Hacker News were the same wound showing; thank you all.
• 241 landmarks joined — the atlas now maps 6,653 across 140 countries. @magpieuae shared the map with “tho not UAE”, and fairly: the Emirates stood at exactly one entry. Now eighteen — Al Fahidi, Fujairah, Sharjah, Ajman, Qasr Al Muwaiji — and Oman grew from seven to thirteen, from Rustaq to Khasab. Guinea-Bissau and Wallis and Futuna enter for the first time (Talietumu, a walled Tongan fort on ʻUvea). By email, Eloy pointed out that Catalonia’s record lives in Catalan — his example, the Castell de Burriac, had no English article at all — so the atlas now reads Catalan, Basque and Galician Wikipedia for Spain, whose count grew to 500; that also answers the Valladolid and Extremadura gaps Stathamdepueblo and Jjota named. Reader reports became pins: Fort Wilkins in Michigan (Elizabeth, by email), Biertan fortified church (rob74) and Montezuma Castle (1970-01-01). Two could not come yet, and honesty beats silence: Coolmain and Kilbrittain Castle in Cork (Christopher Willoughby’s report) have no usable Wikidata item — photo, coordinates or the item itself are missing — and neither, still, does Siyu Fort in Kenya.
• Every localized page now carries its name in both tongues: 大阪城 answers to Osaka Castle in the tab title and under the heading, on all six non-English editions. And the openings that arrived broken — “S. Virgin Islands.” standing alone where a sentence should be — are repaired at the splitter that broke them, so nearly a thousand pages across five languages traded English fallbacks or fragments for whole sentences in their own language.
• The guessing game is now CastleGuessr, and it learned the move its new name promises: pin mode. You know the castle — the photo and name are right there — but where does it stand? One tap on the world map is your answer, scored like the classics: 5,000 points inside 25 km, fading with every kilometre you missed by. Daily and endless play on unchanged.
• And a small mercy on the map itself: a card you hovered but never clicked no longer swallows the scroll wheel — zooming works with a card open, and a card whose text overflows still scrolls itself.

4 August 2026 — German, Spanish, French and Polish now read in their own languages too

• What happened to the Japanese edition yesterday has now happened to four more. 9,844 landmark descriptions that opened in English on the German, Spanish, French and Polish editions have been translated: 1,643 in German, 2,587 in Spanish, 1,878 in French and 3,736 in Polish. Together with yesterday’s Japanese, five of this atlas’s seven editions now tell every story they can in the reader’s own language — and the description under a search result comes in that language too, which is the part a search engine reads.
• The translations keep the habits each language actually uses, not English ones wearing a costume: Roman-numeral centuries in Spanish and French (siglo XIII, XIIIe siècle), French spacing before its colons and guillemets, Polish declining foreign names through its cases (zaprojektowanych przez Guya Maunsella). Every number, date and original-script gloss is carried across untouched, and each translated page names the English Wikipedia article it came from and says plainly that a machine translated it.
• Some 380 pages across the five editions stayed English on purpose. Their source extracts arrive damaged from upstream — cut off mid-sentence, or starting mid-word — and translating a broken sentence only produces a broken sentence in a second language. One such fault was found and fixed while checking this batch: an abbreviation like “the U.S. state of Montana” was being split at the full stop after U, leaving a fragment that read as a whole sentence. The check found its mirror too: extracts that end at an abbreviation — “located in Strelna in St.” — had lost their final words, and fifteen Japanese translations published yesterday had carried that missing name into print. They have been withdrawn, and all those pages now wait for a better source instead.
• The Dutch edition still shows English where its stories are missing, and still says so. Where an edition no longer does, that notice is gone — a warning that has stopped being true is worse than no warning at all.

3 August 2026 — The Japanese edition now reads Japanese

• Since the Japanese edition opened in July, it has carried an honest compromise: where no Japanese Wikipedia article existed, the page told its story in English — and said so. That compromise ends today. 4,122 pages now open in Japanese, machine-translated from the English Wikipedia lead each page has always credited, joining the 2,199 landmarks whose Japanese Wikipedia tells their story natively. All but 98 of the atlas’s landmark pages now read Japanese — the descriptions under search results included.
• The translation stays honest about itself: every translated page says 機械翻訳 — machine translated — in its credit line, keeps every number, date, name and original-script gloss exactly as the source gives them, and links the English original. And the 98 holdouts stay English deliberately: their source extracts arrive damaged — cut off mid-sentence upstream — and a broken sentence translated is still a broken sentence. They wait for a better source, not a braver guess.

3 August 2026 — Tap the map to guess — and keep playing past midnight

• The guessing game grew a second way to answer: the world itself. A night map of every landmark now sits under the photo — tap anywhere and the nearest castle dot becomes your guess, answered as ever with distance, direction and closeness, and every try leaves a colored ring behind showing how you closed in. Typing a name works exactly as before.
• And midnight no longer ends the game: a new endless mode deals castle after castle from the same famous five hundred once the daily is done — or instead of it — with its own played-and-won tally. The daily castle, its streaks and its share grid stay untouched, and the map’s panel now carries a one-word Guess button, so the game is never more than a tap away.
• Also mended: the Dutch, Spanish and French editions of the game page were showing the atlas instead of the game. All seven languages now play.
• By evening the game found its true frame: the board is now the atlas itself. The full-screen night map carries the round, and the photo, guesses and hints float over it in the same panel the homepage uses. On a phone the panel tucks behind the top bar so the whole world is tappable — your latest guess rides beneath the bar while it’s closed, and the reveal opens on its own when the round ends.

2 August 2026 — Spain nearly doubles — the promise to the thin countries, kept

• When this atlas reached the Hacker News front page in July, the sharpest criticism was coverage: Castilla y León showed nine castles where some three hundred stand, and Germany, Ireland, Sweden and Jordan were nearly as thin. Today’s release is that promise kept, and the mechanism matters: the same country-scoped Wikidata sweep that once opened China now runs for Spain, Germany, Poland, Ireland, Sweden and Jordan, joined by a Swedish-language sweep — the 22nd Wikipedia edition the atlas reads. The bar did not move: every entry still needs a photo, coordinates and a Wikipedia article. The atlas grows from 5,814 to 6,412 landmarks.
• Spain grows from 230 to 433 — the largest single-country jump the atlas has had. The Walls of Ávila and the Roman Walls of Lugo, both World Heritage; Seville’s Torre del Oro; Toledo’s medieval gate circuit from the Puerta de Bisagra to Bab al-Mardum; Numantia, where Rome needed twenty years and its best general; the Galician castros from Santa Trega down the Atlantic coast — and across Castilla y León the castles of Ciudad Rodrigo, Aguilar de Campoo, Puebla de Sanabria and dozens more.
• A Warsaw reader’s July complaint is now answered in full. The Citadel joined in July; today the nineteenth-century ring of forts around the city follows it — Fort VIII Służew, Mokotów, Włochy, Wawer and their brothers, with the two Zegrze works up the Narew. Poland stands at 279 landmarks. Thanks again to I claim sanctuary.
• Ireland grows from 107 to 168: Dun Aengus on its Aran cliff edge, Waterford’s Reginald’s Tower — Ireland’s oldest civic building — and the Martello tower where Ulysses opens, an 1804 fortress and the atlas now says so. Sweden grows from 131 to 196: Visby’s medieval wall, the Öland ring forts from Sandby borg to Eketorp, and Torsburgen, Scandinavia’s largest ancient hillfort. Jordan adds Qasr Bshir — perhaps the best-preserved Roman fort standing anywhere — and only three companions, because every further Jordanian site now needs a better Wikidata entry, not a better query.
• The sweep also surfaced what this atlas will not carry: two modern border fences, Franco’s Pyrenean bunker line, a Cold-War data centre in a Stockholm rock shelter, three lighthouses, four Roman cities and a Dublin school were all typed as fortifications somewhere in Wikidata, and every one was excluded by hand. A curated atlas earns its keep by what it refuses. The count stands at 6,412 landmarks across 138 countries — Germany, whose tens of thousands of castles nobody can map one by one, now carries 450 of its most notable.

2 August 2026 — The Andes join the atlas — and Westminster takes its seat

• A thread on Lemmy’s map community did what good threads do: it found the blind spots. The sharpest came from @groet@feddit.org — Peru’s mountains are covered in Inca ruins, and this atlas held four Peruvian entries. The cause was the citadel lesson all over again: Wikidata types the great Andean sites “archaeological site”, a sibling of every class the atlas sweeps, so no castle, fortress or palace query could ever see them. Twelve hand-verified sites join today: Machu Picchu — 1.3 million readers a year, instantly the second-most-read landmark on the whole atlas — Sacsayhuamán, Ollantaytambo, Písac, Choquequirao, Chan Chan, Vilcabamba, Vitcos, Huchuy Qosqo, Waqrapukara, Tambo Colorado and Pikillaqta. Temples, terraces and geoglyphs stay out of scope, and Peru certainly holds more — improving an Inca site’s Wikidata entry (coordinates, a Commons photo, a Wikipedia article) is the way onto the map.
• @ThanksForAllTheFish@sh.itjust.works asked the obvious question: where is the Palace of Westminster? Wikidata types it “parliament building”, so a castle atlas never saw Britain’s most famous palace. It is legally still a royal palace, was a principal royal residence from the eleventh century until the fire of 1512, and now stands where it belongs — straight into England’s top five. The same comment challenged the Palace of Whitehall’s high rank: that rank is honest. It comes from the palace’s own article being read 213,000 times a year — for 168 years this was the monarch’s main residence, with more than 1,500 rooms — and the atlas marks it as the ruin it is.
• @porcelainpitcher@lemmy.today wondered what Melbourne or Sydney could possibly show, and the thread answered itself: Fort Nepean at the Port Phillip heads (suggested by @FluidBeef@quokk.au), which fired the British Empire’s first shot of the First World War, and Brisbane’s moated Fort Lytton (@geolaw@aus.social). Sydney itself adds Bare Island Fort and the Middle Head Fortifications, and Newcastle adds Fort Scratchley, which returned fire on a Japanese submarine in 1942. Australia grows from three landmarks to eight.
• @doubtingtammy@lemmy.ml pointed at a class the atlas had never considered: America’s castle-style National Guard armories, “more castle-like than the fortresses listed”. The three named prove the point and join as castles — the towered 1895 State Armory in Springfield, Providence’s Cranston Street Armory (“the Castle” to its city) and the crenellated 1893 Lynn Armory.
• The atlas now counts 5,814 landmarks across 138 countries.

31 July 2026 — The atlas now reads French

• Every page of this atlas now exists in French at /fr/ — all 5,793 landmark pages, 138 country pages, every ranking, the city day-trip guides, the statistics, data and about pages, and the daily game. It is the seventh language; the switcher now runs English → 日本語 → Deutsch → Español → Français → Polski → Nederlands.
• French Wikipedia covers this subject like almost no other edition: 3,394 landmarks — 58.6% of the atlas, second only to German — open with their frwiki lead, and le château de Versailles reads as fr.wikipedia wrote it. French city exonyms cover 99% of the pool (Londres, Cracovie, Ratisbonne). The rest keep the English story, marked lang="en" and credited honestly.
• French fuses its grammar into the map: every country stores its full prepositional phrase (en France, au Portugal, aux Pays-Bas, à Malte — none of it derivable from the name), landmark articles are read off each frwiki lead with their casing (le château de Versailles, l’Alhambra), and "de" contracts on its own (près du Caire, la renommée du château de Foix). Centuries are Roman (XIIIe siècle), BC is av. J.-C.
• Shipping French also fixed five gaps the Spanish edition had shipped with, from the English fame sentence on every /es/ castle page to the missing compare table on /es/castles/rankings/ and the untranslated near hub — all healed in the same build.
• The About page now carries visible Changelog and Privacy buttons at the top in every language, and the sister-atlas links across the site now read TrainRouter · Sunshine Atlas · Detourmap · FilmMap.

31 July 2026 — The atlas now reads Spanish

• Every page of this atlas now exists in Spanish at /es/ — all 5,793 landmark pages, 138 country pages, every ranking, the statistics, the free-data page and the daily game. It is the sixth language, and the language switch now runs English → 日本語 → Deutsch → Español → Polski → Nederlands.
• What is genuinely Spanish is the landmark itself wherever Spanish Wikipedia has an article — 2,581 of them, 44.6% of the atlas and the best coverage after German: la Alhambra opens with the eswiki lead. The rest keep the English name and story, marked lang="en" and credited to the English Wikipedia by name. The 240 city day-trip guides, the statistics, data and about pages shipped in Spanish the same day — nothing quietly falls back to an English page.
• One Spanish fact is stored rather than guessed: which countries carry an obligatory article — en los Países Bajos, en la India, en el Líbano. Cities use their Spanish exonyms (Múnich, Viena, Varsovia), centuries are Roman (siglo XIII), and the interface is machine-authored pending a native review, which the Spanish footer says out loud.

30 July 2026 — The little word English was missing too

• A day after German and Dutch got their articles, an audit caught English dropping its own: pages said “Castles in Czech Republic” and “in United States”. Seven of the atlas’s countries carry “the” in English — the Czech Republic, the Netherlands, the United States, the Philippines, the Dominican Republic, the United Arab Emirates and the Democratic Republic of the Congo, with “on” for the Isle of Man — and every title, heading, question and cross-link now says so. Three German lines had slipped through the same net (“Nr. 4 in Schweiz” on ranking cards and two FAQ answers); they use the German article table now, and a build test keeps every language honest at once.
• The categories then got the same treatment: 76 landmarks were filed under a chip their own sources contradict. 32 castles are now palaces — Rosenborg, Chapultepec, the Innsbruck Hofburg, Floors and a cluster of Hungarian mansions, all named “Castle” in English only because English has no word for a slott — the sources being the English lead’s own defining noun and, new today, Wikidata’s one-line description. 44 more are now ruins, largely on the Polish and Dutch Wikipedia leads, which state plainly (“– ruiny zamku”, “is de ruïne van”) what the English article never gets around to; Masada, Erebuni and Seoul’s two Baekje earthen fortresses joined by hand. Fortified country houses stay castles — Hever’s moat and portcullis, Appleby’s Norman keep, Broughton, Picton — each now pinned with its reason, and restored sites whose foreign lead is stale (Kaunas, the re-roofed Palace of the Porphyrogenitus) stay standing.
• Two forts moved to the country they stand in, per the atlas’s rule that attribution follows the ground: Fort Du Bus, an 1828 Dutch colonial fort in West Papua, is now under Indonesia, and Fort Piscadera under Curaçao — beside the island’s other forts, which were there all along.
• A Ukrainian fortress had been wearing the title of its Spanish Wikipedia article, disambiguator and all: “Fortaleza de San Pedro (Óblast de Járkov, Ucrania)” is now Fortress of Saint Peter, the 1770 stronghold of the Ukrainian Line. Its address stays what it was — renames never move a URL here.
• The previous changelog entry linked the MCP endpoint as if it were a page; clicking it served a bare API error. It reads as /mcp now — an address for machines, set like one.

29 July 2026 — The little words German and Dutch were missing

• German and Dutch cannot drop the definite article the way this site did: it said “ist Chinesische Mauer” where a German reader expects die Chinesische Mauer. Which names take an article is not guessable — die Alhambra does, Schloss Neuschwanstein does not — so the atlas now reads it from the opening words of each landmark’s own German or Dutch Wikipedia article, where native editors wrote the name with its article in place. Where the article opens some other way, the name stays bare, exactly as before. Sentences that start with a landmark get the capitalised form: “Die Chinesische Mauer steht weltweit auf Platz 1.”
• The long ranking tables — all landmarks by fame, by age and the medieval list — now scroll properly on phones. Below tablet width the 5,793 rows lay out as compact blocks instead of a five-column table in a sideways scroller, and the browser skips the rows far off screen instead of drawing all of them at once.
• Every map landing and game page now ships a proper main heading for readers and crawlers that run no scripts — previously the heading only appeared once the interactive map loaded.

29 July 2026 — Facts the atlas had outgrown

• Several sentences on this site named a landmark or a country in prose, and the data moved on without them. The homepage FAQ said the Palace of Versailles ranked #1; it has been second since the Chinese intake put the Great Wall of China above it. The statistics page said Scotland led the world in castles per head among countries of more than five million people; Slovakia does, with 23.3 per million, ahead of Ireland and Austria. The same page dated the atlas's oldest landmark to the 8th century BC, while the oldest list three clicks away had long since computed Aniba in Egypt, c. 3000 BC. All of these now read the answer off the data, so none of them can drift again.
• The fame ranking described its leader as "among the most-visited historic residences in the world" — written when a palace led, and the Great Wall is nobody's residence. It also counted 192 "Wikipedia language editions", when 20 of those 192 are Wikivoyage, Wikiquote, Wikinews and Commons. Both are now stated the way the page's own table always had: Wikipedia editions and sister projects, plus the readership the article actually has.
• The about page said a landmark needs an article in "any of thirteen languages". That threshold moved to 22 editions in July and nobody updated the sentence; it now counts the editions that genuinely contribute, and so does the homepage. The oldest and oldest-by-country pages said founding years come from Wikidata alone, while their own body text explained — correctly — that the Wikipedia lead fills in where Wikidata is silent.
• The MCP server (/mcp), which answers questions from AI assistants, was describing an atlas of "2,400 castles across 131 countries" with Versailles at number one. It now reads all three from the dataset it serves, so a refresh can never leave it quoting the old one.
• Two league tables on the statistics page were quietly skipping fourteen countries that had crossed the ten-entry mark — Brazil, South Korea, Morocco, Saudi Arabia, Taiwan and nine more. They are counted now, and a build warning fires the next time one goes missing.

28 July 2026 — The atlas now reads Dutch

• Every page of this atlas now exists in Dutch at /nl/ — all 5,793 landmark pages, 138 country pages, 240 day-trip city guides, every ranking, the statistics, the free-data page and the daily game. It is the fifth language beside English, German, Japanese and Polish, and it is a complete mirror: nothing quietly falls back to an English page.
• What is genuinely Dutch is the landmark itself where Dutch Wikipedia has an article — 1,616 of them, 28% of the atlas, and nearly every one a Dutch reader is likely to look up. Muiderslot opens with the nlwiki lead and is called Muiderslot, not Muiden Castle. The other 72% keep the English name and the English story, marked lang="en" so a screen reader does not read English prose with Dutch pronunciation, and credited to the English Wikipedia by name. That fallback never becomes the page description: where there is no Dutch article, the description is built from the data instead, in Dutch.
• Two bits of Dutch grammar are not derivable from a country name, so they are stored rather than guessed: islands take op and not in (op Malta, op Cyprus), and a handful of countries carry an article (in de Verenigde Staten, in de Filipijnen). Cities use their Dutch exonyms too — the day-trip guides are for Keulen, Wenen, Praag and Warschau.
• Two older bugs surfaced while wiring it up and are fixed for everyone: the fame-tier bands on the ranked page had a collapsed conditional that gave Polish and Japanese readers an English or bare label, and the Polish city pages lost the tail of their "see also" line, including two of its three links.

27 July 2026 — “Some descriptions are very sparse”

• Laura Jenny reviewed this map for Bright and made one complaint: “Al zijn sommige omschrijvingen wel heel summier: van ‘gebouwd tijdens de Spaanse inquisitie’ tot ‘kasteel’.” Some descriptions are very sparse — from “built during the Spanish Inquisition” down to, simply, “castle”. She was right, and it was worse than she could see from the outside: 387 landmarks carried no description at all, and thousands more carried one that only repeated the labels printed beside it. Dublin Castle said “Castle”. The Grand Kremlin Palace said “Palace”.
• The landmark pages were never the problem — they have always opened with the Wikipedia article. It was the card on the map, which showed Wikidata’s one-line description and nothing else. So the card now falls back to the first sentence of the Wikipedia entry whenever that description is thin: Houska Castle reads “an early Gothic castle in the municipality of Blatce in the Liberec Region” instead of “Castle”. 3,234 landmarks gained a real sentence.
• A description is treated as thin when it is empty, very short, or only names the building type and the place — “Palace in Vienna, Austria” tells you nothing the chip and the country line above it have not already said. Where Wikidata writes something genuinely its own, it stays: the Alhambra keeps “Palace and fortress complex in Granada, Andalusia, Spain”.
• The second remark in the review — that zooming is occasionally awkward — is fair and still open.

27 July 2026 — A Swedish manor house is not a castle

• Sweden had 89 castles on this map and 19 palaces. Among the castles were Tjolöholm, an Arts-and-Crafts country house finished in 1904, and a long row of Skåne manor houses whose own English Wikipedia articles open with the words “is a manor house”. 128 landmarks have moved to Palace & château, and Sweden now reads 62 castles to 46 palaces.
• The cause is a translation. Swedish slott, Danish slot, German Schloss, Czech zámek and French château all become “Castle” in an English title, because English has no separate word for a grand house that was never fortified. This map read the name before it read the building type, so every one of them arrived as a castle. The tell was in its own translations: the German edition calls the Palace category Schloss, so Schloss Hluboká on one page and Hluboká Castle on the other were the same building disagreeing with itself.
• The arbiter is now the Wikipedia opening line this site already prints on every landmark page — it was quoting “Tjolöholm Castle is a country house” three lines under a chip that said Castle. Where that line says manor house, country house, stately home or mansion, the building is a house, in any country. Where it says palace, château or residence the word settles nothing on its own — Windsor Castle “is a royal residence” and Chinon Castle “is a château” — so those need German Wikipedia, which does keep Burg and Schloss apart, to agree before anything moves.
• Beyond Sweden it lands hardest in France, Germany and Czechia. Brissac joins Chambord, Chenonceau and Cheverny, which were already filed as châteaux; Moritzburg, Hluboká and Orlík follow their own languages. So do Castle Howard, Dunrobin and Balmoral, where “Castle” in a British house name is a matter of status rather than architecture. Neuschwanstein moves too, which will look odd: it is the most-visited page on this site and the most famous castle in the world, and it is a palace — a historicist one, begun in 1869, with no defensive purpose whatever.
• A house built to look like a castle stays a castle. Peckforton is a Victorian country house “built in the style of a medieval castle”, and that is the point of it.
• Seven more landmarks became ruins — ones whose opening line defines them as remains or an archaeological site without ever using the word ruin, among them Leicester Castle and the Yoshinogari site.
• One boundary has deliberately not been touched. Castle and fortress are near-synonyms in English prose, and the same evidence would have moved Caernarfon, Conwy and Harlech out of Castle on the strength of a lead sentence that happens to say “fortification”. That is not a distinction a Wikipedia opening line can carry, so it was left alone.

27 July 2026 — The Great Wall was missing

• China had eighteen landmarks on this map — fewer than Luxembourg, which is three thousand times smaller. It now has a hundred and twenty, and the Great Wall of China is among them. It is, by some distance, the most-read landmark in the whole atlas.
• The cause was not missing articles or missing photographs. It was that Wikidata does not file Chinese fortifications as castles: at this atlas’s admission bar the entire country holds three items typed that way. What it holds instead is 城墙 city walls, 城门 gates, 关 passes, 土楼 and 围屋 fortified clan houses, 炮台 batteries and Tibetan 宗 — and every one of those classes is a sibling of the ones the map searched, never a subclass. So the search walked straight past them. The map now reads the class they all descend from, fortification, inside China.
• What that brings in: Jiayuguan and Juyongguan, the Great Wall’s own fortresses, which Wikidata files as mountain passes. Badaling, filed as a human settlement. Mutianyu, filed as a park. Simatai, Jinshanling, Jiankou and Huanghuacheng. The walls of Xi’an and of Nanjing, and Beijing’s surviving gates. The Old Summer Palace, burned in 1860 and never rebuilt, and Chengde, the Qing summer capital, both filed as gardens. The Silk Road’s abandoned cities — Jiaohe, Gaochang, Tongwan, Suoyang. The Fujian tulou and the Kaiping diaolou. The forts of Macau and Hong Kong.
• Zhonghua Gate in Nanjing deserves its own line. Four courtyards deep, with twenty-seven vaults that could hide three thousand soldiers, it is the most complex city gate ever built — and it carries no classification on Wikidata at all. Nothing about it is obscure; it simply had no type for a search to match.
• Not everything Wikidata calls a fortification belongs on a castle map, and the honest exclusions are worth stating: the fence along the North Korean border, the nuclear shelter under Beijing, a Cold War command tunnel, a working monastery and a single concrete pillbox beside a Shanghai road are all fortifications by classification, and none of them is here.
• Beyond China, the same refresh lifted the countries that were thinnest — Iran, Ukraine, Brazil, Turkey — and added the first landmarks in Sudan, the Democratic Republic of the Congo and the Bahamas. The atlas now holds 5,793 landmarks across 138 countries.
• This finishes something started on Hacker News, where shevy-java counted seven castles in the whole of China. Merging three duplicate country pages in July got that to seventeen, and it was fair to wonder whether the rest was ever coming. It was the classification all along.

26 July 2026 — The atlas speaks Polish

• Polish joins English, German and Japanese as a full edition: the map, every castle page, all 131 countries, the 237 city guides, the rankings, the statistics and the changelog. Castle descriptions come from the Polish Wikipedia where it has an article.
• The language switcher now shows a flag rather than a two-letter code, which was cramped on a phone and unreadable on Windows, where emoji flags render as black rectangles — these are drawn as inline SVG instead.
• The rankings gained a second number. Alongside its place in the world, every landmark now shows where it stands inside its own country — on its page, on the hover card, on country pages and as a sortable column in the compare table. One shared ranking, in every language.

26 July 2026 — Fort Delaware, from a comment thread

• A reader on Neatorama, Cluck2, wrote that the atlas was “missing data even when the data is part of the data set they claim to be using” and named Fort Delaware. They were right: it is typed a fort in Wikidata, it has a photograph and a coordinate, and it is read some sixteen thousand times a year — it was missing only because four Wikipedia editions cover it, and the atlas fills down to eight. The Civil War prison fort on Pea Patch Island is on the map now, dated 1859, the year its granite was finished.
• Its date needed a hand, too, and the fix is general: the year reader had taken “c. 1895” from the only date in its Wikipedia lead — a gun battery built inside the fort in the 1890s. A hand-verified year now also clears the “c.”, which it should always have done.

26 July 2026 — Every page, all the way down in its own language

• This changelog is in Japanese now — all twenty-nine entries, not only the newest — and the Japanese page carries its own headings and date format. That is the visible half of a less visible fix: every ranking and list page was still printing castle names in English on the German and Japanese editions. The fame ranking, the oldest and medieval lists, the by-country pages, the big compare table and all the structured data behind them now read 姫路城 or Burg Himeji wherever the reader’s own Wikipedia has a name for it — about seventy-five places in the page builder had reached straight past the translation. Search still answers to either name.
• The 237 day-trip guides carry Japanese city names — 京都市, ミュンヘン — and the map’s embed snippet now speaks the language of the page that copied it.
• And the daily game is trilingual: 城当てゲーム and Burg raten join Guess the Castle, with the hints, the board, the share text and the castle names in your own language. A Japanese player can type 姫路 and be understood — until now the name matcher threw away every non-Latin letter before it compared anything.

26 July 2026 — Japan, counted the way Japan counts it

• A landmark enters this atlas when eight Wikipedia language editions cover it. That is a fair rule for a world map and a poor judge of Japan, where a castle can be famous for four hundred years and written up in exactly one language. s.ogawa named the measure that Japan actually uses: what a Japanese reader means by “castle” runs as far as 日本100名城 and 続日本100名城 — the hundred castles the Japan Castle Foundation chose in 2006, and the further hundred it added in 2017. Both lists are now a source in their own right, read from Japanese Wikipedia’s own categories together with Wikidata’s membership statements, because neither is complete alone. 74 landmarks joined and Japan went from 155 to 229 — among them Yonago, Naegi on its granite boulders, Tatsuoka, which is the only star fort in Japan besides Goryōkaku, Kaneda of 667 on Tsushima, the Yayoi ditches of Yoshinogari, and the Ainu chashi of the Nemuro Peninsula, which is number one on the list because the list is numbered from the north.
• The gap had exactly the shape they described. They were glad to find Hakodate’s Benten Daiba here — it is on neither list, and they could name the people who would object to it — while Shinoridate and Katsuyama-date, the two of southern Hokkaidō’s Twelve Mansions that the 2017 list does carry, were both missing. An atlas holding the arguable entry and not the canonical two has its rule wrong, not its data. Both are on the map now.
• Underneath, Japanese Wikipedia feeds the atlas’s data now and not only its translations. The sweep that admits a castle documented in German or Czech but not English runs over the Japanese edition too, and over two Wikidata classes that no query starting at “castle” could ever reach: Japanese castle and castle ruin in Japan are siblings of castle rather than subclasses — the same blind spot that hid 38 citadels until July. Where Wikidata held no coordinate, the Japanese article’s own map template gave one: Arikoyama, Kamaha, Hachimanyama, Shōzui and Genbao are placed from it. Where Wikidata held no photograph, Sugiyama arrived with the one at the head of its Japanese article. And Kaneyama in Mino was sitting on Naegi’s summit, forty kilometres east and accurate to the metre — a collision check on every intake is how that kind of error surfaces; its own article had it right.
• Two entries on the lists are still not here: the Shinagawa batteries in Tokyo Bay and 河後森城 Kagomiri Castle in Ehime. Every landmark on this map has a photograph, and no freely licensed photograph of either exists on Wikimedia Commons — which is where they stop, and where whoever uploads one puts them on the map at the next refresh. Three others needed the name the list uses rather than the one Wikidata carries: the Ashikaga clan residence and the Kitabatake clan residence are filed in Wikidata as the temple and the shrine standing on them today — it is their moats and earth ramparts that are protected, and an atlas that lists places of worship as castles is precisely what this one should not be — and Ōno Castle in Chikuzen, the 7th-century mountain fortress, shares its name with the Ōno Castle in Fukui that was already here.
• One more, and it is the question that started this: s.ogawa asked whether the Dōnan Twelve Garrisons were on the map, and they were not. Eleven of the twelve have no Wikidata entry at all, so eleven pins cannot exist — but the group has one, and the group is now a single pin: it stands at the centre of the six municipalities the garrisons were strung across, from Hakodate to Kaminokuni. Its picture is the map of where the twelve stood, English Wikipedia’s own, because no photograph of them exists on Commons except of Shinoridate, which is its own entry. The atlas stands at 3,812.
• One rule changed for Japan as a whole. English articles about Japanese castles state the condition in a sentence the atlas had not learned to read — “Its ruins have been protected as a National Historic Site since 1932” — and a 山城 that is stone bases and earthworks is a ruin under this atlas’s own rule, whatever it was built as. That moved twelve landmarks already on the map, Oka, Hachigata, Tokushima and Hachiōji among them, and settled the state of nineteen of the new arrivals. Two that the sentence would have caught are left alone: Gujō Hachiman and Sumoto both carry a 20th-century keep that their English articles never mention and their photographs are of.

24 July 2026 — What counts as a castle — the rule, written down

• The most-repeated question in the Hacker News thread was never really about a missing castle: it was “what is your definition, and why is it applied so inconsistently?” The atlas had no answer anywhere on the site. It does now — the About page states the rule plainly. Castle & château is a fortified residence, fortress & fort is built for a garrison rather than a household, palace is a grand residence with no serious defensive purpose — and ruin is not a fourth kind of building but a state, which is the honest answer to why the Bastille and Fort Duquesne are filed as ruins with nothing left of either above ground. The page also gives the order the classifier actually works in — Wikipedia lead first, then the landmark’s own name, then Wikidata’s type, then the English description — and says which words stay ambiguous on purpose: château and Schloss mean castle in one valley and palace in the next. Thanks to Bedon292, relevant_stats, drstewart, dbspin, Svip, insane_dreamer and lo_zamoyski, who all asked the same question from different directions.
• Three North American forts the thread named and the atlas still lacked. Fort Pitt Blockhouse — 1764, the oldest building in Pittsburgh and the one piece of Fort Pitt still standing, asked for by someone who found only Fort Duquesne and pointed out its site holds no ruins to visit. Fort Wayne in Detroit — exactly the right catch: the atlas showed the vanished Fort Detroit but not the star fort downriver that is still there. And Fort Anne at Annapolis Royal, Canada’s oldest National Historic Site. All three had sat below the eight-language notability floor, which is precisely why the fame-ordered fill never reached them. Thanks to Brendinooo, garciansmith and yabones. The atlas stands at 3,737 landmarks.
• And the buttons on the map panel and in the daily game now press like the ones on every other page: their raised lip shrinks as the button sinks into it, instead of sliding down with it.

24 July 2026 — A Japanese edition, and cleaner type

• The atlas now speaks Japanese. Every page — the map, all 3,700+ castle pages, the country and ranking pages — has a Japanese edition, with castle names and histories drawn from Japanese Wikipedia where they exist. All 155 Japanese castles, from Himeji to Osaka, read in native Japanese; landmarks without a Japanese article fall back to English. It joins the German edition — switch languages from the map or any page header.
• Headings across the whole site changed from a serif display face to the same clean sans-serif as the body text — one modern, legible typeface everywhere, matching the Japanese pages’ gothic, and one fewer web font to download.
• On phones the map header was rebuilt: the title, a language switch and a light/dark toggle now sit as compact, equal-height controls instead of crowding into a single bar.

23 July 2026 — Castles near you — day-trip guides for 237 cities

• The atlas can now answer the question people actually ask: what castles are near me? 237 city guides — from London and Munich to Kyoto and Jaipur — each listing every landmark within 100 km of the city centre: the closest, the most famous, the oldest, and all the rest in day-trip bands, each with its straight-line distance and compass direction. Border cities show their whole circle honestly — Basel’s reaches into three countries.
• Everything is computed, nothing is written by hand: the cities come from Wikidata (population, coordinates, German exonyms), the castles from the atlas itself, and a city qualifies by having at least seven landmarks in day-trip range — so as the atlas grows, new cities earn their guide automatically. The guides exist in German too, and every country page now links its cities’ guides.
• One departure the new Munich guide made visible: the Braunes Haus — the demolished NSDAP headquarters, which Wikidata types as a palace — is gone, under the same rule that removed the Grafeneck and Hartheim memorial sites.
• A reader in Warsaw caught the atlas showing two buildings that are not there. The Saxon Palace and Brühl Palace were blown up by German troops in December 1944 — of the Saxon Palace only the colonnade sheltering the Tomb of the Unknown Soldier still stands — and both were drawn as ordinary standing palaces. They are marked as ruins now, the same way the Bastille and the Tuileries are. Reconstruction of both is under way and due around 2030; when they stand again, so will their pins. Thanks to I claim sanctuary.
• The same reader asked the sharper question — “even the giant Warsaw Citadel is missing… wonder what was the query?” The query was the problem. The atlas searches Wikidata’s castle, fortress, fort, château and palace classes, and citadel is none of them: it is a sibling class, not a subclass, so every landmark filed only as a citadel was invisible no matter how famous. Of the 93 citadels that meet the atlas’s bar, 38 were missing. That hole is closed, and it was a big one: Arg-e Bam, the largest adobe building on earth; the Citadelle Laferrière, the largest fortress in the Americas; the Citadel of Erbil, on a mound inhabited for six thousand years; Vietnam’s two UNESCO citadels at Tây Đô and Thăng Long; Berlin’s Spandau Citadel; Naryn-Kala above Derbent; the citadels of Namur, Qaitbay, Ghazni, Făgăraș and Erfurt, and the Tower of David at the Jaffa Gate in Jerusalem. And, at last, the Warsaw Citadel — beside Modlin, whose presence made the gap so glaring — together with the Przemyśl Fortress, the great siege of the Eastern Front.
• Another reader found the atlas listing the Schlosshotel in Kronberg but not the actual castle in Kronberg. Exactly right, and an honest explanation: a landmark needs eight Wikipedia language editions to earn a place, and Burg Kronberg has six while the Schlosshotel down the hill has seventeen. Fame is not the same as importance; the castle is on the map now. Thanks to scorbet.
• Two records were also refiled by where the landmark actually stands rather than by the country field Wikidata ranks highest. Ateret Fortress, the Crusader castle at Jacob’s Ford on the Jordan, was listed under Palestine because its country value points at the historic region of that name rather than any modern state — the same slip that once put a castle in the “Kingdom of England”. It stands in the Upper Galilee, and it is filed under Israel now, as is the Tower of David inside Jerusalem’s Old City walls, alongside the Antonia Fortress 750 metres away.
• And one reader admitted the first thing they did was check for Fort Chambly. It was not there — it sat exactly on the notability line and fell off the end of a query. It is there now. Thanks to Francois Dejeuner. With 25 arrivals, the atlas stands at 3,733 landmarks.
• And the homepage is now just the map — on every screen. The long about-section that used to sit below it (browse links, FAQ, the how-it’s-built story) no longer renders; phones never showed it, and now nobody scrolls. The text stays in the page for search engines, and everything it linked is one click away in the map’s panel: all castles, the fame ranking, how it’s built.
• Every button is a proper button now — and it presses. The controls on the map panel, in the game and across the page headers became solid teal with bold cream text, each sitting on a raised lip: hover and it nudges down, click and it sinks flush into its shadow. It is the tactile button from our sister atlas TrainRouter, in the house palette — one button language everywhere.
• The browse page got a search bar: type a name or a country on the all-castles page and the 3,733 cards filter as you type — accents optional, so “chateau” finds the Château de Chambord. Collapsed sections open themselves while you search and step aside when they hold no matches.
• A second founding-date pass: 320 more landmarks now carry the date their own Wikipedia article states — 129 precise (“built between 1813 and 1821”), 191 approximate (“a 12th-century castle”, shown as “c.”). The reader learned new idioms — year ranges, “the 790s”, “early” and “late” centuries — and it now reads German leads too, which dated 116 landmarks that have no English article, Masada (c. 37 BC), the Palace of Aachen (c. 795) and Eltz Castle (c. 12th century) among them. 71% of the atlas is dated now, up from 62% — and the 1,090 landmarks whose articles state no date stay honestly undated.

22 July 2026 — The castle on the 1-cent coin — a third wave from the thread

• Italy’s most embarrassing absence is fixed: Castel del Monte — Frederick II’s octagonal crown, a UNESCO site, the castle on Italy’s 1-cent coin — was invisible to the atlas because Wikidata now types it as a set of museum classes, not a castle; no castle query could ever find it. It is joined by its Hohenstaufen kin Castello Svevo in Bari, the sea-front castle of Trani and the Norman castle of Gioia del Colle, plus Lerici Castle on the Gulf of Poets and Santa Severa Castle — the very item one reader looked up on Wikidata and wondered about. Thanks to _ZeD_, deppep, ipunchghosts and riffraff.
• Africa grew where a reader showed us it should: Lamu Fort and the UNESCO dry-stone enclosures of Thimlich Ohinga join Fort Jesus in Kenya (the Old Fort of Zanzibar and the fortress of São Miguel in Luanda were already on the map). And Jordan gains Machaerus, Herod’s hilltop fortress over the Dead Sea. One honest limit: Somalia’s Dervish forts at Taleh have no usable Wikidata item yet — improving them there is the way in. Thanks to uxhacker and aorth.
• The Citadel of Acre was hiding in plain sight. We first reported it had no usable Wikidata item — wrong: Acre is a castle city, and its citadel is filed under the scholarly label “Hospitaller commandery of Saint-Jean-d’Acre” (the German Wikipedia calls it plainly the Zitadelle), which no citadel-shaped search ever finds. The stronghold of the city Napoleon failed to take in 1799 is on the map now. Thanks to dmblit.
• The Palmerston Forts exist here now: Fort Nelson on Portsdown Hill, the three Solent sea forts — Spitbank, No Man’s Land and Horse Sand — Fort Brockhurst in Gosport, and Plymouth’s Crownhill Fort. Thanks to hermitcrab and CrzyLngPwd.
• Three more by request: Vanbrugh Castle, the playwright-architect’s own “Bastille” in Greenwich; Fort Vasquez, the rebuilt adobe fur-trading fort in Colorado; and Fort Knox — the granite one in Maine, the first fort to carry the famous name and, unlike the Kentucky army post, an actual visitable fortress. Thanks to philipwhiuk, Stealthisbook, modeless and the111.
• One departure: Ljubljana’s National Assembly Building — a working 1959 parliament that Wikidata types as a “palace” — is gone, under the same rule that removed the North Carolina State Capitol. The atlas stands at 3,709 landmarks. Thanks to ajsnigrutin.
• And the hover card stopped running away: it used to follow the cursor, so chasing its links could feel like a duel. It now anchors once, at the castle’s own dot, and stays put while you reach it. Thanks to drcongo.

22 July 2026 — Browse with pictures — and 210 truer ruins

• The all-castles page and every country page are picture browsers now: each landmark as a card with its photo, century, fame rank and a one-sentence story, most famous first — the same format as our sister atlas’s route browser.
• Ruin detection got sharper, in four passes — the last one auditing Wikidata’s own ruin typing for every landmark. The classifier now reads the freshly fetched Wikipedia leads — including German Wikipedia for landmarks that have no English article — and also recognises destroyed and damaged sites: leads that define a site by its remains, narrate destruction that was never rebuilt, or describe a derelict state. 210 landmarks moved to Ruins (472 → 682), the last pass reading the condition wherever the lead states it — often the second or third sentence — plus a hand-audit of vanished sites like Nonsuch and Richmond Palace, from Slovak and Polish hilltop castles to vanished Dutch castles and Madrid’s Buen Retiro. Every pattern is guarded against the “destroyed… then rebuilt” trap, and English Wikipedia stays authoritative where it exists — German Wikipedia calls even well-preserved roofless castles like Caernarfon a “Burgruine”.

22 July 2026 — The atlas now reads Wikipedia in thirteen languages — 1,255 new landmarks

• The sharpest criticism in the Hacker News thread was aimed exactly right: requiring an English Wikipedia article kept whole regions off the map — a Burg documented only on German Wikipedia, or a castello only on Italian, could never enter. The atlas now sweeps twelve more editions (German, French, Italian, Spanish, Polish, Czech, Slovak, Romanian, Hungarian, Portuguese, Dutch and Greek) alongside English, and stands at 3,693 landmarks across 131 countries. Every entry still needs a real photo, a Wikipedia article and exact coordinates. Thanks to pmontra, superhuzza and Bedon292 for pinpointing the cause.
• The regions the thread called out all grew: Italy 179 → 288, France 183 → 276, Spain 88 → 181 (Castilla y León very much included), Poland 77 → 163, Germany 176 → 254, Switzerland 27 → 91, Austria 68 → 133, Portugal 22 → 82, the Netherlands 22 → 76, Romania 19 → 74 and Hungary 13 → 64, with Czechia, Slovakia and Greece close behind — and Australia finally has its dot: Fort Denison, in Sydney Harbour. Thanks to ainiriand, mxfh, stefanka, AStrangeMorrow, notorandit, ipunchghosts, weezing, limaoscarjuliet, cozzyd, comrade1234, hurghada, metada5e, andsoitis and busymom0.
• 285 of the additions are documented only in a non-English edition, and their pages say so plainly: the Wikipedia link is labelled with its language (“in German”), the English-readership half of the fame score honestly counts them at zero — so they enter as hidden gems — and the how-to-add-a-castle guidance now accepts an article in any major language.

22 July 2026 — Eleven landmarks you asked for by name — and cleaner data behind them

• On the map now: the Acropolis of Athens (the ancient citadel itself), Caisteal Maol on Skye, Boldt Castle in the Thousand Islands, Jim Bishop’s Bishop Castle in Colorado (removed again on 2026-08-08 — see that day’s entry), Fort Point under the Golden Gate Bridge, the Dover Western Heights, the Esslinger Burg, and four Romanian strongholds — the Fortress of Arad, Oradea Citadel, Râșnov Citadel and Rupea Citadel. Thanks to YeGoblynQueenne, 00deadbeef, etimberg, Stealthisbook, shagie, mr_toad, frm88 and rob74.
• Hama-rikyū Gardens left the atlas: it is a Tokyo garden — its palace burned in 1724 and was never rebuilt. Thanks to throwaway2037.
• Elmina Castle’s fame was broken by a data gap: its readership had never been fetched, so the oldest and largest European building in sub-Saharan Africa sat a tier below its neighbour Cape Coast Castle. Fixed — with some 74,000 yearly readers and 35 language editions, Elmina now ranks where it belongs. Thanks to prmph.
• China is one country page now, not three: “People’s Republic of China”, “China” and “Tibet” each carried separate entries, which is why the map seemed to show almost no Chinese landmarks — all 17, from the Forbidden City to the Potala Palace, now live under China. The Potocki Palace in Lviv, filed under “Austria–Hungary”, sits correctly under Ukraine. Thanks to shevy-java.

22 July 2026 — Approximate dates join the age rankings, wearing their “c.”

• Until today a landmark dated “13th century” was excluded from the oldest-castles ranking so an approximation could never pose as precise. That was honest but too timid — it silently hid genuinely ancient sites. Approximate dates now rank everywhere: oldest, medieval, oldest in every country, the statistics and the big compare table — always marked “c.”, and sorting after precisely dated sites at the same year. 145 approximately dated landmarks join the lists.
• The same pass fixed an inconsistency: country pages already used approximate years in their “oldest landmark” lines but displayed them bare, as if precise. Every year shown anywhere now carries its “c.” when it is one.

22 July 2026 — Map polish from the thread: hover cards you can click into, colours told apart

• The hover card’s links were unreachable — the card vanished the moment the cursor left the dot. It now waits: a short grace period, and the card holds itself open while your pointer is over it, so “Full story” and Wikipedia are one hover and a click away. Thanks to bwhiting2356.
• Ruins are heather-violet now. The old slate blue sat too close to both the castle cream and the palace teal at dot size — the four categories finally read apart at a glance. And card photos keep their own aspect ratio (up to a cap) instead of a fixed crop that could hide the castle itself — Bamburgh no longer vanishes behind its dune. Thanks to 00deadbeef.
• Cards can no longer be cut off by the map’s edge: a card taller than the frame allows now scrolls within it, and it opens on whichever side of its dot has more room — so “Full story” and Wikipedia are always reachable.

22 July 2026 — Founding years for hundreds more castles

• Wikidata leaves about 40% of the atlas without a founding year. Where a landmark’s Wikipedia lead states when it was built, that year is now read in — 316 more castles are dated, bringing the atlas to roughly three-quarters. Only a date the source actually states is used; the rest stay undated rather than guessed.
• Century-level or hedged dates (a “13th-century castle”, “around 1240”) show a “c.” and are kept out of the oldest-castle rankings so an approximation never poses as a precise date. The map’s year slider now covers most of the atlas, and the open dataset gains a year_approx flag.

22 July 2026 — More from the Hacker News thread: five castles added, the data sharpened

• Five castles you told us were missing are on the map now: Hearst Castle in California, Toronto’s Casa Loma, the Citadelle of Québec and the Halifax Citadel, and Larnach Castle — New Zealand’s only castle, which brings the atlas to 130 countries. Thanks to makeusgreat, guyzero, yabones and kinow.
• Two entries that aren’t castles are gone: the North Carolina State Capitol (a working government building) and Porto’s Palácio da Bolsa (a 19th-century stock exchange, never a fortification). Thanks to allturtles and jcfariadias.
• The Bastille, the Château de Saint-Cloud, Fort Detroit and Fort Duquesne — landmarks that no longer stand — are all filed under Ruins (Fort Detroit had been mislabelled a standing fortress). Thanks to dolmen, mr_toad, garciansmith and Brendinooo.
• Clarence House — a working royal residence rather than a famous palace — left the top-ranked tier; its fame score had leaned on royal-news readership rather than architectural stature. Thanks to jameshart.
• A reminder made clearer on the map itself: this is a curated atlas of the most notable castles, not an exhaustive list — so the map now reads “notable landmarks”, and the homepage FAQ explains how the set is chosen and how to add a castle through Wikidata. Thanks to mxfh.

22 July 2026 — Phone taps, fixed — thank you, Hacker News

• On a phone, tapping a castle now opens its card and keeps its links working. The “Full story” and Wikipedia links used to do nothing on mobile — a tap was read as a hover and closed the card instead of opening it. Reported by samfar90 and BasieP2.
• Where dots crowd together, a tap now picks the nearest castle rather than whatever happened to sit on top — so you no longer have to zoom right in to hit the one you meant. Thanks to prmph.

21 July 2026 — Medieval castles, on their own map

• New page: the medieval castles of the world — every landmark in the atlas with a documented founding year in the Middle Ages, roughly 500 to 1500, the most famous first and then the full chronological table. It sits next to Oldest on the rankings hub.
• It is the cut a plain “castle” search never gives you: the ancient citadels drop out below 500, and the 19th-century revival palaces — Neuschwanstein and its lookalikes — drop out above 1500, by founding year rather than by looks. Dates come from Wikidata (P571); undated sites are left out, not guessed.

20 July 2026 — Every category gets its oldest

• The oldest-castles ranking now crowns each category separately: the ten oldest castles, fortresses, palaces and ruins, each in its own anchored section — Nero’s Domus Aurea (AD 64) leads the palaces, Zerzevan (c. 400) the ruins.
• And a new page alongside it: the oldest castle in every country — one national champion per country, oldest first, with the full alphabetical table. The by-age twin of most famous by country.
• The fame ranking got the same treatment: the ten most famous castles, fortresses, palaces and ruins, each anchored. And everything now lives together on one hub — Rankings, in the menu — six views: fame and age; world, country and category.
• The hub’s centrepiece: one sortable table of all 2,435 landmarks — name, country, category, founding year, fame rank, Wikipedia editions and yearly readers. Click any column to sort it your way, click again to flip, and type to filter by castle or country. No framework, no tracking — a two-kilobyte script on otherwise static HTML.

19 July 2026 — The oldest castles, ranked — and 27 dates set straight

• New page: the oldest castles in the world — every landmark with a documented founding year, ranked oldest-first, from the Bronze-Age Avranlo Fortress and the 8th-century-BC Erebuni to the present. It sits next to Ranked in the menu.
• Building it exposed 26 Wikidata founding years that were really renovation or heritage-designation dates — the Tang-dynasty Daming Palace claimed 2010, Chapultepec 1944, Sanok 2010. All corrected to the documented construction start, and Erebuni Fortress now correctly sits under Armenia rather than the long-gone kingdom of Urartu — the atlas spans 129 countries.
• England, Scotland, Wales and Northern Ireland now show their ENG/SCT/WLS/NIR codes everywhere instead of emoji flags that rendered as a plain black flag on Windows.

19 July 2026 — Room to breathe — pages run wide now

• Every content page widened from a 928 px shell to 1,280 px, the same footprint as our newest atlases. Paragraphs keep their readable measure — it’s the country grids, fact panels, ranking tables and castle photos that use the new room.

19 July 2026 — Ruins mean ruins — the state now beats the type

• A follow-up to today’s label fixes: 142 more landmarks that Wikipedia itself defines as ruins now carry the Ruin chip — Heidelberg (“is a ruin… landmark of Heidelberg”), Vlad the Impaler’s Poenari, Alamut, the Cathar Peyrepertuse, Tantallon and much of the great Slovak and Czech hilltop family (Čachtice, Trosky…). Ruins grew from 119 to 261.
• The rule reads each monument’s own English Wikipedia lead, with guards: “built on the ruins of” doesn’t count, “semi-ruined” doesn’t either, and rebuilt sites the text can’t catch — Kokořín, Nassau, Kosava, Nyköpingshus — are pinned by hand to stay castles.
• The atlas now counts 1,031 castles, 427 fortresses, 716 palaces and 261 ruins; the map filter, legend, statistics and the open dataset follow.

19 July 2026 — Fortresses are fortresses again — 83 labels corrected

• Wikidata often types a fortress as a “castle”, and the atlas used to trust that typing over the monument’s own name — so Kufstein Fortress wore a castle chip. The categoriser now lets an unambiguous name win: 68 landmarks became fortresses (from Hohensalzburg and the Citadel of Aleppo to the Château d’If, a fortress despite the château, and the Moscow Kremlin), and 15 became palaces (the Hofburg, Scone Palace).
• Names only rule where they are unambiguous — a French château or a German Schloss can be castle or palace, so there Wikidata’s building types still decide, and hand-checked exceptions cover oddities like the Kremlin Senate: a palace inside a fortress.
• The map’s category filter, the legend, every page label and the open dataset follow along: the atlas now counts 1,152 castles, 441 fortresses, 723 palaces and 119 ruins.

19 July 2026 — Reader requests: Norway and India grow

• Four Norwegian fortresses joined the atlas after @42@snabelen.no pointed out on Mastodon that they were missing — Kongsvinger, the fortified old town of Fredrikstad, Tønsberg Fortress above Norway’s oldest town, and Oscarsborg, whose guns sank the cruiser Blücher in April 1940. Thank you!
• India grew from 71 to 80 landmarks after @vibhurishi@social.linux.pizza noted the country was thin, closing some conspicuous gaps: Golconda, Delhi’s Purana Qila, Kalinjar, Bidar, Chitradurga, Warangal, the City Palace in Jaipur, Ujjayanta Palace and the wooden Padmanabhapuram Palace — and yes, Ranthambore Fort was already on the map.
• The same data refresh brought 22 further castle ruins worldwide, and the Imperial City of Huế is now correctly listed under Vietnam. The atlas stands at 2,435 landmarks in 130 countries.

18 July 2026 — Guess the Castle — a daily game

• The atlas learned to play: Guess the Castle shows one castle photo a day — drawn from the 500 most famous — and gives you six guesses. Every miss zooms the photo out and answers with the distance and compass direction from your guess to the answer, because every guess is a real castle on the map.
• Streaks and stats live in your browser only (no account, like everything here), and the result copies as a spoiler-free emoji grid.
• The answer links straight to the castle’s page and its spot on the night map — the game is a door into the atlas, not a detour from it.

18 July 2026 — “Ink & Bone” — a calmer, more human look

• The whole site moved from the launch palette (dark navy + amber gold) to ink, bone and verdigris — closer to an antique sea-chart. The honest reason: readers whose taste we respect said the old colours looked machine-made. Matt Muir in Web Curios put it with the most style: lose “the horrible orange/brown look”, gain “some vague semblance of humanity”. He had a point.
• New favicon and social-share card to match, rendered from the live dataset.
• This changelog page, with an Atom feed at /changelog/feed.xml.
• The all-castles index was rebuilt for scanning: a country grid with flags in place of a run-on link wall, columned category lists, and an on-page contents line.
• The same treatment for castles by country and tidier action rows on the ranking pages — and on the map itself, “All 2,400 castles” is now a proper button, so the browsable atlas is one obvious click away.
• The display face changed from Cormorant Garamond to Fraunces — a warmer, wonkier old-style serif — completing the redesign the palette began: headers on every page, the map included, now carry it.
• The site header gained the castle mark, the wordmark settled on a single colour, and the content pages grew wider — prose keeps its reading measure while grids and tables breathe.
• Country chips now set a deliberate country code (ENG, SCT, WLS and NIR included) instead of emoji flags, which rendered differently on every platform — and the Netherlands got its missing NL back.
• The map panel slimmed from five controls to three buttons — Browse, About and Surprise me — and the atlas gained a proper about page: what it is, how it is built, who keeps it. The region quick-jump row went too; the map itself is the region picker.
• The map loads faster: the castle data now travels beside the app instead of inside it — the code download halved, data and code arrive in parallel, and each is cached independently.

15 July 2026 — Honest sitemap dates

• The sitemap now re-dates a page only when its content has actually changed, instead of stamping every page on every deploy — kinder to crawlers, truer to the word “lastmod”.

14 July 2026 — German edition, phone ergonomics, a data fix

• Full German mirror at /de/ — the map UI and all 2,400 castle pages, with German Wikipedia content where it exists (about 84% of landmarks).
• Proper iPhone / touch ergonomics across the map and every page: safe-area insets, bigger tap targets.
• Data correction: Jaigarh Fort’s founding year (1726 — Wikidata carried −1000), plus an override list so future data refreshes keep audited fixes.

13 July 2026 — The most famous castle in every country

• New page: one national champion per country, ranked by fame — the Palace of Versailles comes out world #1; the 131 champions span fame ranks #1 to #2,120.

12 July 2026 — Statistics, embeds and an MCP server

• /statistics/ — citable castle statistics computed from the dataset: country league tables, density and per-capita leaders, founding centuries.
• An embeddable, chromeless version of the map at /embed/, free to put on any site.
• An MCP server at /mcp, listed in the official MCP Registry — AI assistants can query the atlas live.
• The dataset mirrored to GitHub, Hugging Face and Kaggle, and made citable with a Zenodo DOI.
• Unknown URLs now return a real 404 instead of the map shell.

11 July 2026 — Stories, fame and free data

• Every castle page gained its story — the lead of its Wikipedia article, with attribution.
• A fame ranking of all 2,400 landmarks at /castles/ranked/, blending Wikipedia coverage with real readership.
• The free dataset page at /data/ — GeoJSON and CSV, CC0.

10 July 2026 — 2,400 landmarks across 131 countries

• The atlas grew from its launch size to 2,400 castles, fortresses and palaces — Japan fixed, the UK split into its four nations, and a Ruins category (97) added.

9 July 2026 — Launch

• Castlemap goes live: the world’s great castles, fortresses and palaces on one interactive night map, each with a photo, founding century and its own page.