What changed

Every release, what it added and what it fixed.

Site 3.12.1 16 October 2026

Pages repair themselves. A plugin built for another site can rewrite pages that share an address with these (pricing, about, privacy) so that their body is its own marker, which shows as raw text once that plugin is gone. Every page is now drawn from the site plugin whatever its body says, and on updating, and twice a day while an administrator is signed in, every page is checked and put back: its body, name, address and status restored, a missing one built again, a page template another plugin set removed, pages that are nothing but another plugin's marker moved to the trash, and the front page pointed back at the home page if it was pointed at one of them. The settings screen says what was done and shows each page's state.

App 3.25.0 27 September 2026

Against the record. The Reliability tab starts the model from a past year with that year's World Bank output, population, inflation and unemployment, runs it forward with no lever played, and scores growth, inflation and unemployment against what the World Bank recorded: the share inside the 80% band, the typical miss against carrying the start year forward, the lean, by years ahead, the world year by year and the largest misses, and all of it again without 2020 and 2021. What flatters it (today's structure) is said beside every score. The site fetches the history one series per request and keeps it thirty days, and the analyst cites the result when asked how sure it is.
The stress test. The Company tab builds the shocks the company's own files are most exposed to, runs each alone against a world with nothing changed on one future with no luck in it, prices the company line by line in every year and ranks them by the worst year, with the supply each one puts at risk beside the money because the model flags going without and does not price it. Any shock runs in full as the scenario with one press, and the report carries the table.
What would change the answer. Asking the analyst what would change an answer most, or to stress test a country or the company, runs shocks built from the company's files or the country's own trading ties and ranks them.
Revenue by segment, with each segment's range across the futures, in the Company tab and the report.
Watches. A kept answer watches the countries it turns on. When the feeds refresh, a party that has turned high or critical on the live desk, fifteen points above where it stood, flags the answer at the top of the Analyst tab and puts a count on the tab; the watch can be switched off per answer.
One world for the team. Any answer can be shared with everyone on the site allowed to save and taken down by its author or an administrator; an administrator can publish the company files for every member to use in place of their own; saved scenarios have a Share switch, and the team's appear under their own heading with who shared them.
Corrected: a lever set to start now acted from the second year the run plays, so a one year measure did nothing at all and a two year one lasted one. It now starts in the first year played. And a tariff, a sanction or an export control now moves trade in that first year: the first year had been read against the seed figures in a way that discarded any change in trade.
Corrected: the company summary with no scenario on screen now says the run is the baseline instead of showing a table of zeros, and the analyst says the same; line charts leave a gap where a figure is missing instead of drawing it as zero.

Site 3.12.0 27 September 2026

The method page describes the backtest and answers why there is still no single accuracy figure; the analyst, data room and organisations pages describe the stress test, watches and sharing with the team; privacy lists what the team can read and what a watch keeps; the roadmap moves the backtest and team sharing to shipped and names what the backtest still lacks.

App 3.24.0 27 September 2026

The company across every future. Each future of the scenario is set against the same future of the baseline with its own output, prices, exchange rates and the levers of the way it played, so the Company tab shows the spread of the net effect, the share of futures that lose, the share that lose more than a twentieth of revenue, the worst and best, and the company's result in each way a sentence could go. What is kept at the median is named.
Your countries today: every country in the company's files read against the Live desk, ranked by how much is happening there times how much of the company is there, and every supplier's likely sea route against the live state of its chokepoints.
Options to examine: for every flagged or priced up supply line, the countries with the capacity for that input, ranked by stability in the run, how little of the scenario reaches them and bloc; and the larger markets where demand holds up best.
Signs to watch. Every way a sentence was weighed into lists the signs in the world that would show it happening (a party joining, talks, an escalation, a delay), read from how its levers differ from the sentence as said, with what the Live desk reads for the parties today. On the Futures tab and in the analyst.
The analyst reads today. A question about what is happening now, in a place, in the company's countries or in the world, reads the Live desk's pressure, the headlines naming the place, the chokepoints, and the model's structural view beside them.
Reports. An analysis and the company reading each open as a document of their own, laid out for reading, print or PDF, with the run and the files they rest on at the head.
Kept analyses. Any answer can be kept whole, in the account when signed in (fifty at most and a megabyte in all, rebuilt on the server from the known shape) or in the browser otherwise, reopened with the figures of its own day, reported, or deleted; and handed on as a link that asks the same question again on the reader's own run.
Corrected: the company summary no longer prints a currency line of zero when no borrowing or cash was uploaded, and the spread chart always shows the baseline, so a run where every future loses no longer draws the baseline label in the middle of the losses.

Site 3.11.0 27 September 2026

The analyst, data room and organisations pages describe what 3.24.0 adds: the company across every future, today's reading of its countries, options to examine, signs to watch, reports, kept analyses and links that ask again.
The install notes name the current packages.

Site 3.10.1 25 September 2026

A site icon: the app's own mark, an orbit open to the east with the world inside it and the horizon through both, drawn for a browser tab, a home screen and the WordPress Site Icon. The tab icon the site draws for itself now matches it, and stands aside when a Site Icon is uploaded, so a page never carries two.
The mark in the header and the footer is that drawing, alive: the orbit turns, the world breathes and a point of light runs along the horizon.
Every block moves with the scroll. The hero falls away into the world as you leave it. Plates rise up out of the page, tilted back until they are a third of the way up, while the words beside them drift the other way. Headings slide in from their section mark as their rule draws. Cards rise as a wave, column by column. Bands open from a card to the full width, and the ink band behind the manifesto opens like an iris while the manifesto lights a word at a time as you read down it.
The gallery pins on a wide screen and runs sideways while you scroll down, leaning into the speed of the scroll and settling when it stops, with a rule under it showing how far through you are.
Every table, list and ledger arrives a row at a time; a question opens by growing to its height where the browser can; the glyphs on the cards keep moving (dashed lines march, points pulse) and draw themselves again under the pointer; a milestone still to come glows.
Everything above stands still under reduced motion, and the page can no longer be widened sideways by anything moving in it.

App 3.23.1 25 September 2026

The mark in the header is the site icon, alive: the orbit turns, the world inside it breathes, and it spins up under the pointer.

App 3.23.0 24 September 2026

The Analyst tab. Ask anything in ordinary words, in as many parts as the question needs, and the answer comes back with its plan, every figure it used as a numbered evidence row with its kind and where inside the instrument it came from, conclusions that cite their rows and say how sure they are, what would change the answer, and what would make it surer. A what if runs against the baseline on the same seed and can be put back; the evidence downloads as CSV and the whole analysis as JSON.
Written answers are checked on the server sentence by sentence: a sentence must cite a row, every row it cites must exist, and every figure in it must be in those rows. What fails is removed and listed with the reason.
The Brain tab. A knowledge graph of sources, figures, countries, blocs, treaty groups, levers, channels, outcomes, events, live feeds and your own files, with two neural networks trained on it in the browser, one on the world's figures and one on the baseline run. Hidden units are nodes and weights are ties; every output is scored on countries the network did not see; explanations split a prediction exactly into what each figure adds, each with its source; every lever is measured on the engine one at a time.
The Company tab. A data room of seven kinds of file, each column named and explained, every row checked with its reason when refused, kept against the account or in the browser. Every run then says what it does to the company line by line, with the chain from your figure to the run's figure written out, and a ranked list says what to upload next and where to get it. A price more than doubled is costed at a doubling and flagged as rationing rather than read as a bill.
Administrators can put world figures in from the Company tab, with the accepted range and a public source named for each field.
Each way a sentence could go is set against the same futures of the baseline, which drew the same luck, rather than against the baseline's median of all of them. The Futures tab, the briefing and the analyst print the same figure for the same way.
The engine carries currency and stability as bands for every country, so the company's exposure compares the scenario and the baseline on the median of every future rather than on two representative paths with different luck.
Every control in every section has its label tied to it, so a screen reader names it and pressing the label moves to it.
The tabs sit in four families, with Intelligence between the world and the run, and all sixteen fit the strip at 1600 pixels. The instrument can no longer be scrolled sideways by focus landing on a panel parked off screen.

Site 3.10.0 24 September 2026

Four new pages: the analyst, the brain, the data room and one for organisations, and the menu now leads with them.
Scrolling stories: a plate held in view while its chapters scroll past, each chapter telling the plate what to show. Under reduced motion the plate is redrawn still at each chapter.
Three new plates, drawn live: the brain learning and a country running through it, the analyst working a question through to checked prose, and the data room turning a file into a change.
Pages turn into one another where the browser can, with the header held still, and fall back to the fade elsewhere. Small labels resolve out of noise once, a needle under the menu follows the pointer home, plates are uncovered as they arrive, and cards rise toward the pointer.
Corrected: the World Bank refresh fills seventeen fields, not sixteen, and the thirteen others survive it.

App 3.22.0 23 September 2026

Each clause of a sentence is weighed on its own. "China invades Taiwan and the US sanctions China" is a war with its own ways (limited, as said, widened, frozen, a long war) and a sanctions package with its own (holds, answered, evaded, lifted), and the two vary independently across the futures.
The levers in the panel are the sentence as said, exactly as the first future plays it: a war with no length given now shows its five years there instead of a lever that never ends. A Western partner outside NATO that is attacked (Taiwan, Japan, Korea, Australia) can be joined by the United States in the widened way.
Shows of force are their own kind: a missile test, an exercise, a diplomatic break or a bomb programme is weighed as defused, as said, a clash or sanctions, not as years of fighting.
A war with no attacker named, aimed at the focus economy, is laid by the leading power of the bloc most at odds with it rather than by every country on earth.
A civil war is fought inside one country. Fighting aimed at nobody in particular used to be read as fighting everybody, so "civil war in Sudan" put every country at war and cut the world's energy and chip output by a fifth or more; it now weighs on the country that has it.
Military spending from an event is a rise on what a country already spends. The engine holds it as a share of GDP and the events wrote their rises as levels, so "a little more spent on security" set most countries below today and peace set a country's spending to a fixed low figure. "Raises defence spending by two points" is now a change and "to three percent" a level, and the weighing scales a level by its distance from today, so half a rise is still a rise.
The weighing no longer reads the live feeds when a run does not start from them, so the same scenario on the same seed splits its futures the same way whatever the feeds did in between.
Tried levers reach a weighed run: What would help and the step's probes add their candidates to every future, and a probe plays each sentence as said. A fifth sentence runs as said instead of being dropped. A way that no future drew in a small ensemble reads as missing, not as a total loss, and says to raise the number of futures to see it.
The chain and the step's lever list describe the way the representative future actually played, and the chain names it. A kept run in Compare and a saved scenario carry the sentences they came from, whether they were weighed, their seed and their focus, so loading them back runs the same thing; an empty set of sentences saved on the server no longer comes back as a list.
The weighing switch in Expert is honoured by the ask box itself: off, the sentence runs as read, with no call to the model. A run whose settings cannot be built says so instead of leaving the app waiting for a run that never started.
The model's weighing: shares are normalised on the server, so an answer in percent or all zeros still splits the futures; blocs named in any case (Mercosur, the Arab League, OPEC+) map to the names the engine knows; fighting inside one country is kept; more room is given for the answer.
Retaliation draws are independent across years and countries; they used to share a stream that repeated for some pairs.

Site 3.9.2 23 September 2026

The pages that describe the ask box say what it does now: it reads the sentence and weighs the ways it could go before it runs, shows nothing first, and lays out every way afterwards under Futures.

App 3.21.0 23 September 2026

The ask box decides inside. Nothing is laid out while you type and nothing is asked back: press Simulate and the sentence is read, weighed and run. The reading panel is gone.
Every sentence is weighed into the ways it could actually go before it runs, each a full set of levers with a probability. A war can stay limited, run as said for several years, widen as allies come in and the aggressor is walled off, freeze into a ceasefire with the tariffs and sanctions left up, outlast the run, or, between two nuclear powers, cross the threshold. A tariff can hold, be answered in kind, be bargained away within two years or spiral into a trade war; sanctions can hold, be evaded, be answered or be lifted; export controls can hold, leak, be answered with controls on minerals or be joined by allies. Supply cuts, disasters, outbreaks, financial shocks, new technology and domestic policy each have their own ways.
The weights move with who the parties are (rival blocs, nuclear arms, oil exports, alliances), with the pressure the live feeds show on them, and with the sentence's own words: limited weighs the held back ways up, all out weighs the wider ones up, and a length or size the sentence states is kept. A war with no length stated runs about five years as said, with a long war weighed as its own way, instead of fifteen years at full weight.
The engine splits the futures between the ways in exactly those shares and every future plays its way to the end of the run with its own luck, so the fan and every tab are the ways mixed by how likely they are. Two sentences are weighed independently of each other. A lever edited by hand is weighed again from its new value; Expert has a switch to run the levers exactly as set.
Futures opens with The ways it could go: each way, its probability, why it was weighed so, and how it ended against the baseline for the world, the focus economy and every country it names, as cards, a table or lines over time, with where the ways land together. The chain says which way its representative future played, the briefing names the harshest and the mildest way, and Reliability lists the weighing.
With a model key the model does the weighing: three to six ways with probabilities and reasons, read against the current world, checked against the lever schema on the server. If it is slow or answers badly, the browser's own weighing is used. The model may now name a bloc or everyone as the party that acts.
The bottom of the app is no longer cut off. It fills exactly the window down to about 420 pixels tall (it would not shrink below 640, so a short laptop window lost the foot and the timeline), follows a phone's collapsing address bar, and tightens its header and foot under 720 pixels. In a short map the legend folds its notes so it never runs into the layers panel.
The night note in the legend gives the day the sun is drawn for, not only the year.

Site 3.9.1 23 September 2026

The app page no longer loses its bottom edge for a signed in owner: the admin bar is not loaded there, and its 32 pixel bump, which pushed the foot and the timeline below a window that does not scroll, is cancelled. A window too short to hold the app now scrolls instead of hiding the rest.

App 3.20.0 23 September 2026

The map on the Live desk can always be zoomed out and reset. The desk rows are a fixed height, so the map panel no longer grows past the screen and pushes its controls out of sight; zoom in, zoom out and Reset sit in a column on the right edge of the map, vertically centred, at every width. The keyboard works on the map once it has focus (plus and minus zoom, 0 resets, the arrows pan the flat map and turn the globe), a shift double click zooms out, and the globe bar has its own Reset.
The panel picker is rebuilt: thirty panels in six families (Overview; Conflict and security; Markets and trade; Hazards and the planet; People and health; Air, orbit and eyes), each a tile with its name, what it shows, its size on the desk, whether its sources are live and a switch. A row of lenses edits any lens without leaving the picker, with the count on each; a find box narrows the tiles by name, note or source; each family shows how many are on with Show all or Hide all; Everything, Nothing and The lens as it came reset the lens in one press; the arrows on a live tile move it earlier or later on the desk, with its place shown; Done closes it. The Panels button on the desk shows how many are on.
Every section in every tab is laid out the same way: the header groups its tabs into three families (Now, Run, Trust) with a rule between them; the reliability badge moved from the header into the strip under it, beside the run figures; figure cards read label, figure, note in three weights; each section has a marked head, its controls on one tinted strip aligned to one height, and an empty state in a dashed frame; the views that are read for one country (Changed, Chain, Causes) carry a context bar with the country and measure; tables align figures to the right with their headings, keep row headers plain, stripe their rows and mark rises and falls.
No placeholder text anywhere: every search box and settings field has a visible label or a description instead, in the app and in the settings screens.
The app is proof against the theme it is shown in: buttons, labels, checkboxes, tables and headings inside the app are reset to its own look, so a theme that styles every heading large or every checkbox centred no longer reaches in.
The edge of night no longer draws a straight line along the top of the flat map near an equinox, where the curve crosses a pole; the compact map legend says where the sun stands and on what date.
The ask box reads who does what to whom. A place named after on, against, from, to or with, or in front of goods (Chinese cars, US steel), is the one acted on, whatever order the sentence comes in. "A 25% tariff on Chinese imports" read as China taxing everybody; it is now the focus economy (the United States unless you chose another) taxing China. "China puts a tariff on US imports" is China on the United States, "Europe taxes Chinese cars at 40%" is the EU together on China at 40, "sanctions on Russia" and "an embargo on Russian oil" are aimed at Russia, and "an attack on Taiwan by China" is China on Taiwan. US and UK in capitals are found when a word follows them, Europe is the union, and German, Thai, Kazakh and Filipino are read as their countries.
Words are matched as words: "expo" is no longer found inside "exports" (which read every chip export ban as a building programme), nor "emp" inside "employment" or "asi" inside "Asia". Where two readings claim the same words the longer phrase wins, so "cyber attacks on the UK" is a cyber attack and not a war. A lever between two parties said outright ("bans chip sales to China") wins over a loose event that shares a word with it, and "bans" or "restricts" no longer halve the export control they describe.
The West, the other alignment blocs and everyone can lay a tariff, a sanction or an export control. They used to resolve to nobody, so "the West sanctions Russia" ran as nothing at all. A tariff laid by a bloc can now be answered in kind by the country it hits, in one line in the events. Lever cards and readings name the party in words (the West, everyone) rather than a code, and Applied by offers everyone, the blocs and the groups.
The last long dashes are gone: an empty figure in a table reads n/a, the rebuild status in the settings uses a dot, and a figure that was never recorded reads none.
The Chain says so when nothing has reached the chosen country yet. Trade reaches the country taxed a year later, so the first year now reads as not yet moved, with a pointer to step on, instead of "through 0 channels".

Site 3.9.0 23 September 2026

The app page loads the app and nothing else: the theme's stylesheets, the block library, global styles, custom CSS and the emoji script are dropped on that page only, so the theme cannot restyle the simulator. The admin bar stays.
The settings screens describe each field instead of showing placeholder text.
The plate behind the pages no longer draws a straight line at the top when the edge of night crosses a pole.

App 3.19.1 23 September 2026

No Japanese lettering anywhere: the vertical line of type beside the rail is gone, the accent colours are named in English only, and the notes in the stylesheet say the same things in one script.

Site 3.8.1 23 September 2026

No Japanese lettering anywhere: the large faint characters behind the sections are gone, with their font.

App 3.19.0 23 September 2026

The slow tier reaches the browser whole. The payload the app reads copied a fixed list of thirteen sources from the server, written three releases ago, so every server side source added since (the headlines, the cyber catalogue, the volcanoes, the outbreaks) was fetched and never handed over: the desk said "waiting for the server" for ever. The payload now carries every source in the registry.
Three standing registers, read whole by the server: the UN Security Council's consolidated sanctions list (every individual and entity, the regime, the date, the country each is tied to), the UNHCR's population statistics (refugees and asylum seekers by the country they fled and the country hosting them, and the internally displaced), and the Uppsala Conflict Data Program's candidate monthly release of organised violence (every event with at least one death, placed and dated, with the sides and the best estimate of the dead). Each is a panel on the desk, a layer on the map (rings sized by listings; two discs and a ring for the displaced; dots sized by the dead and coloured by the kind of violence), a term in the pressure index for the dead, an alert for the deadliest events, a line in every dossier under On the registers, and six base layers under Live registers.
Forty headline feeds in eleven desks: Diplomacy joins with the UN press office, the US State Department, ReliefWeb, the Council on Foreign Relations and Carnegie; Defense One and Bellingcat join the security desk.
The weather stations sit at the capitals, not at the countries' centroids: Moscow is read over Moscow. The largest cities are read as well, deduplicated against the real capitals, and a reading the budget did not reach keeps its last value for half a day, marked with when it was taken.
The weather field is computed once per set of readings as a picture of the whole planet, bucketed by station, and painted on either projection by lookup, so the globe keeps turning with it on. Country names are drawn light on photography and dark grounds and dark on light ones, whatever the theme, and they draw before a run as well as after one.
The Night switch cycles: shade with the curve, the curve alone (with civil and nautical twilight dashed beside it), off. The month step puts the sun on the first of the month the label names.
The country reader takes the longest name at a place, so South Sudan is not read as Sudan and Guinea-Bissau is not read as Guinea, and it tries every occurrence, so Mali inside "normalisation" no longer hides the Mali later in the headline. The UN's formal names (Democratic People's Republic of Korea, Syrian Arab Republic, Russian Federation and the rest) resolve to the model's codes.
A source whose fetch the host killed is left alone for ten minutes rather than asked again by the next visitor; the UCDP fetch tries the newest release names first and keeps the one that answered with events; registers and views of them carry the right legend note; the site inspector prints outlooks, wind direction and rain for a weather station and words rather than a number for the kind of violence.

Site 3.8.0 23 September 2026

The horizon: every page head is drawn on. A strip of the ruled world with the edge of night crossing it through the seasons, the sun marked, and a few things passing over in orbit, faint, behind every title. The country outlines ride on every page for it.
The sources page lists the three registers and the eleven desks; the site's temperature ramp is the app's exactly; a browser without IntersectionObserver still draws every plate once under reduced motion.

App 3.18.0 23 September 2026

Country names are off by default and drawn as an atlas when they are on: ranked by size, tracked, haloed against the ground so they read over any colour, the small countries falling back to their codes, a budget of names that grows as you zoom, and on the globe faded toward the limb rather than cut off. No ground tile carries a name into the picture any more.
The edge of night is the curve it is. The terminator is walked as the circle round the sub-solar point, so it is smooth at the poles and at the equinox; it is drawn as a line on the flat map and the globe, with civil twilight as a fainter dashed one beside it; and the globe's shading is computed the same way as the flat map's, so the shade and the line always agree. Both follow the clock of the run: stepping by month, week or day moves the season, so December bows the curve to the solstice while the hour of the day stays the real one. The legend says where the sun stands and on what date.
Weather you can see. A Weather field overlay paints temperature across the map from every station, blue through green and yellow to red with the ramp in the legend, the wind as an arrow at each station scaled by speed, and rain as rings; it draws on the flat map and the globe. On the desk the weather panel opens with a strip of the whole planet from coldest to hottest with the mean, and every row carries a three day range bar with a rain mark.
The weather reads the largest cities as well as every capital. Two hundred and forty of them by default, from the cities pack the map already ships, skipping any within forty kilometres of a capital, so the field has a reading every few hundred kilometres where people live. The count is a setting; zero reads capitals only.

Site 3.7.0 23 September 2026

Every page is drawn on. Twelve plates, each a small instrument drawn live from the ideas the simulator is built on rather than a screenshot or a stock image: the ensemble fanning out from one world, the five layers of a place, flows between real nodes, the edge of night through a year, six panels of the live desk, the network with the geography taken away, the ruler of the run opening to finer steps, a number opened into its terms, the zoom from country to place, the weather field, the sites that open, and the atlas globe. Each leans toward the pointer, runs only while it is on screen, is drawn once and left alone under reduced motion, and is drawn in the colours of the theme it sits on.
The home page rebuilt around them: five plates between the sections, a gallery of all twelve that opens the page for each, the features as tiles with hand drawn line glyphs that draw themselves on arrival, a band of ink across the page for the manifesto with the edge of night sweeping behind it, one large faint character behind each section as its mark, and a rail beside the steps that fills as you read. The copy now describes the live desk, the sites, the weather field and the edge of night, and the stats count the sources and the cameras.
Plates on the reference pages too: the ensemble on How it works, the decomposition on Method, the layers, the edge of night and the sites on The map, the desk on Sources, the network on The network, the flows on Flows, the weather field and the atlas on The layers, the step on The model.

App 3.17.0 23 September 2026

The boot object reached the wrong file. WordPress printed the app's settings (the asset path, the REST root, the nonce, the keys, the default ground) just before the last script rather than the first, so on a real site every earlier file started with an empty object: the geometry packs asked for a relative address and got the page back, the server road had no nonce and was refused, the keys never arrived, and the states, counties, cities, towns and every infrastructure layer could not switch on. The harness had always printed it first, which is why none of this showed there. It is printed before the first file now, and the app adopts it late as well, so a cached page heals itself.
Everything on the map opens. Every point the map draws is a thing you can rest on and click: ports, airports, fabs, LNG terminals, datacenter campuses, exchanges, refineries, mines, power stations, research clusters, spaceports, military bases, cable landings, pipelines, chokepoints, satellites, military and civil aircraft, cyclones, volcanoes, weather warnings, outbreaks, launches, protests, alerts, outages, weather and air readings, cities and towns. The inspector shows every figure the site carries, what is around it (the nearest city, cameras within reach, every live event near it, the weather over it now), its country's standing in the run at the timeline's moment, and the one push the model has for it, added to the scenario as levers. The search key finds every site by name.
Flows land where they land: goods at the country's largest container port, chips at its fab cluster, compute at its largest datacenter campus, energy at its LNG terminal or refinery, capital between exchanges, data at a cable landing, people through the busiest airport, arms at a base, research between research clusters, and the political flows between capitals, with the centroid only where a country has none of these. Capitals are placed from the city list rather than the centroid.
Night on both projections. The half of the world where the sun is down, computed from the clock with the twilight eighteen degrees wide, shades the flat map as well as the globe, and the globe's shadow follows the real sun rather than a slow fake one. The switch sits on the View bar.
The default ground is none, the drawn world. A site that never chose a ground is moved to it on upgrade; a site that chose one keeps it. And the globe now draws its skin from the label free twin of every ground: the sheet builder had been reading the labelled provider, so names were still being baked into the sphere.
Six more live sources on the server, four of them with a browser road: tropical cyclones from NOAA's National Hurricane Center with intensity, pressure and movement; the Smithsonian's weekly volcanic activity report; the next launches and the last few from Launch Library, at their pads; extreme and severe weather warnings from the US National Weather Service and Europe's MeteoAlarm; the vulnerabilities CISA says are being exploited in the wild; and the WHO's disease outbreak news placed at the country each names. Each is a panel on the Live desk, a layer on the map (cyclones as spirals with their heading), a term in the pressure index where it bears on a country, and a line in the dossier's timeline.
Weather over every capital the model carries, with a three day outlook, on both roads, and a weather panel on the desk that filters and orders them (largest economies, hottest, coldest, windiest, wettest outlook). A cyber desk joins the headlines.
ReliefWeb moved to its second API and now requires a pre-approved application name; the plugin uses the new address, takes the name under Settings, Keys, and says plainly when the API refuses the request instead of printing a bare 410.
Two hundred and fourteen live cameras, a hundred and thirty of them video: a hundred and five long running streams added across traffic, city, port, airport, border, volcano, wildlife, ski, polar and orbit, each confirmed by the operator's own listing and checked again by the server the moment it is opened. A still camera shows a clock counting down to its next frame, so a picture its owner refreshes every minute reads as what it is rather than as a photograph.
The layers panel never rebuilds itself for a geometry fetch any more: the fetching, finished and failed notices are written in place under the strip, so a click, a double click or a slow host no longer sends the list back to the top, and a row never selects its text. The hover card is larger, solid and above everything.

Site 3.6.0 23 September 2026

The map page describes the site inspector, the anchored flows and night on both projections; the sources page lists the six new live sources and the weather over every capital.

App 3.16.0 23 September 2026

A Live desk. One tab with the world as it is this minute, arranged in panels around the same map the World tab draws: headlines by desk from thirty public newsroom feeds, the alerts a duty officer would want (red and orange disaster alerts, large earthquakes, a geomagnetic storm, a country gone dark, a sea lane in trouble), a pressure index, markets and rates, prediction markets, internet outages, the sea lanes scored, military aircraft by country, aircraft in the air, what is in orbit, earthquakes, space weather, weather, natural events, humanitarian situations, protest reporting, air quality, six live cameras, a watchlist, and the run from today. Eight lenses arrange the same panels for a reader with one question: World, Geopolitics, Security, Markets, Technology, Energy, Planet and Calm. Every panel can be hidden, opened large or read about, and the choice is kept in the browser.
A pressure index. Where the world's attention and trouble are right now, per country, from twelve live signals at once: coverage, conflict, protest and disaster reporting, headlines naming the country, disaster alerts, humanitarian situations, internet outages, military aircraft overhead, prediction markets about it, large earthquakes, open natural events, and the model's own structural instability as the last term. Every term is capped and named, so a country's dossier says why it is where it is. A reading of today, not a forecast.
A dossier for every country: today's pressure and the signals behind it, a seven day timeline of every dated thing in the feeds that names the country, the ensemble's median path for it from the timeline's moment (a week, a month, a quarter, a year, three years out) for tension, stability, growth, inflation and unemployment, its cameras and the markets about it, with a way onto the map, into the profile and onto the watchlist. Any country code on the desk, any headline's chips and the search key open one.
Three new live sources on the server's slow tier: headlines from thirty public newsroom feeds in nine desks, read for the countries they name (the settings screen can replace the list, one feed a line); the most traded open questions on Polymarket, with the price of yes as a probability and the day's volume; and where the internet has gone dark in the last day, by country, region and network, from IODA. The two that allow it are also fetched by the browser when the server has nothing. Internet outages are a map layer as well.
The sea lanes scored. Each of the twelve chokepoints carries a disruption score from what is happening within reach of it: conflict coverage, alerts, military aircraft, and traffic through it as a sign the lane is open. A disrupted lane is an alert on the desk.
The run from today, on the desk: how many priors the run took from the feeds, and for the countries under most pressure the run's median tension and stability now, a week, a month and a year out, beside their pressure score. A sentence box under it runs a different tomorrow from the same today.
Country names off the globe. Every ground that carries names in its tiles is swapped for its label free twin on the sphere, including the topographic, atlas and Esri street grounds and the keyed ones, so no text is ever baked into a tile and blurred. The app's own crisp names are the only names on the globe, a Names switch sits on the globe bar, and switching it off gives a globe with no names at all.
Rest on any layer row and a card says what the layer is: what it draws, its source, whether it waits for a closer zoom or a key, where its figures came by and how old they are, and whether it is on. Touch keeps the info button.
A hundred and nine live cameras instead of forty: sixty more public authority still cameras across North America, Europe, Asia, Africa, Oceania, the poles and South America (volcano observatories, NOAA and Antarctic stations, park services, road and transport agencies, river cameras, wildlife reserves), and ten long running city streams. A camera that runs a channel is no longer opened blind: the server asks the channel what it is streaming first, the player opens on that stream, and when the operator is between broadcasts the panel says so plainly instead of the player saying unavailable. Every camera has a Next camera button.
Two curated layers: a hundred and five military bases and stations coloured by the country that runs them, and twenty nine spaceports sized by launches a year. Both sit in the place inspector.
The ask box reads forty one more kinds of event and several hundred more phrases: an assassination, a regime change, a secession, an alliance, a bomb programme, a missile test, a diplomatic break, military exercises, a naval blockade, a cable cut, a solar storm, an oil spike, an oil glut, an oil find, a refinery hit, a clean energy boom, carbon removal, a compute boom, an AI bust, an AI incident, an AI pause, nationalisation, liberalisation, a tax change, wages, healthcare, schooling, family policy, a crime wave, a trade corridor, a shift away from the dollar, capital flight, a rate move, an inflation surge, deflation, a technology shock, longevity, a cold snap, an impact event, a genocide, a host year, a legalisation and press freedom. Every phrase now belongs to exactly one event, so a coup is a regime change rather than unrest and a tax cut is a lower rate rather than a higher one.
The command palette reaches the desk: every lens, every panel, a dossier and a watch entry for every country.

Site 3.5.0 23 September 2026

The map page describes the Live desk, the pressure index and the dossier; the sources page lists the three new live sources.

App 3.15.0 22 September 2026

The globe no longer blurs or flickers on a street ground. Text inside a tile is pixels, and pixels wrapped onto a sphere blur and shimmer as it turns, so on the globe a ground that carries its own names is drawn from its label free twin and the names are drawn on top, crisp at any angle. And a new sheet of imagery now opens with the last picture scaled into place rather than flat sea, so the swaps a spinning globe makes every few seconds no longer show holes filling in one tile at a time.
Thirty grounds instead of seven, in four groups: streets nine ways, terrain nine ways, the earth from orbit as it was yesterday (true colour from two NASA satellites, last night's lights, Blue Marble, sea temperature, snow, chlorophyll, all from NASA's own tile service with no key), and six that want a key. OpenStreetMap needs no key; the settings screen says what the foundation asks instead.
Over the ground: transparent layers over any ground, any number at once. The rain radar and the infrared cloud picture as of the latest frame, the fires a satellite saw yesterday, sea marks, railways, place names for a ground that has none, borders and roads, and live traffic flow and incidents with a TomTom key.
Keys: MapTiler, Thunderforest, Stadia, TomTom and NASA FIRMS, all optional, each locked to the site. A layer that wants a key the site does not hold is listed and says so.
The news, by theme and by place. Six themed questions to the same index that draws the news map, every half hour: conflict, protest, disaster, economy, energy, health. Summed inside each country they are seven base layers under In the news today; summed inside each state and province they are a disc per unit; inside a selected unit the inspector prints them. Coverage, not incidence, and the legend says so.
Every run can start from what the world is doing now. The news pressure by theme, the open disaster alerts and the UN's list of humanitarian situations become small, named priors on the starting state: a little more tension where conflict is in the news, a little less stability under protest coverage, a red alert or an open situation. Each fades by half a year, the baseline shares them so What changed is still the levers alone, and the Reliability tab lists every one with its size and its source. A switch on the settings screen sets the default and the Expert panel switches it for a session.
Aircraft in the air: every aircraft the OpenSky receivers can hear, the two thousand fastest of ten thousand, with heading, every fifteen minutes. Active fires as points with their radiative power, with a FIRMS key.
"China bombs USA" reads as an attack by China on the United States, which it did not: bombs, missiles, airstrikes, shelling and drone strikes were not in the attack vocabulary and the sentence fell through to a loose reading that ran as a small disturbance. The loose reading itself no longer lists levers to try or asks for a sharper sentence; it prints what it did and says a second sentence adds to it.
A scenario that opens as a shock lands inside the first weeks of its first year when stepped by day: prices, rates, the currency, tension and risk are placed within weeks and output, jobs and debt follow over the quarters. The engine solves years; the This step tab says what the shape inside the year is.
The model provider, when a key is configured, reads a digest of the world as the feeds last saw it before it turns a sentence into levers, sizes the levers to the sentence, never returns nothing, and answers with its reading in its own words: how it read the sentence, which levers it chose and why, and what in the current feeds bore on it. The ask box prints that before anything runs.

Site 3.4.0 22 September 2026

The map page describes the thirty grounds, what lies over them, and the keys.

App 3.14.0 22 September 2026

Live layers no longer wait on one slow server request. The server refreshes its sources inside a time budget and saves each one as it lands, so a slow feed cannot hold up a fast one. When the server has nothing for a source that allows it, the browser fetches that source directly. Every live layer in the panel now says where it stands: live, cached and how old, fetched here, or unreachable and why.
Cameras built in. A curated set of forty public live cameras, across cities, ports, airports, parks, volcanoes, wildlife and orbit, works without a key. The settings list and a Windy Webcams key add to it rather than replacing it. Video cameras open as a live stream; still cameras refresh at the rate their owner publishes.
Satellites always draw. If the server holds no orbital elements the browser fetches them itself, and the panel shows the epoch of the elements in use, so the age of what you are looking at is on the screen rather than assumed.
Cities can be checked. Click any city or town, or find it with the search key, for its population, its country, the local time, the current weather and air quality, and the country's figures at the year on the timeline, with a zoom to it.
States, provinces, counties and districts open an inspector: the area, the places inside it, the infrastructure that sits in it and the live events happening in it.
The layers panel gains a search box, an on now strip that lists what is switched on with one click to switch it off, and groups that collapse. The legend no longer sits over it: the panel is bounded by what the legend leaves, and on the globe the legend and the controls sat one of their own heights above the bottom, which is fixed.
The same figure now agrees with itself across the tabs. The monitor table reads every measure the run carries from the run at the year on the timeline, the same year the map and the inspector are on, so a country no longer has one inflation rate on the monitor and another in the panel beside the map; measures the run does not step stay at the base year and every header says which. The monitor's world output and population cards print the same figure the strip at the top does. The profile's base year blocks are titled with the year, and a row called growth that was population growth is called that.
Units put right in the measures catalogue: energy is exajoules a year, not terawatt hours; oil output is millions of barrels a day, not thousands; fab capacity is a share of world capacity; food self sufficiency is a ratio; the row that was called government effectiveness on a scale of a hundred is the government type on a scale of three, which is what the dataset holds.
The renewable share of electricity was a pattern rather than a figure, with Norway, Brazil and Canada at six percent. It is now a curated figure for a hundred and seventy one countries, and the World Bank refresh covers it, so it corrects itself further on any site that runs the refresh.
A shut strait now costs shipping. The ports lever could only add capacity, so a chokepoint closing was quietly read as more port capacity. It runs negative now: capacity lost makes every import dearer and takes a share of exports off the ships, and a Gulf closure is aimed at everybody's shipping and at Saudi and Emirati oil rather than at the focus economy.
The ask box reads twenty six more kinds of event: drought, blackout, a gas cutoff, a bank run, a default, a housing crash, a chip shortage, an arms race, a satellite shot down, an internet shutdown, brain drain, a baby boom, austerity, a stimulus, basic income, a wave of automation, a nuclear accident, sanctions lifted, an embargo, a water dispute, a megaproject, a petrol ban, an election shock, a corruption crackdown, a pandemic and a rare earths ban. Each clause of a sentence is read on its own, so an event in one half and a lever in the other both land; the same lever aimed at the same parties is one push rather than two; a place name is a whole word, so austerity no longer reads as the African Union and normalises no longer reads as Mali; and a three letter code counts only in capitals, so and is a word and AND is Andorra.
Map labels no longer print on top of each other, and at depth every city and town in the picture is drawn rather than the largest few hundred on earth.
The Sources tab ends with the live tier as it stands this minute, each source with the road its figures came by, how old they are and what went wrong, with an Ask again button on anything that is not answering, and a block on your own data that says what to upload and where.
CSV export on every section of every tab; a section that never declared its columns gets the rows it computed.
Settings: thirteen themes (Sakura, Yuki, Moss, Ume and Ocean are new), twelve signal colours (blossom, orange, lapis, spring green, wisteria and lead are new), a choice of corner, and solid panels over the map for an old machine.

Site 3.3.0 22 September 2026

A page on bringing your own data: the exact field list the upload accepts with the unit of each, the shape of the file, where it goes in, what the checker reports, what happens to a bad row, how a figure of yours is marked and how it reaches the model, and which fields to supply for which question. The field list is read from the simulator's own intake contract rather than kept as a second copy, so it cannot drift.
The hero has a line drawn globe with orbital rings that draws whether or not the simulator plugin is present, and a light that follows the pointer across it.
Motion throughout, all of it additive: headlines that set themselves a word at a time, sections that arrive with a stagger, figures that count up, a thin line at the top of the page that is the scroll position, a menu that condenses as the page moves, buttons that lean toward the pointer, hairlines that draw in as a section arrives, and a fade between pages. Every one of them stands down under reduced motion, and every page is complete with the motion file never loading.
Hover states with intent: underlines draw, cards lift by a hair, arrows travel.
Two new block types for the reference pages, a table and a definition list, so a field contract can be printed as the table it is rather than as prose.
The camera entry on the map page now matches the app: a curated set is built in, and the settings list and a Windy key add to it.
The figures in the ticker on the front page are read from the simulator's own dataset when the app plugin is installed beside this one, so the site cannot print a number the app would disagree with; the list written into the page is only the fallback for a site without the app.
Fixed: the plugin header said one version and the code said another. Both say 3.3.0.
Every page is rendered from the command line and opened at desktop and phone width on every build, checked for console errors, horizontal overflow and anything that failed to draw.

App 3.13.0 September 2026

Every block in the app can now be shown more than one way and narrowed at least one way. A block used to be one drawing of one slice of the data, chosen once by whoever wrote it, and the only answer to somebody who wanted it ranked instead of listed was to write another block. A block is now declared rather than drawn: what it is called, what can be asked of it, and the several ways it can be shown. Thirty six of them across the thirteen tabs, a hundred and twenty six views and ninety three filters, and what you choose is remembered for that block so a person who always wants the table does not have to say so twice.
The monitor reads four ways. The table it always was; a heat grid that shades every country against the middle of its own column, which answers "who is unlike the rest" in one look; a distribution that answers whether the field is one crowd or two; and cards, which is what a narrow window can actually show. The leaderboards gained a distribution and a scatter, so two measures can be put against each other rather than merely ranked, and it says how closely they move together.
A new section on the monitor asks how the world divides on any measure: by region, by alignment, by income tier or by sub region, added up as a total, an average, a weighted average or a count, with a concentration reading and a Gini figure beside it.
The country page gained four sections. Every measure on file against the field you choose, which can be the world, its region, its alignment, its income tier or countries its own size, with rank and percentile for each. Where it stands on one measure, with its nearest neighbours either side. Everything the median path carries, which is seventeen measures a year rather than the five the run publishes a band for. And its trade partners with a concentration index, which is the number that says whether an export control would bite.
The futures tab now reports the odds at any year rather than only at the horizon, tracks how each one changes as the run goes on, and lets the branches be read as cards, as a table, as lines against the ensemble median, or by what marks each one out. The events in the sample path can be searched and filtered by kind.
The why tab can be read at any year, sorted by size or by which terms help and which hurt, shown as a waterfall, a table, a split, or against every other country so a local cause can be told from a global one. The same terms across the run can be read as lines, as a heat grid, or as a table of what each one added.
Reliability, sources, the briefing and the working panel all gained search and filters. The source catalogue can be narrowed by domain, by kind and by whether it refreshes itself, and read by domain, as one table, by provider or by kind. The working panel can be read by section, as a table, by source file, or as every individual reading.
The world map gained a drawer of readouts underneath it, closed by default so the map keeps the room. The map is a picture and a picture cannot be sorted, so the numbers behind it now sit under it with the same views and filters as everywhere else.
Fixed: the first step of a run reported that nothing had happened. Year zero of a result is already a year after the base year, so comparing it with the step before it was comparing it with itself. The first step is now measured against the base year the world actually started from, which is the number a person expects to see at the moment they open the panel.
Fixed: a browser set to reduce motion still waited out the arrival stagger, up to a sixth of a second a block, which is the same staircase the setting asked not to see arriving in one jump instead of sliding.
The ask box no longer refuses a sentence. It used to read a sentence for a lever by name and, finding none, tell you to name one instead, which is the wrong answer to "China invades Taiwan": that is not a lever, and it is a thing this model can carry, because a lever is only ever a way of pushing on the state. Thirty kinds of event are now read as the set of pushes they are: invasion, civil war, a nuclear weapon used, a chokepoint closed, a cyberattack, an earthquake, a famine, a pandemic, AGI arriving, a fusion breakthrough, a moon base, a financial crisis, a currency crisis, a peace deal, a trade war, a four day week, and the rest. The reading says which pushes it used rather than implying the model knows how an invasion goes.
A sentence that names neither a lever nor a known event still runs. It is read for what kind of thing it is, mapped onto the nearest pushes the model holds, and marked as a loose reading with a note saying exactly what it did. Aliens landing in Nevada runs as lost output and financial stress, and says so. Fifty one sentences are checked on every build, and every one of them has to come back with levers the engine will accept.
Seventeen levers that the catalogue carried but no sentence could ever reach now have words: infrastructure, ports, education, research spending, the retirement age, nuclear build, strategic reserves, desalination, crop policy, chip stockpiles, bank capital, price controls, foreign aid, AI regulation, battery cost, global rates and financial stress.
Fixed: the ports and logistics lever never reached the engine at all. The catalogue calls it one thing and the engine was listening for another, so every scenario that used it ran as though it had not been set.
Fixed: a country lever aimed at everybody was silently dropped. "The world spends more on research" resolved to no country and therefore to nothing; it now applies to every country at once, as do levers aimed at a bloc or a coalition.

3.12.0 September 2026

Fixed: the globe on the opening screen of the simulator was an ellipse. The script sized the bitmap from the panel and then wrote that size onto the canvas as an inline style, while the stylesheet carried a cap on the canvas height, so whenever the panel was taller than the cap the element was displayed shorter than the bitmap it held and the browser squashed it. The cap is on the panel now, the canvas fills whatever the panel turned out to be, and the radius comes off the shorter side, so it cannot be anything but a circle. A test reads the pixels and measures the drawn disc rather than trusting any of that.
The opening globe can be turned by hand. It takes a drag, carries the momentum of it afterwards, and picks its slow rotation back up once that has run out and a moment has passed, so it never fights the hand. The pitch stops short of the poles, which is the one way a freely turned globe becomes unusable.
It also has something happening on it: a dozen places, and routes between them that travel one after another in the signal colour. The opening was a shape turning over, and the first thing somebody sees should be the thing the product is about, which is that places are connected and something happening at one of them arrives at the others.
The tab strip keeps its tabs for as long as it can and becomes a menu when it cannot. Twelve tabs need about eighteen hundred pixels once the brand, the reliability chip, the search key and the run button have taken their share, so at sixteen hundred the last tabs were cut off with nothing saying so. Which of the three happens is measured rather than guessed at a width somebody picked: the strip is laid out and measured, it tightens the air around each word if that is what makes it fit, and only when it would be showing less than about four fifths of itself does it give way to one button carrying the name of where you are. Pressing that lists everywhere you could be, with what each one answers beside it.
Fixed: that measurement made the whole instrument slow. Deciding whether the tabs fit changes the size of the tabs, which woke the thing watching for a change of size, which decided again, and the two took turns for as long as the window was open. The answer is worked out once, kept on the header, and taken again only when the room or the tabs themselves have changed. Running the clock forward a month at a time went from a crawl to immediate.
Fixed: the tabs did nothing while the opening was on screen. The opening held the stage until one of its own two buttons was pressed, so every tab above it looked live and changed nothing. A tab is a place, and pressing one takes you there now; if nothing has been run yet the baseline starts on the way, because a tab with no run behind it is an empty room.
On a narrow window the search key loses its word, the brand loses its line of description and the save button steps aside, because the one control nobody can do without on that screen is Run.
A browser set to reduce motion now has the stagger removed as well as the movement. Blocks were still waiting their turn to appear, up to a sixth of a second each, which is the same staircase the setting asked not to see, arriving in one jump instead of sliding.

3.11.0 September 2026

Rows instead of boxes, which is the change that fixes most of what was wrong. The sections were built out of cards in a grid that fitted as many columns as would go, and a grid of cards only works when every card holds about the same amount of text. This content does not: one entry runs forty lines and the next runs three. Five columns of that gave one tall card, four short ones and a quarter of the screen left blank. Every one of those sections is now an index, a title and a body on three columns, separated by hairlines. A row cannot leave a hole and its columns cannot drift.
Everything is about a third smaller. The headings were set at a size where three words filled a screen, which is not authority, it is shouting.
The light theme is neutral. It had warm paper under cool grey text with a blue field behind it, which is two temperatures fighting, and no amount of care spent on any one of them fixes that. There is one near white, one ramp of true greys, and the signal colour is the only colour on the page.
Fixed: the globe in the hero was an ellipse. The canvas was measured once and then only on a window resize, so the reflow when the web fonts land left the drawing surface at the old shape and the browser stretched it. It is measured from its own box whenever that box changes now, and the radius is taken from the shorter side so it can never be anything but a circle.
Fixed: the footer directory ran all its links together on one line. They had been set inline so each could carry an underline on hover, and inline is exactly what a column of links must not be. The whole footer was rebuilt: proper columns, set small, under one hairline, with the name at the size a name should be rather than at fifteen rem in outline.
The field behind the page was twenty six tight lines in a colour that was not the page\u2019s, which read as ruled paper. It is about half as many now, in the page\u2019s own grey, flat at the top of the screen and with more relief toward the foot, with one slow band of light travelling across them. The pointer bends the lines rather than pushing them apart, because a bulge looks like a ripple and a bend looks like terrain.
The spacing scale, the ticker speed and the index were all brought down with everything else. Twenty seven pages checked at desktop and phone width after all of it.

3.10.0 September 2026

The simulator stopped being a page of this site and became a document of its own. It carries its own header, its own menu and its own light and dark switch, so wrapping it in the site\u2019s header meant two brands, two menus and two theme buttons on one screen, and a strip of chrome taken off the top of a map. It now gets the whole window with nothing around it, and every link to it opens in a new tab so the site is still there behind it.
Fixed: the two light and dark switches on that page moved different things, which is what having two of them does. There is one now.
Fixed, in the simulator: there were two implementations of the theme switch, one that moved the theme attribute and one that moved the attribute and the values a richer theme writes on the element. Whichever ran last decided how much of the screen turned over, which is how a switch ends up changing the text and leaving the field behind. The older one now clears what the newer one may have left, so the instrument lands in one state whichever of the two ran, and a test presses the button and checks that every value moved.
A new typeface. Geist carries everything on the site that is read as language, at every size from a seven rem headline down to a caption, with Geist Mono for anything that is a measurement and Instrument Serif in italic still carrying the emphasis. One family set well reads more modern and more elegant than two set adequately.
The index was set at a size that suited a manifesto rather than a menu. It is a menu.
Where your own data goes in. A new screen in the simulator\u2019s admin takes a CSV of corrections to the base year: four columns, thirty settable fields, each with its unit and the range it admits. The range is a refusal boundary rather than a style guide, because a unit mistake reaching the model produces a confident wrong answer, which is the most expensive thing this can do.
It checks before it stores. The first button reads the whole file and shows what would happen without writing anything: every accepted row with the value the world holds now beside the one you are giving it, and the percentage between them, which is how a decimal point in the wrong place is caught. Bad rows are refused one at a time with the reason, so two hundred rows with three faults store a hundred and ninety seven.
Fixed: an upload could be accepted, stored and reported as stored while nine of the thirty fields it set were dropped on the way into the world, because the merge still only admitted the sixteen the World Bank publishes. No error anywhere and the wrong number on screen, which is the worst shape a fault can take.
Everything lines up. A new audit measures, on every tab, whether the blocks in a pane start and end on the same two lines and whether a column of figures shares one right edge. It found that a branch in the futures list shrank to its own content rather than filling the block, so a list of branches was a list of different widths, and that every claim in the working panel sized its own table, so the columns jumped sideways each time a claim was opened. Both fixed; the audit now reports nothing.

3.9.0 September 2026

The site was redrawn from nothing. It is a plate now: the kind of large format sheet a survey office issues, with a number in the corner, a scale down the edge, registration marks where the grid starts and hairlines holding the figures apart. Twenty seven of them make the document and each one says which it is.
All of that furniture is real rather than decorative. The rail down the left edge is the scroll position, and the name on it changes to whichever section you have reached, so on a reference plate of two thousand rows the rail is a place rather than a scroll bar. The number in the corner is where the page sits in the reading order. The marks at the corners of a card are where its grid begins.
A contour field is drawn behind every page: a stack of iso-lines that drift, lean away from the pointer and tint toward the signal colour as it passes. It is the one piece of decoration on the site, it costs a single canvas, and it is what makes a near black page read as depth rather than as a switched off screen.
Emphasis became a different voice rather than a heavier weight. A word the copy marks is now set in italic serif at a larger optical size in the signal colour, against display type tracked tight to minus four and a half percent. The collision between the two is the typographic idea of the site.
Headlines are split into words, each behind its own mask, and rise into place one after another. Figures roll up to their value. The strip of readings under the hero is driven by the scroll rather than by a loop, so it speeds up, slows down, reverses and skews with it. Catalogue rows are dealt out one at a time rather than appearing as a wall.
A precision cursor that lags the pointer and names what it is over, magnets that lean a control toward the pointer, a light that follows the pointer across a card, and a shutter of six bars that closes over the page between plates with the name of where you are going written across it.
An index that opens as a circle growing out of the button that opened it, holding the whole document at a size you can hit with a thumb. Twenty seven pages will not fit in a menu bar, and pretending otherwise is how three quarters of a site ends up unreachable.
The light and the dark are both drawn rather than one being inverted from the other. Light is the plate as printed on bone paper; dark is the same plate on a light table at night.
Every one of it is switched off by the reduced motion setting, and every page is complete with the motion file never loading at all. Twenty seven pages are checked at desktop and phone width for console errors, horizontal overflow, unstyled content and anything left invisible.
In the simulator: the dark themes were rebuilt. A dark palette is not a light one inverted, and this one had been treated as though it were. The steps between the panes were too close to see at the bottom of the scale, the hairlines with them, and the quiet ink was quiet enough to be unreadable. Every step is now wider than its daylight counterpart.
The signal colour is lifted, at the moment it is applied, against whatever pane the chosen theme actually turned out to have. Vermilion picked to sit on white is four to one on a slate pane, which is a label you lean in to read. Rather than keep a second hex for every accent in every dark theme, the accent is measured against the pane and raised until it clears.
Filled buttons stopped using the signal colour as their fill. White on vermilion is 4.2 to 1, which fails for a nine pixel label, and a nine pixel label on a button is exactly the text somebody has to read. The fill is now the accent taken down until white on it clears, computed for whichever accent is chosen, so amber travels a long way and indigo hardly at all.
An eighth theme: Carbon, a neutral dark with no blue in it, because the three darks before it were all cool and a grey field suits a room lit by anything other than daylight.
All of it is held by a new test that walks every tab in every dark theme, finds what each run of text is actually drawn on rather than what it was meant to be drawn on, and fails on anything below the contrast threshold for its size. It found twenty three distinct failures on the first run.

3.8.0 September 2026

A link to exactly what you are looking at. It carries the view, the layer, every overlay, every flow, the year, the selection, the network scale and where the camera is, and whoever opens it lands in that view rather than at the opening. It is a fragment, which browsers do not put in a request, so nothing about it is sent anywhere.
It carries what you are looking at rather than what you ran, and the instrument says so when it opens one: your view over their own run. A link that pretended to carry a scenario it had not run would be the one dishonest thing this could do.
The map saves as a picture that explains itself: the reading and its unit, the projection, the year, how many overlays and flows are on, how many futures the run holds, the date, and the attribution. A screenshot of a map has no scale, no source and no date, so nobody downstream can tell what they are looking at. This one is still readable a year later on somebody else\u2019s desk.
The network view has a second legend. Sixteen kinds of dot with no key is a picture nobody can read, so it now lists every kind present with a count beside it, and appears only when there is something other than countries to explain.
Fixed: the camera could not be placed at a given longitude and latitude, because the offset was computed with the sign the wrong way round. Nothing used it before, and a shared link would have landed hundreds of degrees away.
Three new reference pages. The network, which explains the six hundred and fourteen entities, the ten kinds of tie, the three scales, and what the impulse actually does, including why it is not two billion nodes. The flows, all twenty one, split into the ones that come from the model, the ones that ask a different question of the same trade, and the ones that are inferences and say so. The indices, with the rule the four composites are built to and what an index cannot do.
The site is thirty pages, and the reading order picked up all three without anybody editing the footer, because the footer is built from the reading order rather than kept beside it.
The link and the picture are both covered by tests: a link is made in one page, opened in a fresh one, and every field including the camera has to come back the same.

3.7.0 September 2026

Every layer, overlay and flow was checked properly for the first time, and the check found real failures. The earlier sweep counted ink on the canvas, which the basemap supplies on its own, so every overlay passed whether or not it drew a single pixel. The new check turns each thing off, hashes the canvas, turns it on and hashes again. A layer that changes no pixel is a layer that does nothing, whatever its checkbox says.
Fixed: the tension overlay had never drawn anything, in any scenario, since it was written. Tension is an index from nought to one, and the layer tested it against eighteen and divided it by nine, which are the numbers for a nought to a hundred scale. No country has ever scored eighteen on a scale that stops at one. The thresholds are now set against what the figure actually reaches, which is a median near seven hundredths and a maximum near a fifth.
Fixed: the arms flow drew no arcs for the same reason, a cut at a fifth on a figure that rarely gets there.
Fixed: a layer turned on from anywhere other than the panel drew nothing at all. Loading the geometry it needs was wired to the click handler, so a layer set from a saved scenario, a shortcode attribute, the boot configuration, the command palette or a link silently did nothing. The drawing code now asks for what it needs itself, once, and repaints when it arrives.
Towns wait for the map to come closer before they draw, because forty four thousand of them at world zoom is a band of ink rather than a reading. What was wrong is that the panel said nothing about it, so a layer holding back looked exactly like a layer that was broken. The row now says "zoom in" while it is waiting, and stops the moment it is drawing.
Twenty four new readings, all of them made from figures the instrument already holds: share of world output, real policy rate, the misery index, debt service, output in money rather than in index points, people in cities, how young a country is, working age people, military and research spending in money, total emissions rather than per head, energy per head. Plus five that read the physical layer as a country reading for the first time: datacenter capacity, cables landing, container capacity, refining capacity and generation sited in each country.
A new group of four, Structure, which asks what a country\u2019s position in the network looks like rather than what it holds: how many trade partners it has, how concentrated that trade is as a Herfindahl index, the share sitting with its single largest partner, and how far it is from the nearest chokepoint. The concentration reading is the one that says who is fragile, and the largest partner share is the one that says who has leverage over them.
One hundred and eighty nine layer readings became two hundred and seventy eight.
Eight new flows: students, travel, assistance, knowledge, embodied carbon, virtual water, investment and medicine. Embodied carbon is the map that says who is emitting on whose behalf, which a per head map cannot. Virtual water is the water that went into food that then crossed a border, which is how most water actually moves. Thirteen kinds of flow became twenty one, which is 1,386 distinct flows across the directions and groupings.
All of it is now covered by a test that fails if any layer, overlay or flow stops drawing: thirty eight overlays, seventy one base fields and twenty one flows, checked one at a time against the pixels.

3.6.0 September 2026

The network stopped being a picture of countries and became a graph of everything the instrument carries. Six hundred and fourteen entities across sixteen kinds: the two hundred and thirty eight economies, their five alignments and sixteen treaty organisations, thirteen chokepoints, thirty eight subsea cables, twenty three pipelines, thirty four LNG terminals, thirty container ports, fifty three datacenters, twenty five exchanges, thirty airports, twenty five refineries, twenty four critical mineral operations, thirty power stations, twenty five research clusters, and the five goods the model actually clears.
Two thousand seven hundred and sixty one ties across ten kinds, each with its own meaning and its own conductance: trade, alignment, membership, sited in, lands at, transits, supplies, depends on, powers, feeds knowledge. A tie that means nothing is worse than a tie that is missing, so every one of them is a relationship somebody could go and check.
Three scales, each a superset of the one before. Countries is the trade network the engine runs on. Physical adds the hardware trade moves through. Everything adds compute, finance, materials and the goods, which is the scale at which a shock can be followed from a mine to a datacenter.
The graph conducts. Click any node and an impulse spreads from it, divided at every node it passes through so nothing is created, decaying with every hop. What lights up is what that shock actually reaches, and a panel lists it in order with the share of the original that arrived. A shock at the Strait of Hormuz reaches the United Arab Emirates, Qatar, Bahrain, Oman, Iran and Kuwait first, which is the right answer and nobody wrote it down anywhere.
This is the one thing a network of this kind genuinely shares with a nervous system, and it answers the question a diagram cannot: not who is connected to whom, but who finds out when something happens here.
Fixed: the moving dots along the ties read a country reading by node index, which stopped being a country index the moment the graph held cables as well as countries. Fixed: a cable whose every landing sat far from a country centroid floated with no ties at all, which said something false about a line that certainly comes ashore somewhere.
The network\u2019s three rows of controls stack down the right of the plate rather than wrapping over the graph itself.
The graph, the three scales and the impulse are covered by tests: that each scale contains the one before, that nothing floats at any of them, that an impulse reaches part of the graph rather than all or none of it, and that a shock at a chokepoint reaches the countries that actually depend on it.

3.5.1 September 2026

Fixed: closing the left panel on its own erased the map. The two side panels are never removed from the page when they are put away, they collapse and stay where they are, so the grid always holds three children. Closing one dropped the layout to a single column, the inspector was still eight hundred pixels tall, it took the only row, and the map was left with none of it. Each panel is now placed in a named column, so how many of them happen to be showing cannot move the others, and the single row is pinned so nothing can take the map\u2019s height again.
Fixed: the map could be drawn into a two hundred and sixty pixel strip. A measurement of zero was being treated as a size and clamped to a floor rather than refused, which is what produced a world squeezed into a sliver of a full height slot. A measurement that is not a size is now refused, whatever is on screen stays, and the size is taken again on the next frame.
Fixed: the globe\u2019s imagery was soft, and street labels crawled from one pixel to the next as the planet turned. Two causes. The sheet covered the whole world even though a globe can see at most half of it, so nearly all of the memory that decides the zoom was spent on places that were off screen, and the zoom was consequently one or two levels below what the screen could show. And the sheet was read one pixel at a time with no filtering, so at any magnification a label broke up, and which pixel each point of the screen landed on changed frame by frame as the globe moved.
The sheet now covers only what the globe can actually see, clipped to the window rather than to the planet, so the same memory buys the resolution the screen needs. On a sphere zoomed in, the tile zoom went from four to six. The sheet is read bilinearly everywhere except during an active drag, where nothing is legible anyway. Measured across three sphere zooms, imagery is now oversampled rather than magnified at every one of them.
Sheets are released the moment they are replaced rather than left for the collector, so walking the zoom no longer leaves a string of large buffers alive at once.
The two columns beside the map gave up thirty two pixels each. The map is wider by sixty eight pixels with both panels open, and it does not stop there: the panels hold their width on a larger screen and the extra goes to the map.
Every base layer, every overlay and every flow in every direction was swept and checked to draw: 189 layer readings, 38 overlays and 13 kinds of flow read three ways each. All of them draw.
The panel behaviour and the globe\u2019s sharpness are both now covered by tests that fail if either comes back.

3.5.0 September 2026

The comparison page was one chart and a table. It is now a reference column you choose, every world measure drawn as a tile at once, a difference beside every value rather than two nine digit figures to subtract in your head, a table of which countries moved and by how much, a lever diff that says what actually differs between the columns, the branch mixes side by side, and a written read of what the comparison found.
Kept runs can be renamed, reloaded into the workspace, removed on their own, and are numbered automatically if two would otherwise carry the same name. The page warns you if the columns were not run on one seed, because then part of every difference on it is luck and the arithmetic cannot say how much.
The futures page answers in odds rather than in bands. The chance of a year of world contraction, of inflation above four percent, of energy half again its base price, ranked, each with how many years of the median path were already past that line. It says plainly that these are read off five quantiles rather than counted path by path, so they are worth reading as one in five and never as twenty one percent.
A chart of how much less is known each year: the width of the middle eighty percent as a share of the median, per measure, with a line saying which way the band leans and therefore which tail is worth planning for.
Opening a branch now opens the branch. Its four series against the ensemble median, what makes it itself in standard deviations on each axis, how many futures landed in it and where the focus economy ended.
The causes page adds the whole run rather than one year of it. Every term summed across every year, ranked, with the year each one peaked, and a line saying how much of the movement cancels. Below it, the same term in the six countries this one trades with most, which is the only thing on the page that separates “this is this country” from “this is the world”.
The monitor carries forty eight columns instead of eighteen, grouped and chooseable, including the four composite indices and four readings taken from the run itself. Sorting works on every one of them, and the table draws four hundred rows at a time rather than stopping at four hundred.
Fixed: the labelled proportion bar used in eight places across the instrument had no stylesheet at all, so every one of them rendered as a line of run together text. Rankings, trade partners, variance decomposition and the sensitivity list all read as bars now.
Four more live sources, all open and unauthenticated: NOAA space weather, GDACS disaster alerts, United Nations humanitarian situations and urban air quality over thirty cities. Space weather is the only live reading that changes what another layer means, because a geomagnetic storm raises drag across the whole low orbit population at once, and the status strip says so when it does.
The tab strip fades at the edge it can still travel toward, scrolls on a plain wheel, and brings the selected tab into view. Eleven tabs stopped fitting a laptop some releases ago and nothing had said so.
Every page on the site now says where it sits: which section, how far through it, what came before, what to read next, and what else is beside it. Twenty seven pages is a document rather than a menu.
The method, glossary, about, developers, accessibility, satellites and what it is not pages were roughly doubled. Method now walks through what happens when you press run and answers the six questions that get asked most; the glossary carries forty terms and a list of the words this instrument deliberately avoids.
More motion, and none of it load bearing: every block of substance arrives rather than being found already there, the opening line sets itself a word at a time, the call to action leans two pixels toward the pointer, and the field behind the page moves at a tenth of the scroll. Anything an arrival could have missed is simply shown after three seconds, because a missed animation would otherwise be missing content.

3.4.0 September 2026

Fixed: the menu and the footer were being rendered at the theme’s type size rather than their own, which is why the links were twice the size they were drawn at and the bar was truncating words. The chrome now holds its own typography against the theme, on a short and deliberate list of properties, and only inside elements this plugin owns.
The bar folds when its links stop fitting, measured rather than guessed. A breakpoint is a guess about how much room a set of words needs, and the guess is wrong the moment a theme, a translation, a larger default size or one more page changes the type. Nothing is ever cut off now, because nothing is ever asked to be smaller than it is.
Either side panel in the app can be put away on its own, with a handle on its inside edge and a tab where it was. Focus still takes both at once. The map goes from 904 pixels to 1600 on a normal screen.
The flat map got its panel back. The legend, the zoom and the licence line used to sit in a band underneath, which cost a hundred and fifty pixels of map to hold three small boxes; they now sit on the plate’s own corners where every map in the world puts them.
Streets at full detail is the default ground. It is the one basemap that carries its own place names, so the map reads as a map before a single layer has been switched on.
Fixed: the ground went soft whenever the globe was moving, and blinked out entirely on a zoom. The sphere is now built at the screen’s own resolution rather than at half of it, only an active drag lowers that, and a new sheet is built beside the one on screen and swapped in only when it is ready. The ground never goes away and the place names in a street map stay sharp.
Six more overlays: container ports by the boxes they move, airports by passengers, refineries by what they distil, critical mineral operations by share of world production, the largest power stations coloured by what turns the generator, and the research clusters that are upstream of all of it. Three hundred and fifty curated sites in all.
Four more kinds of flow: remittances along the path people took, freight weighted by distance rather than value, bandwidth over the cables, and arms as an explicitly marked inference. Nine kinds became thirteen, and 594 flows became 858.
The network keeps more of each country’s ties, marks the ones both sides count among their own, and strands nobody: a country whose partners are all too large to count it among theirs used to float with no edges at all, which said something false. 927 ties became 1,388.
A country picker in the inspector, because clicking one on the map only works when you can see it.

3.3.0 September 2026

Fixed: the layers panel jumped back to the top every time you ticked something, because every switch rebuilt the whole list. A switch now repaints its own row and nothing else, so the list stays exactly where it was, and the two changes that really do alter the panel’s shape carry the scroll position across.
Fixed: the Open the simulator button was being cut off. The bar was one flex row, so every page added to the menu took a few pixels off the end of the button. The links and the actions are now two separate boxes: the links shrink and then fold into the burger, and the actions never give an inch.
Focus mode. Press F on the map and the scenario rail and the inspector go away, giving the world the whole window. The same key, or escape, brings them back.
Eight new pages, all of them reference rather than pitch: every layer and what each of the five readings is for, every lever with what it acts through, the model block by block, the satellites and why the navigation constellations are deliberately absent, the sources, a developer page with the shortcode and the routes, the keyboard, and an accessibility statement that starts with where the product falls short.
The reference pages read the simulator’s own catalogues rather than keeping a second copy, so they cannot drift from the thing they describe. With the plugin absent they say so and show nothing instead of printing numbers they cannot stand behind.
Motion, and a rule for it: a thing moves to say what it is, where it came from, or that it just changed. Section rules draw themselves, figures count up to what they are, the light on a pane follows your pointer along its top bevel, registration marks converge as a pane arrives, a rail down the side says where you are in a long page, and leaving a page is a handover rather than a flicker. All of it stands down under reduced motion.
In the app: changing projection is now a handover rather than a cut, a figure that moved because the year moved lights for a moment so you can see which three of fifteen actually changed, and a switch you just flipped marks itself.

3.2.0 September 2026

It opens on the globe, and the globe is now a real one. Satellite imagery, street maps, topography, relief and bathymetry are drawn on the sphere itself rather than only under the flat map, sampled per pixel through the inverse projection so the coastlines curve toward the limb the way a planet’s do instead of tearing at the seams.
Day and night, computed from the clock. The terminator is where the sun actually is, with the real eighteen degree twilight band, so the lit half of the ball is the lit half of the world while you are looking at it.
Fixed: the globe would not tilt. Dragging up and down rolled the picture round like a steering wheel instead of tipping the planet, because the pitch was mixing the two screen axes with each other rather than mixing depth with up. It now tilts properly, all the way to either pole, and there are buttons for the poles and the equator.
Fixed: the boundaries, the towns and the villages could not be fetched on some hosting. They are large files, and several hosts refuse to serve a large script from a plugin folder even though it is there. There are now three roads to the same data, the app takes whichever works, and if all three fail the message names every URL it tried. A new admin screen tests each file and says which one your host is refusing.
The physical layer. Submarine cables drawn as great circles through their published landing points, long distance oil and gas pipelines with whether they are actually running, LNG terminals by capacity, datacenter campuses by announced megawatts, exchanges by what is listed on them, and the thirteen chokepoints with the barrels, the transits and what happens when each one closes.
Satellites, computed here. Orbital elements arrive from CelesTrak every six hours and the position you are looking at is propagated in your browser by SGP4, the standard model, for this exact second. It agrees with the reference implementation to within nine metres over a full day. Stations carry their ground track for one orbit.
Six more live sources, fetched by the server so no key exists to leak and no visitor ever talks to them: navigation interference, protest reporting, markets, military aircraft, natural events, and orbital elements.
Four computed indices: instability, resilience, chokepoint exposure and scored hotspot. Each is a weighted composite of fields already on the map, recomputed from the run on screen rather than stored, and opening a country shows every term and its weight.
Seventeen more levers, taking it to forty four. Public capital, schooling, research, retirement age, reactors, strategic reserves, desalination, agriculture, stockpiles, bank capital, price controls, ports, transfers, capability regulation, storage costs, world rates and financial stress. Every one has a real channel through the model, and every one that costs money is paid for out of the primary balance.
A command palette. One key, one field, and thirteen hundred commands generated from the catalogues rather than written by hand: every layer, every flow, every country, every lever, every view.
The network view now wraps round the planet. Same force layout, put back on the earth, with ties drawn as great circle arcs over the surface and the far side hidden behind the horizon.
The reliability score now says why it is what it is. Every term that took points off, how many it took, and the one thing that would recover them, with the largest marked.
A sources page, in the app and on this site, built from the same catalogues the panels are built from so the two cannot disagree. A hundred providers, sorted into measured, curated, derived and modelled.
Fixed: the world map on this page was stretched to whatever shape its panel happened to be, which made Africa wide on a desktop and tall on a phone. One scale for both axes now, always.

3.1.0 September 2026

A third projection. Flat is the atlas and the globe is the earth; network drops the geography entirely and lays the same countries out as a force directed graph of who actually trades with whom, so the shape on screen is the world economy rather than the world map. The layer, the year and the selection follow you between all three, and the graph is laid out fresh from the state on screen, so changing the year moves the countries.
Five hundred and ninety four flows, up from six. Nine kinds, each computed from the state on screen rather than drawn from a list, crossed with three directions and twenty two ways of narrowing it: everybody, five blocs and sixteen named groupings. Outbound chips from the West and inbound energy to China can be on the same map at the same time.
Three new kinds of flow to go with them: compute, capital goods and minerals.
A camera layer. Cameras draw at their own positions and refresh at whatever rate their owner publishes. There is no open worldwide camera network that works without a key, so nothing is shipped pretending to be one: the list is yours, entered in the settings, with Windy Webcams added on top if you have a key for it.
What is happening now can become what you are asking about. Click an earthquake or an active emergency and the app offers to write it into the scenario it is holding: a magnitude scaled productivity shock, water stress for a drought, a pandemic term for an outbreak, tension and military spend for an armed conflict. Add it and run, or add it and keep editing. The translation is a reading rather than a measurement, and the panel says so before you accept it.
A live strip under the header: how many earthquakes today, how many emergencies open, how many places in the news, how many cameras. It refreshes itself every five minutes while the page is open, and opens into the full list.
Fixed: the network layout collapsed into a single hairball, because the largest economy is in almost everybody's top partners. Attraction is now normalised by how many ties each country has, so the graph reads as a world rather than a star.

3.0.0 September 2026

Rebuilt as a Japanese instrument three hundred years on. An engineered white surface lit from within rather than printed on, panes that catch the light along their top edge, registration marks at the corners the way an instrument face carries them, and one signal colour, vermilion, for the thing you did. Light is now the default, with the same instrument at night a click away and the choice kept in your own browser.
New typefaces throughout: a Japanese geometric gothic for the interface, a Japanese monospace for every figure, and a mincho kept back for the display lines alone.
The flat map is now Web Mercator, which is the projection every tile server on earth already speaks. That is what lets the zoom keep going past the country layer to regions, counties, cities, towns, roads and roofs, with everything landing on the same pixel.
Satellite imagery, street maps, topographic maps, shaded relief and bathymetry can now be drawn under the data. All open sources that need no key, all off until you turn one on, and whichever you choose names itself under the map.
A hundred and sixty nine ways to read the world, up from twenty nine. Every field can now be read as it is, ranked, against the world median, against its own region's median, or as the change since today. None of them is a new dataset: they are the same numbers asked a different question.
Forty four thousand towns, down to five thousand people, as a second tier below the cities.
Step through a run by year, month, week or day. Every step says what moved across people, stability, resources, capability and the economy; why, in the engine's own named terms; what else could be happening instead and with what share of the futures; and what would change it.
What to change now: twelve candidate changes, each added to your scenario at the step you are standing on and the whole model run again on the same seed, ranked by what it actually does. It is the model's answer rather than a rule of thumb about it.
The map no longer opens on a projection of growth. It opens on income per head, which is a fact about the world rather than a forecast of it, and every comparison now reports across people, stability, resources and capability before it reports an economy.
The example sentences and the placeholder are gone from the input, and the gallery of them is gone from this site. The box is empty because the question is yours.
Fixed: panels in a scrolling column were shrinking to fit and quietly clipping their own contents.
Fixed: an element level reset was out-specifying the classes it existed to serve, so buttons lost their borders and their padding.

2.3.0 September 2026

The map is now divided below the country. About a thousand states, provinces, prefectures, departments and Eurostat regions in the first tier, and four and a half thousand United States counties and European districts in the second. Point at one and it names itself; click it and it is selected.
A new layers panel. One base layer chosen from forty three, then every overlay and every flow as its own switch, so any number of them can be on at once rather than one at a time.
Four and a half thousand cities of a hundred thousand people or more, sized by population, drawn more thickly the further you zoom in.
New overlays: capitals, maritime chokepoints, semiconductor clusters, nuclear weapon states, the modelled tension index, a graticule and an alliance wash.
Two new live feeds: humanitarian emergencies currently marked ongoing, from ReliefWeb, and where the world's English language news is being filed from, through the GDELT Project. Earthquakes now start at magnitude 2.5 rather than 4.5 and are drawn at their real position, and current conditions are pulled for twenty eight capitals rather than ten.
A fourth flow: leading edge semiconductors, from the eight clusters that hold the capacity to the economies that consume it. Trade, people, energy and chips can all run at the same time, each in its own colour.
Fourteen more ways to colour the world: policy rate, exports, the split between farming, industry and services, population growth, births per woman, density, governance, nuclear weapons, land area, oil production, energy produced and energy used.
The scenario library is gone. There are no templates to load: the sentence box is the way in, and the How it works page now shows what people actually type.
Seven more use cases, and a page of its own for the map.
Fixed: the year beside the timeline did not follow the slider when you dragged it.

2.2.0 September 2026

Rebuilt as an instrument rather than a brochure. Monospace for every label, number, button and status; square corners throughout; hairline rules you can actually see; one signal colour on near black, and no gradient anywhere.
The home page now opens on a live world map panel rather than a decorative globe, framed the way a product shows its own screen.
The simulator wears the same language: the same typefaces, the same squared chrome, the same single colour.
The default colour is now a signal green. It is still one field in the settings, and changing it restyles the site, the map, the mark and the tab icon together.
Fixed: a reset on paragraph margins was out-specifying every component, so spacing under the hero buttons and in several panels was collapsing.
Fixed: bordered grids left empty cells at some widths, drawing a box around nothing.

2.1.0 September 2026

Ask in a sentence. Type what you want to change in plain words, and the simulator turns it into levers, shows you how it read them, and runs it. The grammar runs in your browser with no key and no network; a model provider, if you have one configured, is only asked when the grammar finds nothing.
The simulator now opens on a world rather than a loading message. A globe, a sentence box, and the first run happening behind it.
Every scenario is now run against a baseline on the same seed automatically, so the difference between the two is your lever rather than the luck.
New: What changed. Your scenario against the world carrying on, for the whole world or one country, with the chain behind any row one click away.
New: consequences you may not have expected. Countries your scenario never named, ranked by how far they moved from their own baseline, with the term that moved them.
New: The chain. What you changed, how it travelled, what moved, what it came to, drawn as a flow rather than a list. Every figure is this run minus the baseline.
Simple and Expert modes. Simple is the sentence, the levers and one button; Expert opens the ensemble size, the horizon and the seed.
Confidence renamed to Reliability throughout, because a percentage next to the word confidence reads as a promise.
Colour in the comparison marks the direction of a change, never whether it is good news.

2.0.2 September 2026

The world state is now printed into the page itself, so the simulator is already holding it when the page finishes loading and never has to ask the server for it. The first load cannot be broken by permalink rules, a security plugin or a firewall any more.
If a request does fail, the screen says what to do about it and lists what each road actually answered, instead of showing a parser error.
The site has been redesigned: a new typographic system, a deeper palette with a warm second accent, a full bleed data strip, and sections that arrive as you reach them.
Seven new pages: who uses it, a scenario library, what it is not, a glossary, questions, a roadmap, and a clearer separation between the product pages and the reference ones.
A new mark and tab icon, drawn from the accent colour rather than uploaded.
Fixed: the hero globe never drew, because its script loaded before the geometry it needs.
Fixed: sections that arrive on scroll stayed invisible on a page where the script had not run.

2.0.1 September 2026

Fixed: on sites where WordPress had not rebuilt its permalink rules, or where a security plugin blocks /wp-json, the simulator answered with a page instead of data and refused to load. It now falls back to a second route that is always reachable, and says what to do if both fail.
Fixed: the page holding the simulator kept the theme's own header and styling while every other page was dark. It now wears the same header, menu and layout as the rest of the site, with a switch to turn that off.
The permalink rules are rebuilt on activation, so the very first load works.

2.0.0 September 2026

Every country on earth: 238 countries and territories, up from 73, each with a provenance mark.
Real country geometry. A choropleth flat map and a 3D globe with horizon clipping, drag to rotate, day and night shading.
A world monitor: every country in one sortable table, leaderboards, alliance groupings, country profiles and live open data feeds.
Four new levers: carbon price, water stress, trade agreements and debt restructuring, plus AI compute as a controllable good.
Coalitions. A tariff, control or sanctions package can be imposed by an entire bloc at once, and the engine weights the load by the share of world output imposing it.
Renewable share, emissions, water stress, connectivity and nuclear status now feed the model rather than sitting in a table.
Written briefings from your own runs, through a model provider you configure, with a monthly cap and a choice of who may spend it.
Full width layout, and a public site plugin that builds the home, pricing, method and legal pages.
Fixed: a sanctions package used to register against the country imposing it, which cratered the sender's currency.
Fixed: the trade matrices were shared between runs, so a previous scenario's tariff wall could still be standing in the next one.

1.0.0 September 2026

First release: 73 economies, 23 levers, Monte Carlo ensemble with common random numbers, branch clustering, exact causal decomposition, scenario comparison and the confidence page.

Product

Reference

Learn

Company

What changed

Site 3.12.1 16 October 2026

App 3.25.0 27 September 2026

Site 3.12.0 27 September 2026

App 3.24.0 27 September 2026

Site 3.11.0 27 September 2026

Site 3.10.1 25 September 2026

App 3.23.1 25 September 2026

App 3.23.0 24 September 2026

Site 3.10.0 24 September 2026

App 3.22.0 23 September 2026

Site 3.9.2 23 September 2026

App 3.21.0 23 September 2026

Site 3.9.1 23 September 2026

App 3.20.0 23 September 2026

Site 3.9.0 23 September 2026

App 3.19.1 23 September 2026

Site 3.8.1 23 September 2026

App 3.19.0 23 September 2026

Site 3.8.0 23 September 2026

App 3.18.0 23 September 2026

Site 3.7.0 23 September 2026

App 3.17.0 23 September 2026

Site 3.6.0 23 September 2026

App 3.16.0 23 September 2026

Site 3.5.0 23 September 2026

App 3.15.0 22 September 2026

Site 3.4.0 22 September 2026

App 3.14.0 22 September 2026

Site 3.3.0 22 September 2026

App 3.13.0 September 2026

3.12.0 September 2026

3.11.0 September 2026

3.10.0 September 2026

3.9.0 September 2026

3.8.0 September 2026

3.7.0 September 2026

3.6.0 September 2026

3.5.1 September 2026

3.5.0 September 2026

3.4.0 September 2026

3.3.0 September 2026

3.2.0 September 2026

3.1.0 September 2026

3.0.0 September 2026

2.3.0 September 2026

2.2.0 September 2026

2.1.0 September 2026

2.0.2 September 2026

2.0.1 September 2026

2.0.0 September 2026

1.0.0 September 2026