Case Study · Geopolitics · Cartography
The Living Atlas
The idea
This site is designed as an atlas, and an atlas without a real map is a bluff. The Living Atlas is the call: a fully modern political map of the world, rendered in the site’s own map-room palette, with toggleable layers: physical geography, cities, maritime chokepoints, political blocs, and a century of border changes under a year slider. The homepage frontispiece keeps its engraved Ortelius map from 1570; the /atlas page is its living counterpart, and the namesake project of the whole site.
The conviction behind it is that layers are how geography becomes argument. A base political map is trivia; the same map with rivers, straits, and alliance tints toggled on becomes a machine for asking questions: why cities cluster where they do, why a handful of narrow waters carry so much of the world’s trade, how the map of 1914 dissolved into the map of 1991. Print atlases have always known this; the web lets the reader work the layers themselves. It’s also, frankly, the project where my geopolitics reading and my engineering habits finally point at the same target.
The foundation
The v1 stack is deliberately boring and deliberately self-contained. Geometry comes from Natural Earth, the public-domain cartographic dataset. Tippecanoe cuts it into vector tiles, which are packed into a single PMTiles archive: a static file, served from this site’s own hosting, that clients query with HTTP range requests. MapLibre GL renders the tiles in the browser, styled with the same design tokens as every other page, so the map reads as a plate of this atlas rather than an embedded widget. No external tile servers, no API keys, no third-party requests: the whole map is files I build and host myself, which matters both for longevity and for the site’s no-tracking posture.
Vector tiles mean the map is drawn, not photographed: every zoom level is the same geometry restyled, and the tile pyramid stays tractable because level contains at most tiles. The projection is Web Mercator, which maps longitude and latitude to
a conformal projection whose price is the famous one: local scale grows as , so Greenland balloons and the poles flee to infinity. For a zoomable political map the tradeoff is right (shapes stay locally true, and the standard, cheap tile pipeline runs on it), but the distortion deserves labeling rather than silence, and the atlas will say so on the map itself.
The first layers
The v1 layer batch, chosen so each one earns its toggle:
- Rivers and lakes: the hydrology that explains most of the political map beneath it.
- Cities: populated places, scaled by importance as you zoom.
- Maritime chokepoints, a hand-curated set of the straits and canals that concentrate world trade: Hormuz, Malacca, Suez, Bab el-Mandeb, the Turkish Straits, Gibraltar, Panama. Small dataset, disproportionate explanatory power.
- Political blocs: membership rendered as country tint (NATO, the EU, and company), the layer that turns the map from reference into argument. Membership is dated, so pulling the year slider paints the bloc as it stood that year, on that year’s borders: the European Communities are six countries in 1957 and twenty-eight in 2019, and NATO in 1955 sits over a map that still says Tanganyika and Ceylon.
- A century of borders: the CShapes 2.0 dataset of historical state boundaries, 1886–2019, under a year slider. Drag it and watch empires shatter, decolonization sweep, and the Soviet Union assemble and dissolve. This is the layer I most want to exist.
- Language families: every language in Glottolog with coordinates, nearly eight thousand dots colored by top-level family (CC BY, Hammarström et al.). This is the layer where the linguistics minor pays rent: one glance shows the Austronesian expansion flung across two oceans, the Atlantic-Congo saturation of sub-Saharan Africa, and the improbable density of New Guinea (more distinct dots per square degree than anywhere on Earth). Click any dot for the language’s name and family.
- Airports, OurAirports (public domain), drawn as rings so they read apart from the solid city dots: 1,176 large airports at every zoom, plus 2,148 medium fields with scheduled service once you’re close enough to want them. IATA codes label the majors. An infrastructure layer, which is to say a layer about where the world thinks its connections belong.
- Terrain and seas, the physical atlas underneath the political one: three hundred named mountain ranges, plateaus, plains, and peninsulas from Natural Earth, labeled at the zooms Natural Earth itself recommends; the oceans, seas, gulfs, and straits lettered in the wide letterspaced caps marine names have carried since Ortelius; and deserts as a faint ochre wash (the one terrain class a flat map can shade honestly, since a desert is a region, not a ridgeline).
- Submarine cables: 530 cable systems and 1,400 landing points, the internet’s actual geography. The layer is deliberately a dated snapshot: TeleGeography’s current data is commercial-only, but their November 2022 GeoJSON was published under CC BY-NC-SA 3.0 (which is irrevocable) and is preserved by the Internet Archive. So the toggle says “(2022)”, the license lives in a sidecar file, and staleness is documentation rather than error, the same rule the atlas applies to lines of control. A cable map three years old is still the right mental furniture: the chokepoints it reveals (the Red Sea funnel, the Luzon Strait, Egypt as a terrestrial shortcut) move on decadal timescales.
The v2 batch widened the atlas considerably:
- Relief, a shadows-only hillshade (later retired for live Terrarium tiles; see the roadmap): first computed from the public-domain ETOPO 2022 elevation grid and baked into a 5 MB raster PMTiles file. Shadows only, because a shadow layer is honest in both themes: it deepens the map room and engraves the parchment without fighting either palette. The terrain toggle also now distinguishes ergs from regs (sand seas stippled, stone deserts flat: a hand classification, since Natural Earth doesn’t type desert surfaces), shows glaciers hatched, washes the rainforest belt (RESOLVE Ecoregions) and lowlands, and labels islands, peninsulas, and depressions.
- Seas got their own toggle: each named sea, gulf, and strait carries a subtly different depth of green, so the water reads as regions rather than void.
- Chokepoints grew from 14 to 54, each ranked (global, regional, emerging, dormant) with rank driving prominence: corridors and gaps that are really lines (the Suwałki Gap, the GIUK line, the Wakhan Corridor, the Darién Gap) are drawn as dashed strokes rather than pins.
- Autonomous areas: 191 of them, from China’s autonomous regions and Russia’s republics through Åland-class special cases (Gagauzia, Vojvodina, Zanzibar, Aceh) to the 99 largest US Native reservations from the Census, which belong in exactly this company. Clicking one, or clicking a dependent territory’s ring, opens a sheet: what the autonomy actually is, the year it began, the office that governs, the population and its year. Where the unit is also a first-order subdivision, that sheet reads its population and capital from the subdivisions file rather than keeping a second copy.
- Disputed areas as a layer of their own: Bir Tawil, the Hala’ib Triangle, Abyei, Aksai Chin, Essequibo, and some seventy more, hatched and labeled.
- Peoples and languages as areas: the GREG dataset (the digitized Soviet Atlas Narodov Mira) provides the only open global ethnic-group polygons in existence, so the atlas uses them the only honest way it can: labeled “(1964)”, a dated snapshot like the cables, with curated fixes (the Garifuna are named). The same polygons, dissolved through a hand-built group-to-family mapping, become the language-area layer.
- UNESCO World Heritage sites (1,852, via Wikidata), ski resorts and major military bases (hand-curated, ~100 each), a majority-religion tint beside the blocs, admin-1 subdivision lines, a native-place-names toggle (self-hosted glyphs for a dozen scripts), country hover highlighting with halos that make microstates findable, and a click-through dossier sheet whose entries are placeholders awaiting their author.
The v3 batch, built out over 2026, is mostly about answering a click with a real number.
- Nuclear, 291 sites carrying every one of the 417 operating reactors the IAEA lists (379,608 MW across 31 countries), scraped from PRIS, plus construction sites, permanently shut stations, the fuel cycle, and the weapons complexes and test ranges. Colour says what a site is for and fill-versus-hollow says whether anything on it is generating, so Japan’s licensed but dark fleet draws at its true capacity as a hollow ring. The per-site sums reproduce PRIS’s own per-country totals exactly, which is the check that matters.
- Protected areas, 396 reserves in 140 countries with their IUCN category. The honest part is the geometry: 200 are drawn to a real boundary, and the other 196 have no outline under a licence that lets me redistribute it, so they draw as a hollow ring at their position rather than as a shape I invented. The legend says so and every centroid feature repeats it in its popup.
- Wildfires, a 24-hour NASA FIRMS snapshot: 465,170 satellite detections clustered into fire events, the 340 largest drawn as discs that scale with detection count, plus 150 fires that NIFC and GDACS name and measure, drawn as rings that scale with burned area. The two marks are separate on purpose: a count of hot pixels is not a burned area. It is also dated, and a late-July capture is a portrait of the boreal and southern-African burning season rather than of wildfire in general.
- The AI supply chain, end to end: 31 data-center clusters, 62 fabs, and the lithium and rare-earth mines that feed them. The fabs are the point, each carrying what it makes (logic, DRAM, NAND, analog, packaging), its smallest node in volume production, and whether it is operating, under construction, or merely announced. That last distinction is the whole argument: read the map and Taiwan, South Korea and Japan hold the memory and the leading edge, while the American and European sites in the headlines are mostly hollow rings targeting 2027, 2028, 2030.
- Ski areas, rebuilt from OpenStreetMap: 642 areas, 626 with a real lift count, the disc scaling with it. Three numbers are deliberately missing, and I measured why before dropping them: OpenStreetMap’s piste ways sum to between 0.86 and 2.6 times published figures, and its mapped station range reads 262 m of vertical at Bogus Basin against a real 546. A number the site cannot defend is worse than a blank.
- Chokepoints, redrawn as gates. The old geometry ran along the axis of travel, so Malacca was a 1,027 km line down the channel, which says nothing about why anyone cares. Each is now a line drawn across the flow, shore to shore, so its length is the width of the opening: Malacca is 24 km. Every gate carries that width, measured from the drawn line where the line is the narrowest crossing and cited where it is not.
- Measured water geometry. 789 named lakes, seas and rivers used to open the sheet with nothing but a name. A lake or a sea is a polygon, so 411 of them now carry a geodesic area and an outline length computed on the WGS84 ellipsoid; a river is a line and has no area, so 347 rivers carry a measured centerline instead. The 31 that carry no measurement at all are the 29 hand-traced sketches, whose extent is published and whose lengths are not, because measuring a five-vertex sketch to three significant figures would be a lie about what the sketch is, plus Drake Passage and Luzon Strait, which Natural Earth draws as lines closed on themselves rather than as areas. The row is labelled “Drawn extent”, never “Area”, because it describes the polygon the map draws rather than the lake, and the coastline paradox is measured rather than asserted: the same lakes at 1:10m come out 1.38 times longer on the median outline while their areas move less than a percent. Rivers get “Mapped centerline”, because Natural Earth names reaches, not rivers, and the Nile measures 2,530 km of its 6,650 under that name.
- What a click answers. The v3 pass was mostly about that. A subdivision gives its capital, population with the census year, area, establishment date and the office that runs it, for 489 first-order units; the office ships and the office-holder does not, because 489 names cannot be kept true and the only source with that coverage is measurably stale. A city gives its civic office in the local language, and the office-holder only where two independent sources agreed and only ever beside the date it was captured. A disputed area gives who administers the ground and who holds the recognised claim as adjacent rows, because the gap between those two rows is the dispute. An airport gives passengers, movements, terminals, runways and the year it opened, for the 390 hubs worth clicking, each figure carrying its own year.
- Enclaves, exclaves and counter-enclaves, out of the borders switch and onto their own toggle, on a closed six-term vocabulary so an exclave is never called an enclave and Llívia can be both. Baarle is derived rather than copied: read the two municipal multipolygons and work the structure out, and you get 22 Belgian enclaves, 7 Dutch counter-enclaves, and the parcel near Ginhoven that sits in Belgium proper and so is a plain enclave, which is the distinction most published Baarle counts miss. Five abolished cases are carried marked with the year they ended, Dahala Khagrabari among them: the history is the thing the layer exists to teach, and showing it as live geography would not be true.
- Military bases, 392 of them, where the useful question is not what kind of installation it is but whose flag is on the ground: filled marks are the 171 with a foreign or allied force, hollow are national bases at home. Nothing in it is derived from imagery, fitness data or flight tracking, and sites whose location is not already public are omitted rather than inferred.
- Volcanoes (300 in all: 297 from the Smithsonian’s Holocene list, plus Yellowstone, Long Valley and Toba, dormant supervolcano calderas that last erupted tens of thousands of years before the Holocene and are flagged as such) and recent earthquakes (a USGS 30-day snapshot), both dated rather than live, because the atlas makes no third-party request at runtime.
- The lines the tiles could not carry: Somaliland’s boundary with Somalia, the Cyprus green line and the UN buffer zone, and the two UK Sovereign Base Areas, which are real territory that the base map simply omits. The Somaliland line is dashed where it is a de facto boundary and dotted across the Sool sector, which SSC-Khaatumo has held since 2023.
The v4 pass was about what the map ASSERTS when you click it.
- What each admin-0 unit actually is. The countries layer paints 258 polygons and 197 of them are sovereign states; the rest are 33 dependencies, 10 self-governing parts of another state, 12 units of indeterminate status, 3 disputed territories, 2 leases and Kosovo. All 258 used to hover, label and open a sheet headed “Country dossier” alike, which called Baikonur and Scarborough Reef countries. They are now separated by Natural Earth’s own TYPE and SOV_A3, read by two rules the data forced. Parentage comes from SOV_A3 and never from a TYPE filter, because Natural Earth types Israel “Disputed” (its borders are unsettled, not its statehood) and Cuba and Kazakhstan “Sovereignty” (they hold sovereignty over leased ground). And a unit typed Disputed or Indeterminate gets a status word and never a named parent, because Natural Earth gives Palestine the sovereign code IS1: printing “Dependency of Israel” would assert the one thing this atlas exists to refuse. Labels follow the split, a state at full size and everything else a size down.
- Antarctic claims. Seven states have claimed sectors of Antarctica and Article IV of the Antarctic Treaty holds every one of them in abeyance: nothing is recognised, nothing is renounced, and no new claim may be made. The layer draws what was asserted, computed from the meridians each claimant’s own instrument names rather than traced from anyone’s map, with the three wedges where the Argentine, Chilean and British claims cover the same Peninsula ground drawn hatched as their own geometry, Marie Byrd Land labelled as the sector nobody claims, and Article IV printed on every sheet. The apex stops at 89.9° S because Web Mercator sends the pole to infinity, and the projection inflates these more than anything else on the map, which is the one place the globe button is not optional.
- Indigenous policy in the dossier. Click a country and, for 39 of them so far, the sheet carries a collapsed block: how the state recognises Indigenous peoples and under which instrument, the land and resource regime, the self-government arrangements by name, the ILO 169 ratification, the 2007 UNDRIP vote with any later endorsement, and the pages actually read. A country without a record shows no block at all.
- One toggle for autonomous and indigenous areas. Canada’s federal register of Aboriginal lands (3,364 parcels) rode a switch of its own beside the autonomous areas layer, which asked a reader to know that the same question had two answers in two places. They are one switch now, one key, one click behaviour, and the reserves draw in the same green every other indigenous territory on the layer uses.
- The 1964 layers, made readable. The language-family and ethnic-group maps now build their key from the colour map inside their own data file, so the swatch in the panel is the colour on the map by construction: the twelve largest families by drawn area, then a count of the tail, because 39 rows would push every other key off the screen. Hovering names the people and its family without a click, the largest areas carry a label, and the fills came down from 0.38 to 0.3 and 0.26 so the hillshade underneath still reads.
- Notes that do not promise. Fifteen click sheets used to end in a bracketed line saying a note was coming. The note section is off unless a build turns it on, so a sheet with nothing to say now says nothing.
The border policy also sharpened, and this is now the only place it is spelled out: the map itself stays a map, and the conventions live here and in the sources & methods. Ukraine is drawn at its internationally recognized 2014 borders, with the line of contact as a single dotted gray line, taken from a dated DeepStateMap snapshot (July 2026) and deliberately undifferentiated (no legal distinction drawn along its length); Kosovo–Serbia is drawn as a disputed dashed line; Western Sahara stays indeterminate, carrying both its de jure Morocco boundary and the de facto Moroccan berm as dotted lines. De facto boundaries are solid; claims and disputes are dashed.
On disputed borders
Any political map is a set of claims, so the atlas states its conventions up front. De facto control is drawn solid; claims are drawn dashed, the standard cartographic convention, applied consistently rather than per-controversy. If the atlas ever carries lines-of-control layers for active conflicts, they will be dated snapshots with cited sources, never a live assertion. A line labeled “as of” a date is documentation; the same line undated is an editorial I didn’t mean to write.
Where the map itself is
The map is not embedded in this page. It wants the full width of the window, its own layer controls and its own keyboard, so it lives at /atlas/ and this page stays the case study standing behind it.
The roadmap
The ledger of what the atlas cannot do yet, kept here so nothing silently disappears. Two items graduated off it in the third build-out: relief was regenerated from the 30-arc-second grid to zoom 7, so individual Alpine ridges resolve; and sea-depth shading now comes from real ETOPO bathymetry rather than a name hash. (A third, a hand-rolled projection inset that redrew the world in Robinson and Gall-Peters, was later retired; see the note below.)
A later build-out cleared more: native names for cities (now 576, and read at runtime rather than baked into the tiles, so adding one no longer waits for a retile); subnational religion so Nigeria reads as a Muslim north and a Christian south rather than a flat “mixed”, now across sixty-one countries where one national label misleads; pipelines (since folded into the Oil, gas & pipelines toggle, one switch for the whole hydrocarbon story) and data centers as toggles; a curated post-1964 addendum marking the largest shifts the 1964 ethnic map cannot show (Cyprus, Bosnia, Kosovo, Karabakh, the Baltics, Xinjiang); deeper ski and base coverage; and a curated subdivision-capitals table that lifts the hover from roughly half to seventy percent. The deep-time basemap also moved from a copyleft source to Cliopatria / Seshat (CC BY 4.0), and the map font gained CJK, Hangul, and Canadian-syllabics glyphs, so 中国, 日本, and ᓄᓇᕗᑦ finally render in their own scripts. What remains:
- Even higher-resolution relief. Land relief is now a live hillshade tinted from Mapzen /
AWS Open Data Terrarium terrain tiles, which replaced the self-hosted ETOPO bake: that read
worse in both themes and went blank past its zoom cap. Going deeper is now a matter of the
Terrarium overzoom and the exaggeration curve, not a rebuild and no longer bound by a
per-file size cap; the baked pipeline (
scripts/build-relief.py) still renders the sea-floor bathymetry the Seas layer draws. - Mercator and globe only, by design. A hand-rolled inset once redrew the world in Robinson and Gall-Peters beside the live map, but a static canvas next to a reprojectable globe read as clutter rather than clarity, so it was retired. A genuinely reprojected main map would need a different rendering engine (a D3-geo canvas layer, or a future MapLibre that supports it); until then the map offers Mercator and the true-proportion globe, and the projection trade-offs live here in prose.
- A comprehensive modern ethnic map. The shifts-since-1964 layer is a curated addendum, not a replacement for the 1964 snapshot; the full modern equivalents of the ANM polygons are commercial, and Wikidata-derived areas remain sparse. What the addendum does have is a stated rule for what belongs on it, which is the difference between twenty-five cases and an arbitrary eleven: post-1964, documented with a count, large enough to change which group predominates over an area a world map can draw, and settled enough in direction to state as fact. The exclusions are written down too, so an omission is visible rather than silent.
- The last unhosted scripts. Bengali, Sinhala, Lao, Khmer and Thaana names stay romanized: the glyphs are unhosted, and the map engine does not reorder those scripts correctly even when they are. Sixteen cities are affected and each keeps its real name in the data, waiting. Where only the local script is blocked and the national one is not, the country’s language fills in instead of the romanization, which is why Chennai reads चेन्नई and not சென்னை. Japanese and Traditional endonyms borrow the Simplified stroke styling.
- Country dossiers: the sheet answers a click with sourced facts for every unit, and the one thing still to come is Greg’s own note on a place. That section is hidden rather than stubbed until the notes are written, so the map promises nothing it is not carrying.
- Surveyed, not sketched. Oil and gas fields, bases, and the post-1964 outlines are approximate regions curated by knowledge; a surveyed-boundary pass is a future project. The pipelines came off this list: they are now 1,510 routes traced by Global Energy Monitor (CC BY 4.0), from the Global Gas and Global Oil Infrastructure Trackers, in place of the 25 corridors this atlas drew by hand. They had been ruled out on a licence read off a page footer rather than off the licence itself, which is the mistake worth keeping on the record here.
Retrospective
This page will outlive many versions of the map, so the status line is worth keeping current: the atlas is well past its foundation. Sixty-odd layers are live, every one of them cited on sources and methods, and the work now is depth rather than breadth, which is what the v3 and v4 passes were: making a click answer with a sourced number, and making the map say what kind of thing it has just been asked about.
Three constraints are chosen rather than suffered. Curation is the feature and the
bottleneck: every hand-picked layer is a maintenance promise, and a stale infrastructure layer
is worse than none, which is why each one carries its own capture date and its own refresh
cadence in docs/data-tiers.md. The disputes policy gets tested by real cases rather than
hypotheticals, and holding a convention under pressure is harder than declaring one: the
Palestine sovereign-code case in the v4 list above is exactly that, a place where the
structure of the source data would have had the map assert something the policy forbids.
And scope is permanent temptation (street-level zoom, live data, one more layer), which the
PMTiles architecture happily permits and the calendar does not.