The Atlas dev blog

The Island's People Settled. Its Trade Never Has.

2026-10-20T00:00:00Z

Teluk Tinggi, a stilt town on the island's east coast, buys 596 tonnes of rice a year and feeds most of it to its hens. Each of the island's 900 settlements keeps a ledger like Teluk Tinggi's, and every ledger closes, yet the trade between them has never once settled: whenever it stops, scores of villages, lately hundreds, are still taking orders for more.

The island's economy after trade run 17 (2026-10-19). Every dot is a settlement, sized by the people who live there. Orange: fewer than the world's design asked for, because the land could not feed them. Teal: more, because the land had room. Grey: exactly as planned, as every city of 10,000 or more is. Gold lines are grain, pulses and chestnuts moving between two places, wider for more tonnes a year, drawn straight whatever road, river or sea lane they really take. Look at the west coast, where the plains' grain fans out from Ardbach and Alderheim, and at the teal crowd in the east. Underneath is Atlas's paper map, the island's official relief, of the terrain on 2026-10-20.

Why every village keeps books#

We have stopped that trade thirteen times, each after its eighth round, with between 85 and 478 villages still ordering. So we are changing what the economy is. It stops being a sum we redo from scratch and becomes a record that the world edits, one village at a time.

Atlas is a game set on one island, about a hundred kilometres across. Its settlements were placed by hand and by rule long before anyone asked what they eat. The last post followed their cargo down the rivers. This one goes back to where the cargo starts.

Each settlement has people, and the people have lifestyles: a peasant of the rainforest east needs 170 kg of rice a year and 100 litres of rice wine; a townsman needs a pair of boots, a hundred eggs and twelve baths. Needs call for makers: fields, yards, workshops, each with its recipe of inputs and outputs. The makers need hands, and the hands are more people with more needs. A settlement's ledger is that loop, closed. Every kind of thing it makes, uses, buys and sells, in kilograms, litres, head of cattle and pairs of boots.

The ledgers matter because a second program builds the world from them. Our town builder reads each one and lays the settlement out: so many houses, so many poultry yards, a woodcutters' camp, a market. If the ledger says 166 poultry yards, the town builder has to find room for 166 poultry yards. A wrong number is a wrong village.

Teluk Tinggi, by the numbers#

Take one ledger. Teluk Tinggi is a stilt town at a river mouth on the east coast: 542 people in 112 households, and 62 of those households live on piles over the water.

Teluk Tinggi's ledger after run 17. What it keeps is counted in establishments: a fishing ground worked is one crew's water. "Places" are the other settlements it trades with.

Walk its boardwalks and the ledger becomes a smell. 166 poultry yards make it a town of hens, geese and ducks. Every year 2,541 chickens leave it alive for seven places that keep fewer birds, and 1,351 geese and 1,260 ducks for two. Its fruit goes to 15 neighbours and its vegetables to 17. It has no fields to speak of, 20 hectares of gardens and orchards, and buys 141 cows a year from seven villages.

Its 596 tonnes of rice a year are more than a tonne for each of its people. We took that for a bug until we traced it. 465 tonnes of it is feed: a poultry yard keeps 100 hens, 20 geese and 20 ducks, so Teluk Tinggi's yards keep 23,240 birds, and in the rainforest east the only grain there is to give a hen is rice. The stilt town is a poultry town, fed by four rice villages.

Teluk Tinggi's trade flower: the 709 tonnes a year the island trade brings it. The ring is split by goods category, each by its share of the tonnes. Every category of 2 % or more grows a petal to the place that sends it most, as wide as its tonnes. Grain is 85 %, all but 5 tonnes of it rice, from the four rice villages; the petal points at the largest, Danau Makmur, with 196 tonnes. The cattle come from eight places, the most from Muara Putih. Underneath is Atlas's paper map of the terrain on 2026-10-20.

Then the odd ones. Teluk Tinggi sells 832 tonnes of firewood a year, all of it to a single buyer: Labuhan Harjo, the city next door on its sea wall, where a thousand people are to live over the water. Most of the wood is cut by four crews of farmers in the slack weeks between rice harvests. In the rest of the island those crews are winter woodcutters; the town builder asked us, quite reasonably, what a winter crew is doing in a rain forest. And it sells 24 tonnes of stockfish to four fjord villages in the Snowbound North, at the other end of the island. Nobody wrote "tropical fishing town feeds the Vikings" into a design document. The trade found it the cheapest way to fill a Nordic larder.

Labuhan Harjo's trade flower: 11,886 tonnes a year. More than half is firewood, 6,619 tonnes from 15 places. Its petal points at the largest, Bukit Sari, with 1,679 tonnes; Teluk Tinggi's 832 tonnes come fourth. Its grain comes from 14 villages, the most from Hanh Vuong.

Matching 900 ledgers#

No ledger balances alone. Teluk Tinggi has too much fruit and not enough cows, and so does everyone, in their own way. The island trade matches them.

It works in rounds:

From stock. Every settlement's surplus is matched to every deficit of the same kind, the cheapest pair first. The cost is the real journey: ox carts and mule trains on the roads, boats where the rivers allow, coastal ships between harbours.
By order. What is still short is ordered from villages with spare hands: a village with idle land grows rice for a neighbour, a village with woods sends a crew to cut for the town. Land comes first, then the farmers' slack-season crews, then anyone's spare hands, then the towns' crafts.
Again. Every village that took an order, or whose carts and boats changed, has its ledger balanced again with its new work. Its new work has new needs: the woodcutters want axes, the axes want iron.

Run 17's result is the map at the top: 73,942 flows of 288 kinds of goods, and not one of the 900 ledgers out of balance. On the way it also reckons the island's shadow side: 304 pilferers, 36 fences and 23 thieves' dens steal 506,063 hours of work a year, and the watch, raised against them, protects 313,535 more people a year.

Ardbach's trade flower: 34,905 tonnes a year, 84 % of it firewood. The firewood comes from 16 places, every one of them 48 to 92 km away as the crow flies; the petal points at the largest, Padang Putih, with 3,801 tonnes. Ardbach's grain comes from 47 places, its straw and manure from 80.

The land said no#

A ledger can ask for anything. The island cannot give it.

When we first added up what the ledgers asked of the land, the forests alone were asked for 314,826 hectares. More than half of that was hunting grounds, which had been claiming forest as their own: 162,637 hectares, 60,000 of them in the desert. Game now runs on any wild land, shared with the woods and the herds, and the desert's households burn brush.

The fields were worse. The world's design asked for 517,532 people. Their bread needed far more field land than the island's slopes hold, even with contour strips to 21 percent, terraces where the culture terraces, and the plains' fertile soil.

So the people moved, on paper. The design keeps its target for every settlement and never changes it. Beside it, the economy now keeps the settled people: those the land can feed. Every settlement under 10,000 lost the same share, the cities kept theirs, and villages with idle land gained the hands to farm it.

Each realm's people, as designed and as settled. The plains and the temperate middle lose most, the rain forest east gains a third.

The island now holds 480,809 settled people. 82,577 are missing from settlements below their target, and 45,854 have arrived at others above it. The extremes are small villages. Hanh Vuong, a stilt village in the east, was designed for 81 people and settles 281; the town builder's own site critics spotted the mismatch before we did.

Two numbers from this part surprised us most:

Chestnuts carry 44 percent of what the grain realms eat. Chestnut groves on the slopes too steep to plough, measured in the grain they replace, as in the Cévennes, where the chestnut was the arbre à pain, the bread tree.
19,808 hectares of field land lie idle while 13,343 tonnes of grain come from abroad. If a coastal ship carries 15 tonnes, that is about 890 shiploads a year. The fields are given and not farmed, because no buyer's order reaches them.

We costed what more land could still do: chestnut groves up to 40 percent slopes, the idle fields, terraces in every culture, irrigation, sheep and goats on the rough ground, ships. The land could settle about 48,800 of the missing people. Then we chose to keep things as they are for now. It is a good enough island to build on, and the next part explains why "for now" is the important phrase.

Eight rounds, 336 villages#

Trade run 17, 2026-10-19. Grey: settlements whose ledger was balanced again in that round. Orange: villages that took new orders. Round 1 touches nearly everyone; round 8 still touches half the island.

Each round, fewer villages take new orders, but the decline flattens. Round 8 still ordered 17.4 million units of goods (kilograms, litres and pieces, counted together) from 336 villages, and balanced 459 settlements again. Then the run stopped, because eight rounds is where we told it to stop.

Every trade run whose log we kept and that went all eight rounds. The early runs had a simpler economy and fewer orders; from run 11 to run 17, more than a third of the island's villages are still ordering when the run ends. Run 18, whose rounds stop ordering from villages that cannot make more, ends with about a quarter.

Two words need telling apart here. A ledger closes when what its settlement makes, buys and sells covers what it needs, at the journeys and orders of that round; after every round, all 900 do. The trade settles when another round would change nothing: no village takes a new order, no cart or boat moves, and no ledger has to close again. That has never happened. In the last round of each run:

Run	Villages still taking orders	Ordered in the last round, against the first
3	85	0.25 %
4	90	4.1 %
5	100	0.02 %
7	215	6.1 %
8	247	2.6 %
11	478	9.6 %
12	352	8.3 %
13	336	4.1 %
14	333	5.1 %
15	348	4.9 %
16	336	5.4 %
17	336	4.6 %
18	212	0.52 %

Each full run's eighth round. The orders are kilograms, litres and pieces counted together, so the share says how far from still the trade was, not what it weighed. Run 5's last round was nearly quiet, and still a hundred villages took orders.

Not one of them settled. Three things keep it moving:

Orders only grow. A round adds orders and never takes one back. If round 2 asked a village for too much rice, the village grows too much rice forever, and the surplus becomes someone else's cheap supply.
New work makes new needs. A crew cut for a buyer wants tools, the tools want a smith, the smith wants charcoal.
Each round is expensive. A ledger takes 0.3 to 7 seconds to balance, 2.6 on average. Run 17 balanced about 6,000 of them. A run takes hours.

For a sum we do once a week, that was fine. It stopped being fine when the world started changing one village at a time.

What changes when one village moves?#

Say the town builder lays out a settlement and finds room for 240 people where the ledger asked for 300, or for two fewer farms. That is a new fact about the world. It changes that village's ledger, which changes what it buys, which changes its suppliers' orders, which changes their ledgers.

Rerunning the whole island for that takes hours, and it gives a slightly different island each time, because the path matters. So we asked: what would it take to rebalance only what moved? We weighed five ways:

One village at a time, its change rippling along its trade partners until it fades. Cheap, but the result depends on the order things happened in.
A cheap island-wide step, run many times from where the last one stopped, able to take orders back as well as give them.
An island-wide check for imbalance, then spot rebalancing only where a threshold is passed. This one came from the team.
Regional markets: market towns hold the island's trade between them, and a village only deals with its market's books.
Prices, or the living world doing it over game time: stocks run low, prices rise, people move.

We chose the third, built on the second's machinery. The last two are for when the world runs.

A record, not a recalculation#

How the economy now works. The full run still exists, as the reset and the audit. Everything after it is a version with a parent.

The record. The trade's whole state goes into one database: every settlement's ledger as the trade sees it, every order, every flow, the carts and boats each ledger was balanced with, the errands neighbours run for each other, and the cost of the cheapest journey between every pair of settlements (about 810,000 of them). Its version is the fingerprint of its content, so two identical economies have the same version on any machine. Every version is kept in our storage, so any program can read the economy as it was at any point. That includes the town builder, the website that shows every settlement, and one day the game.

The check. It reads the record and today's world, and changes nothing. It flags a settlement whose own inputs changed (each has a fingerprint of everything the economy reads about it), anything a round would still order or should take back, ledgers out of balance, and carriers that moved. A change smaller than half a year of one workshop's work is not worth flagging.

The step. It balances again only what the check flagged, and runs rounds over those alone. It can take orders back now. Where the island has a kind unsold, the step withdraws ordered work, at most half of the excess at a time, from the most expensive delivery first. The damping is there so the economy settles instead of swinging. Every step writes a new version with its parent and the reason it was made: the town builder laid Teluk Tinggi, say, or someone built a glassworks.

The events. When the town builder lays a settlement, what it built becomes a fixed fact: the people its houses hold and the fields it placed. The economy settles around those facts and never asks the town builder to change them. A settlement is built once per design, and the loop stops.

The first step, and what it broke#

We tried it on 20 October. First the record: run 17, rebuilt from the tables it left behind. All 900 ledgers came back as they were: the same 73,942 flows, the same 40,013,394 units bought abroad. Two flows came back a hundredth of a tonne apart. That was the first good news. It also showed something we had not known: no run could ever have reproduced run 17 from scratch. Theft is reckoned from the counts the previous run left, so every run carries a little of the one before it. The worker processes also each drew their own random hash seed, so two runs on the same inputs could break a tie differently. We fixed that.

Then the check. One pass over the whole island takes 51 seconds. On a full run's record it flagged 602 settlements: 217 that a round would still order from, 423 holding ordered work that nobody buys, 205 whose carts and boats had moved. This is the never-settling, finally on a list.

Then the step, asked to polish that record without any change in the world. It did not settle. Over four rounds and 80 minutes it balanced 764 settlements again. Every round it ordered more from 217 to 463 villages and took work back from 205 to 423. And it made the island slightly worse: grain bought abroad rose from 13,435 to 14,677 tonnes a year, firewood from 15,906 to 17,632.

The reason is a rule we wrote too simply. The step took back ordered work wherever the island as a whole had a kind left unsold. But a surplus in the far west does not help a village in the east that loses its supplier. That village falls short and is ordered for again, from somewhere else. The fix is to take work back only where the buyer can be supplied from a surplus that actually reaches it. Before that, the rounds themselves must learn to settle.

Where this stands#

Running today: the ledgers, the island trade, the land budget with its settled people, and run 18 as the economy the town builder lays out.
The rounds, as of run 18: they now learn which villages cannot make more. A village whose woods, grazing, land or hands held it below its order is ordered no more that way, and its buyers' want goes elsewhere. Orders are contracts, delivered to the buyer they were made for. Run 18's eighth round ordered 2.0 million units, against run 17's 17.4 million, but 212 villages still took orders: closer to still, not still. The herds still move, because a village's grazing is shared among its herds by what each is asked. Two earlier runs that day stopped with 49 and 47 villages ordering, and bought twice the milk from abroad. Run 18 buys less from abroad than run 17, 121,169 tonnes a year against 124,331, but more grain, 11,947 tonnes against 9,885, and we are finding out why.
Being built: the versioned record and its check, both working: run 18's record is the economy's second version. And the step that closes again only the villages out of balance. It no longer swings: on its sixth trial, polishing run 17's record, it ordered from 38 villages in its first round and from one in its sixth, and the island bought 121,646 tonnes from abroad instead of 124,331. It takes work back now only where no buyer takes it at the village that made it. It stays an experiment until a change in the world, the town builder laying a village, runs through it.
Planned: the town builder's layouts flowing in by themselves as events; regional markets; prices; and a living world that rebalances as it goes.

The hardest part is not the code. It is deciding what is true. Until now, "the economy" was whatever the last full run printed, and any run could be repeated from its inputs. From here on, the economy is a chain of versions, each repeatable from its parent and its reason. The full run becomes the audit: if a fresh run and the stepped record ever disagree by more than the check's thresholds, one of them is wrong, and we will want to know which.

The question we can't answer yet#

Labuhan Harjo burns 832 tonnes of Teluk Tinggi's firewood a year, most of it cut by four crews of rice farmers between harvests. That is fine until Teluk Tinggi has a bad year. A harvest fails, a road washes out, a bridge collapses: the island changes for reasons no builder decides. When the world runs, imbalance stops being an error in a sum and becomes weather.

Then who rebalances Labuhan Harjo? The economy, or the people who live there?

For close readers. Theft is reckoned once, before the trade starts, from the counts as the last run left them, and held fixed for the run. The thieves of this year's island steal from last year's.

Another. 2,212 travellers sleep in the island's inns, almshouses, barns and hospitals on any given night: pedlars, tinkers, minstrels, pilgrims and beggars. They are in the ledgers as travellers whose round passes through, eating where they sleep, with no house of their own.

And one the economy found on its own. Nobody made Teluk Tinggi a poultry town. About 137 of its 166 yards keep birds for the neighbours: the households of every village within a day's walk fetch their eggs and hens from the nearest place that keeps a poultry yard, and for many of them that is Teluk Tinggi. The rice to feed the birds followed.

1.9 km of Road Ran Inside a Lake. Now It Keeps to the Shore

2026-10-19T00:00:00Z

On the east shore of Girpilio ezeras, a lake of 22 km² in the Great Forest, a cart track ran 43 metres inside the water's outline, on ground 1.2 metres below the level of the lake beside it. All five tracks round the lake strayed in like that: 1,902 metres of road inside a lake, and only 38 of them over open water.

Now the eight stretches that strayed have been laid again on the shore behind, and every other point of the five tracks lies where it lay. On this island the roads give way to the lakes: where a shore is known exactly, the roads keep back from it. Across the same water a ferry now crosses, and in the north a path goes down to a lake that never thaws.

Girpilio ezeras, its water milky jade with clay. Left: the road record before (version 38), with the stretches that run inside the lake's outline, or within 20 m of it, drawn red. Right: those stretches as laid in version 39 (green); the old line is dashed. The picture was drawn while version 39 was still a proposal, and its own label says so. The pale blue that spreads up the river mouths is the lake at its April high water. Look at the south-east corner, by Alksniai: the old track ran along the water's edge for 2 km. A plan view at true scale on the game's terrain; the lit render is not made yet.

Why the lake gets the last word#

Atlas is a game set on one composed island, about 99 by 101 km, with 900 settlements. Their roads are not painted: they were grown, settlement by settlement, by a router that prices every metre of climb, every bridge and every ferry. The result is a network of 2,965 stretches of road between junctions, from paved highways to footpaths, and it is canonical. Nothing regenerates it. When something is wrong, it is corrected forward with a new, recorded step, as at Mavigol, and settlements never move.

Lakes were the vague part. The old terrain knew where its lakes were only to the nearest 32 metres. This week the terrain gained a lake store: every lake on the island traced as a polygon at 4 to 8 metres, with its level, its high water, its name and what kind of water it is. It holds 294 outlines: the three lakes the game already has; 22 planned ones, fourteen of them lakes, two more that join a neighbouring lake, and the rest salt flats or dry floors; the 18 lakes and pools of Saat Sar, a spring-fed chain on the Lava Tip; and 251 other hollows that could hold water. The store has its own story, in the companion post 2,585 Hollows and Three Lakes.

A lake with an exact shore is a promise to the player: this is where the water stops. A road that ignores it walks you into the reeds. The team's whole instruction was one line: "Roads should adapt to lakes, not vice versa." The lakes are fixed, and the roads move.

1,902 metres#

We measured every road on the island against every outline in the store, every 2 metres.

At two of the three lakes the game already has, nothing needed to move. No road comes near Himinvatn, high in the Snowbound North. On Breitmeer, the broad mere of the Temperate Middle, the only road inside the outline is the ferry from Grunwald: 9,015 m of open water, its landings 11 m and 27 m inside the shore, where a ferry's stone mole and jetty belong.

Girpilio ezeras was different. Five cart tracks ring it, and all five strayed into it: 1,902 m in eight stretches, as far as 43 m from the shore. The road record's own line ran as much as 2.5 m below the lake's surface. No village depended on those metres alone, but every village on the lake used them.

The planned lakes are a bigger job and a later one. On the store's first outlines, which are being redrawn as this is written, they would drown 21 km of road on 44 stretches, and five villages would lose their only way out. Those roads wait until the terrain actually holds the water. A road that swerves round a lake that isn't there yet is its own kind of nonsense.

Dry land below the waterline#

Here is the odd part. On today's map, only 38 of those 1,902 m are over water the game draws. The rest run on dry ground. How can a track be inside a lake and on dry land at once?

A cut across the east shore of Girpilio ezeras where the old track ran deepest, 43 m inside the outline. The lake the game draws stops at the dark blue. Behind it the ground dips as much as 1.9 m below the lake's level and stays below it for 96 m. The store counts that strip as lake, and so does physics: water at 12.35 m would pour straight in. The old track (red, version 38) ran down there, on ground 1.2 m below the lake. The re-laid one (green, as laid in version 39) stands 160 m from it, inland, 1.7 m above the April high water. The lake bed under the dark blue is not measured; its depth is drawn nominal. A profile cut from the game's terrain, not a render.

The store traces each lake at its true level over the finished ground. The game's water was drawn from coarser data. At Girpilio ezeras the two agree on 98 percent of the water, at Himinvatn on 95. The rest are low strips like this one, below the lake's level but drawn dry. At Alksniai one of them is about 130 m wide, a flat at 10.5 to 11.8 m beside a lake at 12.35 m. The old tracks ran on those strips because, on the router's coarser map, they were flat dry shore.

Wait, does that mean the game's lakes are too small? Slightly, yes. That goes back to the terrain session, with numbers attached. For the roads it settles the question of which shore to obey. We obey the store's.

Twenty metres back#

The whole fix fits in one sentence. A way keeps its distance from a lake's outline: 25 m for a stone road, 20 m for a cart track, 12 m for a footpath.

On the five tracks, every stretch that came inside that distance was cut out and laid again. The new line starts and ends on the old one, where the old one is clear of the setback by 10 m, and it joins it smoothly, without a kink. Between those points the new line is found the way the router finds any road: it pays for length, for climb, and for any grade steeper than 6 percent, and it is never steeper than the class allows. Every metre within 12 m of another road costs it six times over, so it never runs along one. Everything outside the re-laid stretches stays exactly as it was: 12,099 of the five tracks' 18,121 points did not move.

The rule bends a little where the ground is tight. The planner accepts a line that keeps nine tenths of the setback, and one stretch, north-west of Laukenai, comes within 18.6 m of the outline. The other seven keep their 20.

The longest re-lay, as laid in version 39: 2 km of the east shore track, which ran 686 m inside the lake. Now it keeps at least 21.8 m off the shore and crosses the river a little upstream of its mouth. The pale blue tongue up the river is the April flood. A plan view on the game's terrain; the lit render is not made yet.

The five tracks came out 137 m longer in all, 18,245 m against 18,108 m. The steepest new pitch is 14.5 percent, under a cart track's 18.

Below Girpilis (1,914 people), the town on the west shore, as laid in version 39. The track to Laukenai left the town along the water. Now it climbs a little behind the shore. The town itself stays where it is: its front on the lake is its own business. A plan view on the game's terrain; the lit render is not made yet.

Hug the shore or keep your feet dry?#

The rule left one real choice. Girpilio ezeras is a snowmelt lake: it peaks at the April thaw, 1.6 m over its lip, and a lot of the shore lies under that.

We tried both answers. The dry version priced flooded ground two and a half times dearer, and the tracks fled uphill. One stretch grew 30 percent longer, and they still ran 2.7 km on ground under the high water plus half a metre. The shore version priced it at 1.25 times: 6.2 km of new track instead of 6.6, within sight of the water most of the way, with 3.3 km in the flood band.

We chose the shore. The old tracks sat in that same band, so nothing got worse, and a lakeshore road should look at the lake. A low bank to carry the tracks over the April water, a causeway with culverts where the streams cross, is the natural next step. It is written down as a proposal, not built: the first version stays small.

What a ferry has to beat#

Breitmeer's ferry, both ends: in the road record long before the lake had an outline, and nothing here changes. It was laid by hand, and the outline agrees with it. Note the pale blue on the left: at the March thaw, by the store's reckoning, the flat behind Grunwald's landing stands in shallow water. That reckoning errs high, by design. Plan views on the game's terrain.

The road network already has a rule for when a long way round deserves a shortcut. When it linked neighbouring villages, it built a new road only if the old way by road was more than 1.5 times the new route. Ferries get the same test. Two villages on one lake get a ferry when the way round is over 1.5 times the way across.

From Girpilis to Alksniai, on the far south-east shore, the road runs 6.7 km round the south end. The ferry route is 4.1 km: 39 m from the town to its quay, 3.64 km over the water, and 451 m up from the far landing to the village. That is 1.62 times, so it passes.

The ferry, laid in version 39 of the road record, with its far landing where it stands now: where the store's shore meets the drawn water, a 451 m walk from Alksniai, not out on the flat. Each landing will be what the road builder makes of every ferry: a low stone mole across the shallows to where the water stands 0.4 m deep, then a 6 m timber jetty. The lake has 7.3 km of open water for the wind to work on, and the store calls its waves stormy. The ferryman will have opinions about April. A plan view on the game's terrain.

The far landing moved once. It was first put on the shore nearest Alksniai, and it landed on that flat below the waterline: lake by the store, dry land by the game. The router's own ferry marking would have run the boat aground there. So landings now go where the store's shore and the drawn water meet. The crossing got 210 m shorter, the walk up to Alksniai longer, and the ratio stayed 1.62.

A path down to the heaven lake#

Himinvatn means heaven lake. It lies at 1,622 m in the Snowbound North, 114 hectares of water that the store says is frozen all year: ice-white, turquoise beneath. Three villages stand round it, and none of them comes down to it. Their paths wind high around the lake. Djupadal and Ravntopp are 2.0 km apart as the raven flies and 7.9 km apart by road.

The path from Djupadal to Ravntopp, laid in version 39: 4.8 km, from Djupadal at 2,120 m down to the shore at 1,622 m, along the south-east end of the lake, and up to Ravntopp at 1,763 m: 586 m of descent and 231 m of climb in all, over the ups and downs between. Its steepest pitches reach 49 percent. Footpaths go up in steps cut into the slope past 30 percent, so expect stairs. The two grey dashed lines were tried too and failed the test: Nordsund to Ravntopp would save only 1.15 times, Nordsund to Djupadal nothing. A plan view on the game's terrain.

It passes the same 1.5 test, at 1.63. A lake gets one such path at most, for the pair that gains most, and only where the shore earns it. Two paths along one shore would end up riding each other, and the network has a rule against that too.

Picture the walk from Djupadal: some 500 m of switchbacks down to a lake that never thaws, a stretch along its rim of ice, then the climb to Ravntopp, 3 km shorter than the high way round. Whether anyone skates is up to the people who live there.

Laid, and not yet#

Laid: in version 39 of the road record, the five tracks round Girpilio ezeras re-laid, the ferry from Girpilis to Alksniai, its water open from landing to landing, and the path down to Himinvatn, checked before they landed: every one of the 900 settlements still connected.
Not yet: lit renders of all of it, each view judged by a blind critic before it is shown; every figure above is a plan view or a profile on the game's own terrain. The low bank over the April water is a proposal. The planned lakes, and the roads across them, wait for the water.

Hollows that wait for water#

The three lakes in this post were the easy ones: the water was already there. The planned lakes lie in hollows that are dry today, and on the store's first outlines 21 km of road runs across them. One town, Binh Ninh, will stand in its lake on piles, a stilt village, and its stone road crosses the floor that will become Ho Lang. Saat Sar steps down the lava from 309 m to 17 m, eighteen spring-fed lakes and pools, and the only road to touch it is the Hundred Arches, some 29 m over Ansoosar, its lowest pool.

When those lakes fill, the roads will have to choose again, shore by shore: hug or climb, go round or take the boat. Which of them will earn a ferry, and who will be waiting at the landing when they do?

2,585 Hollows and Three Lakes

2026-10-19T00:00:00Z

The island's map shows 2,585 hollows that could hold a lake, and three lakes. Walk onto any of the twenty-two floors in its hills, Binh Ninh's among them, and the ground lies level for a kilometre or two, as flat as water, with no water on it. One of them now holds a lake: Ho Lang, at the head of this page, on the floor where Binh Ninh stands, its water level with the sill at its outlet.

Every place on the island where water could stand, coloured by what it became: lakes blue, ponds green, dry salt flats pale, no lake grey. Only the three largest blue shapes, in the north, and Ho Lang hold water in the game today. A red ring marks a village standing in the water of its lake. (The lake store's second set, traced on 4 m ground, 2026-10-19.)

A lake bed with no lake#

The island's large relief was not drawn. It was grown: uplift raises the mountain belts, rivers cut them down, and a simulation balances the two over a long run of steps on a grid of nodes 64 m apart. At every step, water finds its way from each node to the sea. Where it cannot, it stands, filled to the height of its outlet.

Then the simulation does what a real lake does to itself. Sediment settles in still water, and at every step the floor of a flooded basin rises by a share of its depth. At the end of the run it tidies: any lake left shallower than 4 m, or smaller than 0.6 km², is filled to its outlet. Three lakes passed. Breitmeer covers 123 km² and stands 62 cm above the sea. Girpilio ezeras covers 22 km², 12 m up. Himinvatn covers 1.1 km² at 1,622 m, in the north-east mountains: the only lake that survived far above the sea.

The twenty-two floors are what the rest became. Twenty-one lie level to within 0.7–0.8 m, a standard deviation, across up to 5.8 km²; the last, 1,706 m up, to 1.2 m. On the 4 m ground, twelve of the sixteen we measured lie within half a metre of the height at which their basin overflows: the mark of a lake filled to its brim. The map records no water on any of them.

2,585, counted from a map#

To find every other place water could stand, we flooded the island from the sea on a 32 m map. The flood fills each hollow to its spill level, and it found 2,585 hollows of at least a hectare and a metre deep. Half are smaller than 1.6 ha; the deepest one per cent go down more than 60 m. The 251 of at least 5 ha and 3 m we traced again on the finished ground, sampled every 4 m.

A quarter of the largest vanished, nearly all of them on rivers. Of the 251, 188 still hold a hollow of a hectare and a metre: 149 of the 157 away from rivers, but only 39 of the 94 with a river running through. At 32 m a river's channel is narrower than a cell, so the map reads the river bed as a dam, and a hollow forms behind it. At 4 m the channel is there, and the water runs out. One hollow that promised 20 ha of lake at 32 m holds 1.9 ha at 4 m, its outlet 4.3 m lower.

One hollow at two grains. Red: the 20 ha lake the 32 m map promised. Blue: where the 4 m ground holds water 30 cm deep or more. Most of the promise drains through a channel that runs along the outline's spine.

Walls that trap their own rivers#

The four largest hollows are the strangest. A sea cliff on this island is raised by the height of its wall, 500 m at the middle of the Weeping Wall, and the raise tapers away over the next few kilometres inland. Behind a tall wall, the land is left lower than the lip: a closed bowl. Behind the Weeping Wall the bowl covers 27 km² and is 136 m deep, and twelve villages stand on its floor. Its river climbs the last 650 m to the lip, then falls 170 m into the sea.

Three of the bowls are not to be lakes: filled, they would drown 24 villages and some 13,000 people. The fourth, a small one in the east, may be one day: a lake held at 75 m behind a notch cut through the sea wall, falling 75 m into the sea, with Karang Pinang on its shore. It is still a proposal.

The rain decides#

A hollow is a place water could stand. Whether it does is a question of water. Each was weighed against the rain over its catchment, scaled by its realm's wetness, and against the evaporation off its own surface.

The Sandstone Desert and the Lava Tip get no rain at all. 60 of their hollows are dry salt flats, white crusts that never hold water.
A lake that loses more to the air than its rivers bring shrinks and turns salt.
A small or shallow one is a pond, or a fen.
Every lake that remains overflows all year. In the monsoon east, Ho Lang rises half a metre in the rains. In the north, Hatunqucha lies under ice for three months and peaks at the March thaw.

Each lake also has a colour, by the rules real lakes keep: turquoise under snow, tea-brown where peat drains in, jade-green in the warm east, milky where clay washes in, acid green over the black sand of the lava.

A name older than its count#

The Lava Tip has no rain, so its one long river is a dry wash. The plan gives it a chain of lakes anyway, fed by warm springs that rise along the lava. Its name is older than any count of it: Saat Sar, the seven lakes. On the 32 m map the stair held eight lakes, each at least 2 m below the one above; at 4 m it holds eighteen waters, five lakes and thirteen pools.

The names run from the moon at the source to tears at the sea. Chandsar, the moon lake, is a lava pit 60 m deep at 309 m. Sheeshasar, the mirror lake, is the largest, 45 ha. Ansoosar, the lake of tears, the last before the sea, turned out to be a pool of one hectare. The chain kept its name.

Its nearest real kin is Band-e Amir, in Afghanistan's Hindu Kush: six lakes behind dams of travertine, the mineral that springs lay down.

Saat Sar from above: the named waters in green with their levels, the unnamed pools in blue. North is up; the sea lies just past Ansoosar. From Chandsar to Ansoosar the stair falls 292 m.

Lakes above their rims#

The planned lakes went wrong once more before they were right. The first plan stood the water two metres over each floor. On the whole-island map, eight of the twenty-two basins have a rim lower than that. On the 4 m ground most of the floors are not closed at all: they drain through channels the 32 m map cannot see. Our sampling boxes had cut the outlines short, and two lakes, Ho Lang and Ho Guong, met through the saddle Binh Ninh stands on.

So each planned lake now stands at its planned depth, never above its basin's rim, and every narrow neck it would drain through is recorded as an outlet to build. An outlet can take one of three forms: a fall over a sill, a wide open outlet, or two channels about an island. An island like that is a site for a fortress, as Trakai's castle stands on its island in Lake Galvė.

The planned lakes, the first plan in red and the corrected one in blue, both traced on 4 m ground. The red outlines' straight edges are our sampling boxes, not shores.

Held within their basins, the lakes reached further than the first plan, and three more villages stood in their water. Sternsee posed a choice: kept small, 71 ha of reeds below a dry Tiefenberg, or let fill to 756 ha with Tiefenberg's bay in the water.

Sternsee held two metres below Tiefenberg (left) and at its planned level (right). The red circle is 60 m across the village's middle.

It was let fill. Tiefenberg, Sandwinkel, Nusskirchen and Rotenenheim become lake villages on piles, beside Binh Ninh and Waynaqullqa: six in all. Their real kin are the prehistoric pile dwellings around the Alps. UNESCO lists 111 of them, nine on the shores of Lake Constance in Baden-Württemberg.

A name for every water#

Each lake has a name in the tongue of the villages around it, weighted by their people and their nearness: the tongues of twelve cultures in all. Some names say what a lake measures, some are poetic, and a few borrow a village's name. Each language's grammar had to come out right:

Name	Language	Meaning
Sternsee	Germanic	star lake, where, the villages say, a star once fell
Ho Lang	Vietnamese	the still lake
Hatunqucha	Quechua	the great lake
Sabkhat al-Sarab	Arabic	the salt flat of the mirage
Petit Lac	Norman	the little lake (petit goes before the noun)
Girpilio ezeras	Lithuanian	Girpilis's lake (the village's name in the genitive)
Himinvatn	Nordic	heaven lake

A name keeps to its place. When a lake was traced again on finer ground, it found its name by the point the name was pinned to. That is how the chain's eight names held as the chain grew to eighteen waters.

What runs, and what is still water on paper#

Runs: the lake store traces 294 outlines on 4 m ground and keeps them, with their levels, their low and high water, their outlets, colours and names. The three lakes the island already had are in the game.
Runs, since 2026-10-21: Ho Lang, the first planned lake built into the terrain itself (Atlas 8c59aabb). Its water stands on the finished ground at the height of its outlet's sill, and the sill is built across the outlet's neck. The lake's low and high water are in the lake store; the game does not keep the seasons yet, so its water stands at one level, the sill's.
An experiment: the judging. Which hollow becomes a lake, a pond, a salt flat or nothing is a set of rules, each tested, that we keep refining. Where a village stands in the water, the team decides.
Planned: every other lake in this post. The other planned lakes, the Saat Sar chain and the lakes in natural hollows are not in the game yet. Binh Ninh's ground is under Ho Lang's water, but its houses on piles are still to be built, and none of the other five pile villages stands in water yet. The roads will bend round each lake as it is built, never the reverse, as they already do at Girpilio ezeras.

A floor that became a lake#

Binh Ninh's 2,370 people live on a floor that has become a lake. Ho Lang's water stands 1.1 m deep at the village's middle, and half a metre deeper in the rains.

Firewood for Winter, in a Word the Plans Didn't Know

2026-10-19T00:00:00Z

At the edge of Lichtenloh, in the weeks its fields let its people go, a crew of three works a sawpit, and the woodpile beside it grows. Thirty-nine crews go out from the village, 1,705 from across the island, and in the southern desert 134 pairs of gatherers drive their donkeys out for brush. The wood is not for home. It is cut from what the woods have left over, for buyers the island's trade finds, and from the east much of it goes west by sea.

They come from 318 of the island's 900 settlements, a third of it: the north, the rain-forest east, the desert south. Until the morning of 19 October, no plan of any of them had a word for the crews. Without the word there would have been no plan at all.

The 318 settlements on the island's paper relief (terrain aeead1e), from the economy's counts at trade run 15. Brown: the winter crews, 1,705 in 276 settlements, a dot's area by its crews. Ochre: the desert's brush gatherers, 134 pairs in 42. Grey rings: the other 582. West of a line 20 km west of the island's centre stand 335 settlements and 3 crews. Ardbach and Kiltarn, in white, keep none; their firewood comes by sea.

In the slack weeks at Lichtenloh#

Lichtenloh is a Germanic village of 649 people in 166 households, a nucleated village a little north of the island's middle. Each of its 39 crews is a farmer and two labourers: 117 people, near one in five of the village. There is never more than one crew for every three farm hands. They cut only in the woods left over once the village's own people are served, and only for buyers the island's trade finds them. A crew needs a woodpile and a sawpit by the village, not a camp in the forest.

It was not always so. In the counts of 16 October, Lichtenloh had 891 people and five woodcutters' camps, out in the 1,005 hectares of woods the counts asked for it: woodcutters by trade, living with their work. Two days later the island's land budget settled fewer people on the land, Lichtenloh among them. In the counts now it has 649, and no camps. It farms 16 hectares and buys its grain, and its 58 carters' yards carry the grain in and the firewood out. A village of woods and carts.

The two are not the same thing counted twice. A camp is three woodcutters and a carrier, out all year, and cuts about 930 tonnes of firewood a year if it cuts nothing else. A crew is a farmer and two labourers, out only in the slack weeks, and cuts 240 tonnes in a winter. One camp cuts about as much as four crews. Lichtenloh's 39 crews cut as much as some ten camps would.

The plan of 16 October found room for none of the five camps. A camp needs 150 hectares of woods of its own, and the plan reported "not enough land in the settlement's territory and terrain". The camps were counted, and stood nowhere.

Lichtenloh, laid twice on the same terrain window by the same generator (dc8149a, roads mycelium v37): above, from the counts of 16 October; below, from trade run 15. Left, the area; right, the red frame, 160 m across. Brown: a winter crew's lot, 10 by 6 metres, for its woodpile and sawpit, 39 laid of 39. Above, the five camps were laid nowhere. The yellow fields went too: on the island's measured land, Lichtenloh's own territory, 452 hectares, holds 18 of plough, so it farms 16 and buys its grain. The lots are drawn as the plan lays them, plain; the woodpiles and sawpits in them are not drawn yet.

Fifty-seven crews from 155 households#

The most crews go out from Naatamaki, a Finnic-Baltic palisaded village in the north-east: 57 crews, 171 people.

Naatamaki laid twice, as Lichtenloh above. From the counts of 16 October, 211 people; from trade run 15, 1,063. Its 6,264 hectares hold 878 of fields, far more than 211 people farm, and the land budget settles the hands to farm them. Of its 57 crews' lots the plan laid 6. For the other 51 it reports "no free ground round the town": the lots stand only at a town's edge, and this town's edge is full. The village has 155 households.

In the southern desert the fire is fed another way. The peasants and townsfolk burn brush: scrub, thorn, tamarisk and palm fronds, gathered on the wild land by two gatherers and their donkeys, for the towns. Tighza n'Tifawin, an Amazigh ksar of 1,320 people, keeps 51 pairs. In the counts of 16 October it had seven woodcutters' camps and a charcoal kiln.

Tighza n'Tifawin laid twice, as above. From the counts of 16 October: 1,839 people, seven woodcutters' camps laid nowhere (no room, as at Lichtenloh) and a charcoal kiln, in black. From trade run 15: 1,320 people and 51 pairs of brush gatherers, 41 of their lots laid, 8 by 6 metres each, in ochre. For the other 10, "no free ground round the town".

Wood for the west#

The crews are thickest where the woods are fullest and the hands are many. Within a day's walk of the settlements more than 20 km east of the island's centre, 21,766 hectares of woods are left over after their own people; for those more than 5 km west of it, 4,135. The rain-forest east sends 515 crews from 75 Indonesian settlements, more than any other people.

The west, the 465 settlements more than 5 km west of the centre, has 13,523 hectares of woods within reach, and its own people use 11,015 of them. Little is left to cut for anyone else, so the west buys: from the middle of the island, from the east, and 14,788 tonnes a year from abroad.

Muara Putih is a stilt village of 93 people at a river mouth in the east. Every year 21,000 tonnes of firewood are loaded onto sailboats at its landing, most of it for Kiltarn and Ardbach, cities of 10,055 and 12,000 people some 72 to 75 km west. About 18,000 tonnes of it come down to the landing by pack train from villages inland, Tanah Sari alone some 10,400. The village's own 11 crews cut about 2,600 tonnes in a winter, a crew some 240, much of it for Karang Makmur.

The economy calls the season the winter weeks, everywhere. In the rain-forest east they are the slack between two rice harvests.

Two words the plans did not know#

Every settlement's plan is laid by the town generator from the economy's counts. For each settlement the counts are a line of words and numbers. Lichtenloh's ends winter_cutters: 39; Teluk Tinggi's has fishing: 20 and winter_cutters: 4. Each word names something the generator lays, exactly as many as the number says. It knew 162 words. winter_cutters was not one of them, nor was scrub_gatherers.

A word it does not know stops the plan. The whole settlement is refused, not just the word. That is on purpose. Every crew the economy lists is in its sums: the firewood they cut warms someone's house. A word skipped quietly would be 39 crews and nowhere to stack their wood, and no one would know. A refusal is loud, and it names the word.

Why did nobody hear it?#

The generator's tests and its example settlements all read its own copy of the counts. That copy was taken at 02:10 on 16 October. The brush gatherers came into the counts that evening, with the forest asks. The winter crews came at 00:22 on 18 October, with the land budget. On the old copy every line read cleanly, every test passed and every example was laid. The island's settled population had not yet been laid from the new counts.

The gap showed on the morning of 19 October, before the first lay on new counts: the water towns' houses over the water. The new counts' words, checked against the old copy's, held four new ones. Two the generator already knew. Two it did not.

What a refusal would have cost#

318 settlements of the island's 900, refused whole: 140,695 of the 480,809 people the counts settle.
Three of the seven water towns. Teluk Tinggi, Telaga Baru and Wassermark were to be laid that night with their households over the water. Each also sends winter crews.
The next step. Laying the whole island's settled population would have failed at every such line.

Nothing published was touched. Every example settlement pinned so far was laid from the old copy, and still is: they show the island as it was before the land budget.

A woodpile, a sawpit, a tether#

Since commit dc8149a (19 October) both words are the generator's. Each winter crew gets a lot of 10 by 6 metres at the village's edge, for its woodpile and its sawpit. Each pair of brush gatherers gets 8 by 6 metres, for their stacks of brush and their donkeys' tether. In the world's vocabulary they are a woodcutters' camp and a harvest camp.

The economy's next counts, trade run 16, bring words of their own: the wayfarers' rounds (pedlars, tinkers, minstrels, pilgrims and beggars, who sleep in inns, almshouses and barns and keep no building) and the hectares each village's holdings farm. Those are read already, before they arrive. One more word could come in any run: bargemen, a river barge and its two boatmen at a landing. No line carries it yet, and the plans cannot read it yet either.

What runs, and what does not yet#

Runs, in the economy: the winter crews since the land budget of 18 October, the brush gatherers since the forest asks of 16 October.
Laid, in the plans: since dc8149a (19 October), a lot for every crew and every pair, at the town's edge. Lichtenloh has all 39; Tighza n'Tifawin 41 of 51; Naatamaki 6 of 57. Where the edge is full the plan lays the rest nowhere and says so. Letting the lots stand further out is the next step, not yet made. The generator's full test suite passed on that commit on 19 October (06:59 to 08:05: formatting, lints and all 35 test suites, with tests for both words); it is on the generator's main.
Not yet drawn: the lots are plain lots on the plan. The woodpiles, sawpits, brush stacks and tethers in them are not drawn yet.
Not yet laid: turf. Where the woods within a day's walk cannot give the households' firewood, they burn peat, dung cakes and straw. The counts carry no word for it, and the plans show none yet.
Not in the game: the game does not lay the island's towns yet. These are the generator's plans.

Between two rice harvests, eleven crews go out from Muara Putih's houses on their piles. Pack trains come down from the villages inland. At the landing the sailboats load, and leave westward.

Forty-Five Metres Over the Rainbow Gate

2026-10-18T00:00:00Z

The camera crosses the Rainbow Gate 45 metres above the middle of its crown, 1,250 metres over the sea, and the arch's stones go by underneath in a fraction of a second. It got there by a flight planned whole, from its first metre to its last, before a single frame was drawn: a camera that simply follows the ground, the way a player's does, cannot fly like that.

This post is about three flights over the island and how we learned to plan them: out of a slot canyon into a low sun, over the crown of a colossal arch, and along the Hundred Arches to the Black Shard. It took two wrong takes, a critic and a plan that would have filmed a wall to find out what a camera flight needs from a world.

Watch the film Open the film

The pass over the Rainbow Gate, 20 seconds (film.rainbow-pass). The camera comes in from the north-east at 950 m, rises through 1,250 m over the crown and sinks to 650 m, 1.4 km beyond it. Its gaze holds the crown until the crown passes under it, then turns back over the bay's west shore to the whole arch, face on. The Gate's crown is a block of stone 27 m square. For the second the camera is over it, the arch's feet stand 1,210 m below it, on two bluffs 895 m apart. The Gate's builder is still another story.

A camera that hops#

The camera a player holds in Atlas orbits a point on the ground. As it moves, that point slides over the land, so every ridge it crosses lifts the camera and every gully drops it. And whatever the camera would dip into, it is pushed up out of, a few metres above the ground under it. In play that is right: the ground is the subject, and the camera keeps it in view. On film it looks like a machine following a line.

A filmed flight has to know what is coming. It rises before a ridge, not as the ridge passes under it; it turns before the canyon bends. So we stopped steering the camera frame by frame. A flight is now planned once, over the real ground, and then every frame of it is rendered.

The three flights on the island (left, the paper relief at Atlas aeead1e) and over their own ground (right, the finished ground sampled every 8 to 16 m). Red is the camera's track through its keys, each with its height above the sea; the crosses are the points it looks at; the arrow its last gaze. In A the track is only 160 m long: almost all of that flight is a climb. In C the track runs beside the bridge, never nearer than 63 m.

Take one hung level over a valley#

The plan's first idea was a string. Take the highest ground within reach of the track, add the clearance, and lay a string over the top: let it settle smoothly, never below. We smoothed and raised it sixty times. The first film was perfectly smooth, and wrong. Over a valley 3 km wide the camera never came down: it hung level, 720 m above the valley floor.

A string pulled taut over a row of crests cannot sag between them. Smooth it and raise it often enough and it settles on the crests' upper hull, which is flat across every valley. So take two does it the other way round. It blurs the land under the track into a gentle line that does dip into the valley. Then, wherever that line passes too close to the ground, it adds back only as much height as it needs, blurred, so the lift has no corners.

The same 3.2 km track up the great forest's valley, the first two takes. The dashed line is take one, flat at 1,000 m. The orange line is take two: it sinks to about 920 m over the valley, never closer to the land than the envelope (the highest ground within 300 m, plus 300 m). With 300 m of clearance there is little room to sink. In the slot canyon below, the clearance is 15 m.

Out of a slot canyon, into a low sun#

Watch the film Open the film

Rising out of a slot canyon of the Jungle Canyons, 18 seconds (film.canyon-rise). It starts 81 m above the canyon floor, looking down the slot to the sea through the canyon's mouth, with the sun behind the east wall. It climbs, the sun clears the east rim 8.4 seconds in, and the camera turns 52° to the left into it, coming out at 700 m in the canyon's mouth. Under the sun, on the horizon, stands the Black Shard, 18.8 km away. The lava it stands on, at most 99 m high, is seen from here almost edge on: a strip 0.3° tall, five pixels of the frame, as dark as the sea around it, so the Shard seems to stand on the water.

The canyon is a deep canyon cut down to the sea on the rainforest east's coast. Where the flight starts it is 160 m wide at its floor, 260 m from rim to rim, and 690 m deep. This flight could not follow the ground at all. It is a tween between three keys: a place and a height for the camera, and a point to look at, for each. The camera passes through its heights on a smooth curve that eases out of the first and into the last.

Across the slot canyon where the flight starts, 21.3 km south of the island's middle, to scale. The camera starts in it, 110 m above the sea, and ends 590 m higher and 160 m further south, in the canyon's mouth.

The first cut of this flight stared into the canyon wall for two seconds. Its gaze moved from one look-at point to the next in a straight line, and a turn of more than a right angle sends that line close past the camera, so the view swings across whatever stands between. The gaze now turns by its bearing and its tilt, the short way round, and still lands on each key exactly.

Looking before rendering#

A finished film is about 500 frames, and each one waits for the ground around the camera to be built before it is drawn: about twenty minutes of the machine for one film, and on a busy night an hour or two in the queue for it. So we look at a plan before we render it.

We rendered the first cuts small and showed them to a critic who had never seen the island. Besides the stare at the wall, the canyon film revealed the sea all at once and ended so high that the canyon was lost behind it. After the Gate, the film had no subject. The arches flew under a flat midday sky. We planned the three flights again from those notes.

Then, before rendering them, we cast each frame of each new plan as rays over the island's own ground: no light, no trees, no water, only what stands in the way. The canyon's new plan would have kept a wall of rock in the middle of the frame for eleven of its eighteen seconds, and ended looking out through a narrow gap, with the sun behind the east wall. The canyon runs south to the sea, and the sun stood 31° east of that line. We had planned the ending on the map, not from inside the canyon.

The canyon flight cast as rays over the finished ground, at six moments of its 18 seconds: each pixel is what its ray meets first, rock (darker nearer), sea or sky; the ring is the sun, yellow where it can be seen and red behind rock. Top: the plan after the critic. Bottom: the plan we filmed.

The same look caught two faults in the other plans. The Gate's gaze tipped down toward the far side before the camera had climbed level with the crown, so the crown slipped out of the top of the frame for two seconds just before the crossing. It then turned back nearly 150° in four seconds, too fast to read at 24 frames a second. It now tips down only 28° as it crosses, and turns back at most 26° a second. The arches plan flew 3.3 km in 20 seconds, past arches 62 m away: a flicker, not a glide. It now flies 2.1 km in 24 seconds.

Along the Hundred Arches to the Black Shard#

Watch the film Open the film

Along the Hundred Arches to the Black Shard, 24 seconds (film.arches-to-shard, version 2). It opens 75 m up on the bridge's own axis at its mainland end, the deck a line leading to the Shard 4.5 km away. Then it comes down beside the arcade, 46 to 81 m above the sea and never nearer to it than 34 m, its gaze on the bridge 900 to 1,380 m ahead. At the end it lifts its gaze to the Shard, 2.4 km away, with the arches leading in to it from the lower right.

The bridge runs 2.8 km across the lava field of the Lava Tip, from the mainland to the lobe where the Shard, a basalt monolith, stands: a hundred arches, nine towers 14 m tall, a deck between 3 and 61 m above the sea. The camera flies about level with the deck, beside it rather than over it: seen from above, a deck hides its own arches. We filmed it at a sun 7° above the horizon, under more cloud than the other two, so the lava can show its own light.

The first final of this flight ended 110 m up, 400 m short of the bridge's end. There the deck comes down to meet the ground, so its last arches sank out of sight, and from that height the glowing cracks in the lava faded into a dark plain: a slab on an empty table, the very thing the critic had warned of. The film already held a better ending, two seconds earlier, where the arcade leads in from the right toward the Shard's foot. We filmed it again to end there.

Then the critic saw each final beside its first cut, blind to which was which. It chose the canyon's final clearly and the Gate's decisively, but this film's first cut, slightly: that cut opened on the bridge's own axis, the line of arches leading straight to the Shard, and it called that the best frame of either. So we filmed it a third time, opening on that axis and then coming down beside the arcade: that is the film here. Shown blind beside the second take, it won clearly, though from 75 m up the bridge in its first two seconds still reads as a plank, its arches hidden.

What runs#

Runs: planning a flight over the real ground, by track and clearance or by keys, and filming it as one finished render a frame, at 1920 × 1080 and 24 frames a second. Each frame waits until the ground around the camera has been built, then ten more frames. With six, one frame in 385 caught the sky's light being changed and flashed hazy; with ten, one in about 1,400 still did, 15.2 seconds into the arches film's second take; that one frame has since been rendered again.
Runs: the films live in the render store beside the island's other renders, each with its flight: the camera of every frame, its light and its size, which films it again frame for frame, and the command that plans it again over the land as it is by then.
Runs: the ray-cast look at a plan before it is rendered is now a step of the glide: every flight is cast over the ground it sees, and the glide warns of a wall in mid-frame, of frames mostly rock and of a gaze turning faster than 25° a second, and says when the sun is in sight, before a frame is drawn (source/raycast.py is this post's copy).
Not yet: the camera does not bank into its turns, and the clouds hold still: our offline renders stop the clock, so nothing in the sky drifts during a film.
Runs: the camera a player holds glides too. As it pans, it looks 1.2 seconds ahead along its way and rises before a ridge instead of hopping with the ground under it; at rest it stands on the ground as before. A click on the land turns the gaze to that point.
Open: the places themselves. The critic found the Gate's stones as flat as tiles, the Shard bronze rather than black and its lava painted on in stripes, the canyon's walls banded like sandstone and without mist, and the bridge's arches knee-high. Those are the island's next work, one place at a time.

The Shard under the sun#

From the canyon's mouth the sun stood 149° round from north, and the Black Shard 154°: under the sun's path, 18.8 km out, beyond 16 km of open sea and the black lava of the Lava Tip.

Who stands on that shore at this hour, and what do they make of it?

How Does a Port 209 m Above the Sea Load Its Ships?

2026-10-18T00:00:00Z

At Labuhan Harjo, a river slides over the edge of a sea wall and falls some 170 metres into the head of a sea canyon. Beside it stands a port city of 10,868 people, with 31 ships, four galleys, 152 fishing boats and about 18,900 tonnes a year to load.

The city stands 209 metres up. At the foot of the wall there is no shelf of land wide enough for a quay, only fallen rock; but off it lie some ten hectares of sheltered shallows, half a metre to four metres deep. This is a post about a procedural island's sea cliffs, the rivers that fall off them, and a game economy that has not yet looked down. We only saw it because we taught the island's maps to show how high a coast is.

The Weeping Wall from straight above, about 10 km across, at 8 m a pixel, drawn from the same terrain as the island map (land identity 8edf4a5515b0be6b). Both blues on the sea are deliberate exaggerations. The long one across the bay is the wall's shadow under a low sun, with the relief doubled. The darker strip along the shore is the coast's height, more on that below. The blue lines are the rivers our trade survey measured. Follow the short one at the top right: it runs south-east from Kuala Baru and ends at the cliff edge beside Labuhan Harjo. Muara Putih sits at the south-west end, where the wall runs out, its feet in the water.

A coast you can only see by its shadow#

From straight above, a 500-metre cliff is a line one pixel wide. Our first whole-island maps, drawn from the game's own terrain, made every coast look alike: a cove with a beach and the Storm Coast's walls of 600 to 700 metres, twice the height of the Eiffel Tower, were the same thin edge between paper and sea.

So the maps cheat, on purpose. Every coast throws a shadow onto the water: 20 pixels of sea for every 1,000 metres of height on the 2,400-pixel island map, so a 100-metre bluff gets two pixels and a 700-metre wall on the Storm Coast fourteen. A single sun would only darken the coasts facing away from it. Our first try did exactly that, and the cliffs facing the light vanished. So the shadow is cast from eight directions at once: the sun's side at full strength, the others between 60 and 88 percent. No real light does this. It is a cartographer's lie that tells the truth about height.

The whole island (map.island-paper, drawn from the terrain at Atlas commit aeead1e). The red box is the close-up above; the red dot is Yawarchaka on the Storm Coast. Look for the wide dark bands on the sea: those are the island's tallest walls.

Two of those walls do something no other coast on the island does.

The Weeping Wall never let its rivers down#

The Weeping Wall is a wall of soft claystone on the south-east bay, where the rain forest thins into temperate country. It rises to about 540 metres in the middle of the bay and stands about 250 to 270 metres where Labuhan Harjo looks over it. The coast rose faster than its rivers could cut down, so none of them ever carved a canyon. Each keeps its floor at the lip and falls from there straight into the sea: a dozen ribbon falls down the face, a hanging valley at the top of every one. The design asks for more: rainbows in the spray in the afternoon, when the sun stands behind anyone looking at it from the bay, and mist forest on the rim where the spray lands. Neither is in the world yet.

One of the Weeping Wall's falls, rendered in the engine (render.inv-03-weeping-wall-fall-high, Atlas commit c65fcf7, 2026-10-17). The falls are still drawn as plain ribbons. The spray, the rainbows and the mist forest are planned, not built.

The island has 13 painted cliff coasts. Only two of them hang their rivers: the Weeping Wall, where every river falls from the very lip, and the Storm Coast in the south-west, where the walls stand 600 to 700 metres and every river leaves them partway up. Both came into the world on 3 October.

190 metres the river survey never saw#

In Cargo Only Floats Downhill, we measured every river the island's trade used. That survey reads each river where the world first lays it down: on the island before any cliff, canyon or plateau is built on top. On that land, the river from Kuala Baru runs gently down to a river mouth at sea level.

The finished island disagrees. When the blog's editor redrew the river post's maps on the new relief, the survey's water matched the ground within 10 metres at 4,817 of its 4,978 points. Five short reaches didn't. All five meet the sea at the Weeping Wall or the Storm Coast.

Two rivers, as the trade survey floated them (blue) and as the island stands today (grey). Near the coast, both stand about 190 m above the water the survey assumed. At the Weeping Wall (left) the ground drops to the sea in the last quarter-kilometre: that drop is the waterfall.

Under the trade's first rules, the old simulation floated 1,323 tonnes a year down the Kuala Baru river, all the way to the lip of a drop of some 170 metres. On the Storm Coast it sent 879 tonnes down to Yawarchaka, a Peruvian-inspired village of 308 that stands at 409 metres, about a hundred metres from where its river leaves the wall at some 190. The river post's new rules took that cargo off both rivers. They fall 19.5 and 14.7 metres in their steepest kilometre, and even rapids, under those rules, carry goods only up to 5 metres a kilometre, so the goods go by land. The rules were right for a reason nobody had seen yet.

So who loads the ships?#

Here is the plain answer, part by part.

What runs today. In the island's record, Labuhan Harjo is a stilt village. Our settlement generator sited it where a river met the sea, on the same cliff-less land the survey reads, on ground it wrote down as 3.29 metres above the sea. On 4 October the settlements became a fixed record: nothing moves or resizes a town after that. Today the same spot stands 209 metres up. The trade simulation knows that Labuhan Harjo is a port. It does not know how high anything stands. So it loads 18,874 tonnes a year onto ships at the city and unloads 10,685, and keeps 90 percent of the city's granary stock and nearly all its warehouse goods by the harbour. That harbour, as far as the terrain goes, would sit 209 metres below the streets, where the wall stands in the water.

Where the land comes down. Two places on this coast are at the water. A kilometre and a half east of the city, Teluk Tinggi, a stilt village of 542, stands three metres above the sea and 75 metres from it; its one approved concept shows houses on posts and a red flag on the jetty. The economy ships almost nothing there: 220 tonnes out a year, 440 in, one ship. Five kilometres south-west, where the wall runs out, the stilt hamlet of Muara Putih has 93 people, no ships of its own, and loads 22,999 tonnes a year onto other people's. Nine tenths of it is firewood. Pack trains bring it down from villages in the eastern rain forest, Tanah Sari, Kampung Asri and Pasir Putih the most, 136 trains on an average working day, and ships carry it west. Their first ports of call are Kiltarn, for some 7,500 tonnes of it a year, and Ardbach, for 6,100: the hamlet that ships the east's firewood west. In the river post, firewood left the island's rivers for carts, from 67,000 tonnes a year by water to 13,000, and Ardbach stood at the foot of the busiest river system. Now ships land firewood at Ardbach from the far side of the island. On the busiest night, 15 ships lie at Muara Putih's quay.

The model counts room, not hands: berths, places for pack trains, stores. It does not ask whether 93 people can load 22,999 tonnes, and a settlement does not grow to its traffic. The numbers say where the goods go, not yet who carries them.

Labuhan means anchorage.

What is decided, and what is open. On 18 October, seven water towns were chosen. Teluk Tinggi is one, with about 300 of its people to live on the water. Labuhan Harjo gains a quarter of about a thousand people in stilt houses on the water, with the harbour's warehouses on piles beside them. The way up is Tangga Labuhan, the Harbour Stair, at the sea canyon's mouth. Two stone stair bridges climb at 35 degrees, each on a single rampant arch: 102 metres from the water to a landing at 74 metres, then 120 metres back to one at 158. From there a flight of 299 steps, cut into the canyon's west wall, climbs to the rim at 209 metres, and a path of 113 metres leads into the town. It is not the most practical way up. It is the fantastical one. The stair is in the island's road record now; in the game its two spans still stand empty, waiting for their stones. There is no hoist and no cart road: people and pack animals will carry the harbour's goods up and down it, about 29,500 tonnes a year by sea, some 100 tonnes every working day. The land below the city has no room for any of it: along the wall's foot, the largest patch of low ground is about a hectare. The water does. Where the river falls it is seven metres deep, but within a kilometre of the city, about 13 hectares of sea lie half a metre to four metres deep within 250 metres of the shore, some ten of them sheltered. None of it is built yet, but the economy already counts the quarter's people: 205 households over the water, each keeping a boat, so some fourteen new boats a year, nearly all from the town's own carpenters. Exactly where it stands is for the town generator to lay, with a plan shown first.

This is not unprecedented. Bonifacio in Corsica stands on cliffs about 70 metres above the sea, and its harbour, la marine, lines the narrow shelf of the inlet below. Its most famous way down, the King of Aragon's Staircase, has 187 steps. Legend says besieging soldiers cut it in one night in 1420. More likely it is older and led to a spring.

One map, every coast's height#

The two maps behind this post are now on the ontology site, where the island's 900 settlements sit on the paper relief. There is also the same island in the tactical map's colours, map.island-biome: green lowland, the snowbound north in white and blue, the sandstone south, the black basalt of the Lava Tip. Every picture records which version of the terrain it was drawn from, and the tool that draws them refuses to draw from an older one. When the island changes, the maps know they are out of date.

Drawing the shadows also turned up a bug. Near the island's centre, a rectangle of ground about 6 by 12 kilometres has straight, ruler-cut edges in the relief. No real land does that. A fix is proposed and not yet started, so today's maps still show it.

A city of 10,868 and a hamlet of 93#

Stand on the Weeping Wall at Labuhan Harjo and look down: a river falling some 170 metres into the head of a sea canyon, the sea filling the canyon below it, and at the wall's foot only fallen rock. Look east: a kilometre and a half away, at Teluk Tinggi, the land comes down to the sea. Turn south-west. Five kilometres along the coast, pack trains loaded with firewood are coming down to a hamlet whose quay holds fifteen ships on a busy night.

Upriver, at the head of the falling river, sits the hamlet of Kuala Baru. Its name means new river mouth. It stands 167 metres up, a kilometre and a half from the lip where its river falls.

In the shallows below the city, the piles of a thousand houses and a harbour's warehouses are still to be driven. Every sack that leaves Labuhan Harjo by sea will go down 1,211 stone steps. Whose back will carry it?

The Holy City Has a Place Now

2026-10-18T00:00:00Z

Every Muslim grave on the island is now laid across the way to one city. In the eleven towns laid so far, all 2,294 graves keep to their town's line within a tenth of a degree, and the lines leave the map together, east-south-east, toward a shore no map of the island shows. Until this week the graves agreed on nothing: first they lay in churchyard rows, then along three lines set by how each town happened to be turned. Those directions are mended, and the graves stand in rows again, turned together to the city. Where the lines meet, the island does not say.

These are the generator's example towns, laid and pinned; the game itself does not lay the island's towns yet. It took four layers of the generator to get here: we misread the first, and a question about the third gave us the fourth.

The island on its paper relief (terrain aeead1e) and the sea to the east-south-east. The dots are 225 settlements: the island's 126 of the Islamic tradition and its 99 Indonesian villages, which the generator also gives mosques and buries the Muslim way. They are coloured by people: Turkish gold, Arabic red, Amazigh purple, Pakistani blue, Indonesian green. Each line runs from a settlement along its qibla, the way its dead face, and leaves the map. Seen from the island, the lines fan by 5.6°: Baranaara in the south-west looks along 103.1°, Rantau Bambu in the east along 108.7°. Eleven of them have been laid towards it so far. For the other 214, the line is the way their dead will face once they are.

Stand at the domed tomb in Casseia#

Casseia is an Amazigh city of 10,800 people on the south coast, in a canyon so closed that it needs no wall. Its burial ground holds 103 graves round a domed tomb, 7 m square. Every grave lies on the same line, 13.4° from north, laid so that the dead, on their right side, face 103.4°, a little south of east.

Ride 6.5 km west to Baranaara, an Arabic village of 766 perched on its slope. Its 56 graves lie at 13.1°. Ride on, and the line turns a little from one burial ground to the next: 13.9° at Qamareen, 14.3° at Farisiem, 15.5° at Yukarikale in the Turkish east.

A Muslim cemetery's graves are not laid at random. The dead are laid on their right side, facing the qibla, the direction of the holy city, so a whole cemetery agrees on one line. Neighbouring towns nearly agree with each other, because the city is far away and lies almost the same way from both. Nearly, not exactly: the turn from town to town is how far away the city is.

How we came to be measuring graves#

We did not set out to survey cemeteries. This week the island's town generator laid example towns for every kind of settlement on the island: 59 of them, each a full plan of its buildings, fields, roads and sacred places. A concept artist works from those plans, so that its concept art shows what the generator really builds.

It was the concept artist who stopped. Preparing the ksar of Qamareen, an Arabic fortified village in the sandstone desert, it found a chapel in the plan, and three stations that our table of names called calvaries: a Christian church and, it seemed, Christian wayside crosses, in an Arabic village. That sent us through every grave and shrine in all 59 plans.

A bearing, a line, and which way the dead face#

The numbers in this post are bearings on the game's own map: degrees clockwise from north, with x running east and z south. A grave is a rectangle, 2.2 m by about 1.0 m in a churchyard's rows and 1.9 m by 0.8 m in a Muslim burial ground. Its line is the bearing of its long side. A line has two ends, so 2° and 178° are the same line give or take 4°, and two lines can never be more than 90° apart.

Which way the dead face is a different thing. In Muslim burial the body lies on its right side with its face towards the qibla, so the grave runs across the qibla, at right angles to it.

Four layers under Farisiem#

Farisiem is an Arabic town of 1,221 in the sandstone south. We have its burial ground at four layers of the generator, each pinned so that it can be laid again exactly.

Farisiem's burial ground at the four layers, north up, 90 m square. The dashed arrow is Farisiem's way to the holy city, 104.3°, drawn on every layer for comparison, though only the last two knew it. First: churchyard rows round an 11 m by 7 m chapel. Second: a Muslim burial ground round its domed tomb, its graves along a direction that came from the town's drawing. Third: the same ground, its graves across the arrow. Fourth: the same ground in rows turned to the arrow, an aisle (tan) from the gate to the tomb.

Layer 1, the plans of earlier this week: 336 graves in rows round a chapel, each grave square to its plot, at 124.6°.

We read that line, and the other nine Islamic towns' lines, as ten attempts at a qibla that had come out ten ways, and wrote this post about it. Measured again, every one of the 1,589 graves in the ten towns was 2.2 m long, the size of the generator's churchyard rows, and nine of the ten burial grounds stood round an 11 m by 7 m chapel. The ten towns' dead were buried as Christians. The island's economy asks most towns for a cemetery, and that request chose a churchyard whatever people lived there. The fix came in the next layer: every people now keeps its own burial ground.

Layer 2: 190 graves scattered round a domed tomb, as the generator's Muslim burial ground should be, all at 53.1°. Now the lines were meant as a qibla, but the direction was fixed in the town's own drawing. Siting turns every town to fit its ground, and the direction turned with it. In these lays every turn was a multiple of 45°, so the ten towns could take only four lines, and they took three: 143.1°, 53.1° and 8.1°. Farisiem and Qamareen, Arabic neighbours 5.8 km apart, buried their dead 45° apart. And the graves ran along that direction, head to foot, where a Muslim grave lies across it.

Layer 3, laid on 18 October: 191 graves at 14.3°, across Farisiem's own way to the holy city.

Layer 4, laid on 22 October: the same ground, in rows. Looking at layer 3, the team asked a plain question. The graves had gone from neat rows to a loose scatter by turning; why not turn them together, and keep them packed? At layer 3 every grave already lay on one line. What scattered them was where they stood: a 3 m grid square to the town's drawing, a third of its places left empty and the rest nudged by up to 0.9 m. Each grave turned on its own spot, and the loose grid stayed.

Layer 4 turns the grid with the graves. They stand side by side along the qibla, 1.6 m apart, in rows 2.9 m apart, packed to the wall, about one place in ten left empty. The domed tomb stands square to the qibla on the same rows, and an aisle runs from a gate in the wall to its door, with no grave on it. Farisiem's ground, 2,814 m² inside its wall, now holds 473 graves: a grave to every 5.9 m², where layer 3 gave one to every 14.7 m². Their line is still 14.3°, across the qibla.

Town	People	Layer 1: churchyard rows	Layer 2: the drawing's direction	Layer 3: across the qibla	The qibla
Baranaara	Arabic	171.8°	8.1°	13.1°	103.1°
Tinaberkan	Amazigh	74.5°	8.1°	13.3°	103.3°
Casseia	Amazigh	122.0°	53.1°	13.4°	103.4°
Multankhana	Pakistani	135.1°	8.1°	13.6°	103.6°
Qamareen	Arabic	177.6°	8.1°	13.9°	103.9°
Farisiem	Arabic	124.6°	53.1°	14.3°	104.3°
Ksar n'Tidri	Amazigh	2.1°	143.1°	14.7°	104.7°
Aman Ifrane	Amazigh	122.3°	53.1°	14.8°	104.7°
Yukarikale	Turkish	113.0°	8.1°	15.5°	105.5°
Kizilpinar	Turkish	141.2°	53.1°	15.9°	105.8°

Each of the first three columns is the line of a town's graves at one layer; the last is the bearing from the town to the holy city. Read down the first column and the lines fill an arc of 108°: those are ten churchyard plots, each turned its own way. The second has three values. The third follows the last, 90° round, to within a tenth of a degree. The scattered style spaces its graves wider than rows do, so the ten grounds hold 989 graves at layer 2 and 1,014 at layer 3. In rows at layer 4 they hold 2,245.

Layer 1: the ten towns' churchyards at one scale, north up, sorted by their line. Kizilpinar, whose name is Turkish for a red spring, has no chapel: it did not fit its ground. Its graves still keep clear of the room the chapel would have taken, and the empty room cuts off three graves to the south. That is a bug. The Muslim burial ground's domed tomb had the same fault, fixed at layer 3; the churchyard's is next on the generator's list.

Layer 3: the same ten, sorted by their line, from Baranaara at 13.1° to Kizilpinar at 15.9°. The dashed arrow in each is the town's way to the holy city. The arrows look parallel; across these ten they differ by up to 2.7°. Kizilpinar has no domed tomb at this layer either, and its graves now close over the room it would have taken.

Layer 4: the same ten in rows, turned together to the qibla, each with its aisle (tan) from a gate to the domed tomb, sorted by their line as above. Every grave stands on its ground's rows. Eight of the ten grounds stand where they stood at layer 3, their domed tombs on the same spot; there, 733 graves have become 1,656. The generator moved on between the layers, and two towns were laid a little differently: Casseia's ground stands 129 m from where it was, Yukarikale's 51 m. Kizilpinar has its domed tomb at this layer.

The mosques knew the way first#

In two of the ten, the courtyard villages of Aman Ifrane and Multankhana, the generator also lays a mosque. From the first layer the mosques were laid from the world, not the drawing. The generator held one world direction for the holy city, south-east, turned it back through each town's turn, and laid the prayer hall along it. Both prayer halls lay along 135°, while at layer 2 their graves lay along 53.1° and 8.1°.

Layer 3 gives these graves and mosques one source: the bearing from the town to the city. Aman Ifrane's prayer hall now lies along 104.7°, its graves across it at 14.8°. Multankhana's lies along 103.6°, its graves at 13.6°.

The other towns' mosques have not caught up. They are laid the way any town lays its temple, square to their plots, and their walls stand off the city's line: Yukarikale's by 6.2°, Ksar n'Tidri's by 43.8°. Their dead face the holy city more truly than their prayer halls do. That is a fault, and it is next on the generator's list.

One town at layer 1 had its churchyard rows agree with its mosque. Multankhana's, on the black basalt of the Lava Tip, lay along 135.1°, the line of its prayer hall.

The churchyards don't face east either#

Follow the chain one link further, because the Christians' graves are square to their plots too. In the early Middle Ages the dead in churchyards were commonly laid facing east, feet to the sunrise. Our churchyards lay their graves square to the plot they stand in, and the plot turns with the town. In the layer-1 plans the Celtic village of Strathwych buries its dead at 97.5°, near enough east–west. Llanrhos, another Celtic village, buries them at 6.5°, north–south, across the old custom. Nothing in layers 2 and 3 changed the churchyards' rows.

Meanwhile: churchyards where none belonged#

The same cemetery request reached the island's other peoples who are not Christian, and a second rule gave every temple a churchyard along its walls, except in the four Islamic peoples' towns. The ontology names no tradition yet for the Indonesian, Vietnamese and Andean peoples. The generator's own rules give them their faiths: Indonesian villages build mosques, Vietnamese ones pagodas, Andean ones huacas, and their dead are meant to lie in their own way, scattered or in mounds. Instead their temples got rows of graves round a chapel.

The biggest was at Teluk Pelangi, the city at the mouth of the river where no boat could go in our river post, behind the Rainbow Gate. At layer 1 it had 534 churchyard graves spread over 600 m of the city. At layer 2, with its own burial ground, the temple rule still gave it 455 graves in rows. At layer 3 churchyards belong to the Christian peoples only: Teluk Pelangi's 85 graves lie scattered at 16.6°, across its qibla of 106.6°, and none in rows.

Our own table of names had made it look worse. It listed every grave as a churchyard grave, whatever the faith. Now a grave outside a Christian people's town is a plain cemetery grave. Qamareen's three stations still stand at layer 3, on the processional way up to its domed hill shrine. The generator lays the same 2.5 m stations and the same domed shrine for every people; only the table called the stations calvaries, and now it calls them wayside stations. It still files the domed shrines of the Muslim towns, on the hill and in the burial ground, with the Christian pilgrimage shrines. The concept artist, for its part, had been leaving such things out all along: its approved bird's-eye concept of Teluk Tinggi shows houses on posts and a red flag on the jetty, and no church.

What runs, what is fixed, what waits#

Laid and pinned: the four layers. Layer 1 is the 59 example towns as they were pinned earlier this week; layers 2 to 4 are the ten Islamic towns and Teluk Pelangi, laid again with each fix. The game itself does not lay the island's towns yet. Laying all 900 waits on a decision about baking the whole island.
Fixed and laid: every people's own burial ground; churchyards for the Christian peoples only; Muslim graves across the qibla; one holy city for the graves and the courtyard villages' mosques; the graves in rows turned together to the qibla, an aisle to each domed tomb.
Decided, and proposed: the holy city's place, and with it every town's qibla. The generator lays towards it in layers 3 and 4; the ontology holds it as a proposal. Its name and its kind are not chosen yet.
To come: the other 48 example towns laid again with these fixes, and pinned.
Faults, next: the other towns' mosques stand square to their plots, not to the city; a churchyard keeps room for its chapel where the chapel does not fit, as at Kizilpinar; the table files the Muslim towns' domed shrines as Christian pilgrimage shrines.
Not decided: whether Atlas's Christians bury facing the sunrise.

A city past the edge of the map#

Until this week the island's holy city was only a direction, south-east, and not a place. Now it is a place, across the sea, on a shore no map of the island shows. It has no name yet.

Every Muslim grave on the island lies across a line that ends there. Casseia keeps 145 boats and no quay. Does anyone from the island make the crossing? For now the holy city is known only by the way the island lays its dead.

Who Built the Rainbow Gate?

2026-10-18T00:00:00Z

The Rainbow Gate is a stone arch 1,165 metres high over the mouth of a cove at the head of our island's great southern bay. Every stone of it can be measured. Who raised it, no one on the island can say.

The Rainbow Gate from the bay, looking north-east into the cove. Engine render at Atlas c65fcf7, 2026-10-17 (render.rainbow-gate-bay).

Its feet#

Its feet stand 895.5 m apart, one on each of the two headlands that close the cove. Each leg is a column of dressed blocks, ten across at the foot, and it goes on 30 m down into the rock.

Where a leg meets the rock it covers 2,005 m². So far the town generator has laid 19,460 buildings in 59 of the island's towns. The largest of them, the church at Damebois, covers 1,984 m².

To one scale, north up: one foot of the Gate, the largest building in the city at its feet, and the largest building laid anywhere on the island. The outlines are from the world query database.

The Bay of the Rainbow#

Behind the arch lies a city of 12,788 people. Its harbour is inside the Gate: the way in from the sea runs between the Gate's feet.

The city is called Teluk Pelangi. In the coast's language that is the Bay of the Rainbow. It is where the river survey's thick red line ends: a river of cargo where no boat could go.

The Shard, the Arches, the Giants#

The island on its paper relief, with the four monuments in red and the towns at their feet in white.

The island has four works of this size. The ontology files each one as a place (poi.artefact.*), with its classes, its realm, and the code that builds it.

The Black Shard is a slab of black stone on the Lava Tip, the island's south-eastern end, 300 m wide and 100 m thick. Its top is cut on a slant, 1,000 m high at one edge and 900 m at the other. Lava runs out of its top and down its faces. A village of 1,200 people stands 313 m from its middle: Kalakot, Black Fort.

The Black Shard from the crater's side. Engine render at Atlas c65fcf7, 2026-10-17 (render.black-shard-front). There are no houses at its foot yet. The town generator has laid 59 of the island's 900 settlements so far, and Kalakot is not one of them.

The Hundred Arches is a bridge 2.8 km long across the lava field north-west of the Shard. It runs in fifteen straight legs that turn on rock islands. It has exactly a hundred arches: 24 in its longest leg, one in its shortest. Nine towers stand along it.

The Giants of the Pulpit are three hooded figures with folded arms, each in its own niche. They are cut into a slate wall, 480 m high at its highest, on an islet in the north. The rock's layers run straight through the figures and their niches: they were cut out of the cliff, not set against it.

The Giants of the Pulpit from the sea. Engine render at Atlas c65fcf7, 2026-10-17 (render.cliffs-12-giants). Two of the three are in view.

The nearest towns are Hvalvagur, 3,423 people, 2.1 km away, and Vestmannahofn, 7,442 people, 3.2 km to the south. Vestmannahofn is the westmen's harbour.

Thirty-eight sightings#

The ontology's proposal register lists 38 approved concepts of things like these: - a hand in a lake whose fingers are moorings; - a bell fallen into a village and used as a market; - rooms behind a carved face; - a stone disc whose shadow covers a well.

None of the 38 is one of the four, and no approved concept shows any of the four. In the ontology the 38 wait under a proposed class, with no place on the map.

Six of the 38 approved concepts, in rows from top left: Five moorings, Fallen bell market, The rooms behind the face, Water in the disc's shadow, The stone they never lifted, Abstract head in the wheat field.

What the ontology keeps#

The ontology deals in facts: classes, places, numbers, and the code that builds each one. Its class for the Gate, fantastic.form.colossal-arch, says that people made it: "a human-made arch of dressed stone". Which people, it does not say. For the monuments we have proposed three kinds of statement, kept apart on every page (specs/ontology-mystery.md in the Atlas repository):

Measured is what the world holds. Every number in this post is measured, and each one comes from a file named in this post's sources.
Held is what the island's people do with these things and call them. The Bay of the Rainbow and Black Fort are held. Later come their tellings: myths, each attributed to a people and marked as told, never as what happened.
Withheld is one line, the maker. It stays blank, and a check fails any page that fills it in.

The team decides the proposal. None of it is on the ontology site yet.

The blank is not a teaser. The team said it plainly today: who left these enormous things around the island is a true mystery of the world. Nobody working on Atlas knows the answer, so there is no answer to reveal later. The numbers around the blank stay exact: the mystery is in the world, not in what we say about it. Where we looked for a pattern and found none, the proposal says so. The Gate lies 7.1 km off the line from the Shard to the Giants.

Six people live at Lambagjogv, 333 m up on the Giants' slate, two and a half kilometres from them.

A Stair at Mavigol That Leads Back Where It Started

2026-10-18T00:00:00Z

Below the village of Mavigol, a flight of stone stairs steps 3.8 m off the paved valley road onto the bank and comes back down to the spot it left. Where it meets the road, its slabs, steps and side stones lie heaped against the kerb.

The valley road below Mavigol, looking south. Steps come down the bank on the left. Where they reach the road, just left of centre, the stonework stops making sense. To the right of it, a straight-edged patch of bare ground pokes up through the paving. Atlas e103060b (2026-10-17), lit, with ambient occlusion and contact shadows.

At the foot of the hill#

Mavigol is a Turkish-inspired village of 243 people, inland, a little south of the island's centre, in the Temperate Middle. It sits on a hillside at 262 m. A footpath ramps down from it, and beside the path people's stairs, a flight of stone steps 1.2 m wide, cut straight down for anyone who would rather not walk the bends. Both end at one point on the valley road, 180 m west of the village and 138 m below it.

The road is a stone road, 4.2 m of polygonal paving between 0.3 m kerbs: the game's Roman road. It runs beside a river that falls 57 m in the 3.5 km from Faltersee, 32 m of it in one kilometre. No boat works it. Roads cross those 3.5 km nine times, all by bridge: eight stone arches, all on this road, and one timber trestle 21.7 m over the water on another. Readers of the river post will remember Yaylahisar's nine crossings.

Left: the island on its paper relief (terrain aeead1e), the red square 6 km across. Right: the square. The river is the river post's survey (blue), the stone road dark; at this scale they run as one line. The red dot is the stair, at x 6,809, z 13,706 (x east, z south, in metres).

Come down from Mavigol by either way and you arrive at the heap.

Nobody drew it#

Our road network was grown, not drawn. A program we call the mycelium added the island's settlements one step at a time; each new village sent a road to the network over the cheapest ground, and each road fused with the ones it touched. What it laid is canonical. The roads are kept exactly as recorded and nothing regenerates them, so a road you remember is still there tomorrow.

The record keeps the roads as strokes: one continuous road line, from one meeting point to another. There are 1,572.

The old tools that built roads from the network tidied up as they went, and the tidying was never written down. One of the things they enforced was a standing rule of the team's: roads never ride roads. Roads join at junctions, part, and never come back along the same road. Where a route would leave a road and rejoin it, the router dropped the stretch between, and the old earthworks dropped any sliver under 8 m between two shared stretches. When we began building straight from the record, the tidying stayed behind with the old tools.

The stair below Mavigol is one of the 1,572: stroke 58.

7.8 metres, out and back#

The meeting point in plan, north up, on a 1 m grid. The red line is stroke 58: from point 0 on the road's centreline out to point 1, 3.8 m from the centre and 1.4 m past the kerb, then back through 2 and 3 to the same spot. The dashed blue stretch is the gap its slab bridge was decided over. It lies on the paving.

Stroke 58 is a people's stair of five points and 7.8 m. It rises 1.5 m on the way out and comes down again to the point it started from. Nearly five of its metres lie on the paving or the kerb. Its far end and its way back run within a metre of the last few metres of Mavigol's own path and stairs. Of the island's 1,572 strokes, it is the only one shorter than 8 m.

Built exactly as written#

Stroke 58 got what any short stair gets:

A tread, 1.2 m wide. Wherever a tread would cover a road it meets, it is trimmed back to the road's edge. Most of this one was trimmed against the kerb.
Steps, because 1.5 m in under 4 m is about 40 percent, and side stones with them.
A slab bridge, decided and never built. The record holds a span priced on this stroke, and the crossing table, given a short stone span on a footway, decided a clapper: flat slabs laid end to end on piers of stacked stone, the Dartmoor kind.¹ Its 2.2 m gap lies on the paving. The builder put nothing there: under the team's rule, a slab or timber bridge over a dry gap less than 0.8 m deep builds nothing.

Each part followed its rule. Every rule ended at the same kerb.

Close up, same frame. The steps on the bank are Mavigol's path, by our reading of the record's lines. Its foot and stroke 58 share the same few square metres of kerb.

Diagnostic colours: pink ground laid beside a way, white paving, cyan an earthen tread, red a bank side or a tread too steep, yellow the kerb, orange-brown the ends of a flight's slabs, light green stones, magenta a face seen from behind, which should never show. The heap is all stair: its tread, a steep red block, a slab end, a side stone, back faces and a twisted kerb. No bridge part is in it.

Every visual change here goes to a blind critic, an AI shown the render with no hint of what changed and asked to report every artifact it sees. This frame scored 1 of 5, with five major faults. The worst: the junction is "a pile of clipping parts", the red block, the slab end and the loose block, back faces, broken kerb pieces and pink bank shards above the paving. Then the path's foot ending in a lip, the way never joining the road; the patch of bank ground punching through the paving; a strip of back faces in the bank at the top left; and a stepped crease in the trough beside the stair. It knew none of the history. It saw what a player would see.

Un-building one stair#

The fix is the old rule, brought back at the builder: a stroke with no stretch of 8 m off the roads it lies on, and none running off to one of its own ends, rides them and is not built. It finds three on the island. Two carry a decided slab crossing; neither crossing was ever built.

The team has decided: riding strokes are left unbuilt, and those two decisions are dropped from the crossings record in a new version. Decided, not yet built, not on main.

The road record keeps stroke 58. In the patch under test the builder skips it as a point of zero length, so no other stroke renumbers.
The crossings record loses two decisions. Nothing on the ground changes from that: both gaps were dry and shallow and built nothing.
A test loses its subject. The example site for a clapper on people's stairs in the temperate realm was drawn from the decisions, and drew one that never built a clapper. That is how a critic came to look at this spot at all. The cell needs another site.

Forty centimetres of paving#

The hunt turned up a second, general fault. A stair or path ending on a road keeps its steps 2 m from the junction, measured from the road's centreline. A stone road's kerb is 2.4 m out, nearly 2.8 m where a stair meets it at 60 degrees, as Mavigol's own stairs do. So a steep flight can begin on the paving, its treads trimmed to the kerb. The fix under test keeps steps off any ground where the tread and side stones would lie on the met road at its level. How many junctions it changes, the tests will tell.

What runs today, and what does not yet#

In the world today - Stroke 58 and its steps, built at Mavigol from the road record as published (mycelium version 37, step 1,565). Its decided clapper builds nothing. The renders are of Atlas e103060b, the day before. - Stairs whose steps can begin on the paving of the road they end on.

Decided, not yet built or on main - Strokes that ride the roads they meet are left unbuilt; the two slab crossing decisions on them, neither of which builds anything, are dropped from the crossings record in a new version.

Under test - Steps kept off the road a stair or path meets. - Tests: a small network where a copy of the Mavigol loop must ride and a 20 m spur must not; Mavigol itself; island-wide sweeps that allow no riding strokes and no steps on a road.

Next - Before and after renders of this spot and of slanted stairs elsewhere, each to a blind critic, and a new example site for the clapper cell.

Does a path that hugs a road ride it?#

Mavigol's footpath does not lie on the road. For its last 14 m it runs beside it, 1.2 to 2.7 m past the kerb and 2.5 to 5.6 m above it, falling at about 40 percent, and drops onto the road only in its last couple of metres. By the geometry, though, it runs inside the cutting the road dug into the hill, the path's bench notched into the road's cut face. Its steps are the ones in the pictures above. Nobody has yet looked closely for the seam.

The rule never mentions the earthworks a road carries with it. If a path can ride a road's cutting without touching its stones, the old rule reaches further than its four words.

For close readers. The rule is written to catch any stroke that never leaves the roads it lies on. Of the island's 1,572, it caught three: Mavigol's, 7.8 m; one of 17.2 m at x 5,531, z −658; one of 19.5 m at x 21,806, z 21,623. Each of the three ends at the point where it began.

The best known stands at Postbridge, over Dartmoor's East Dart: first recorded in the 14th century, thought to have been built in the 13th so that packhorses could carry tin to Tavistock. Postbridge, Wikipedia. ↩

Cargo Only Floats Downhill: The River Trade of a Mountain Island

2026-10-15T00:00:00Z

On our island, 92 percent of everything that travels by river travels downhill. When we finally measured the rivers, that one number decided which boats exist and which bridges have to be rebuilt. It also decided why boatmen on the rapids above Ardbach spend their afternoons walking empty boats back up the bank on a rope.

The first survey: every river our island's trade used, coloured by what its water could actually float. Blue for the river barge, green for only a small flat boat, red for neither. Line width is the cargo the old simulation sent along it each year. Black crosses are fords, orange rings bridges too low for a boat. Look at the west coast, where the busiest river system fans out above the city of Ardbach. Then look east, at the thick red line running straight down to Teluk Pelangi: a river of cargo where no boat could go. The island beneath is a paper relief of the game's own terrain, drawn three days after the survey (terrain aeead1e); the rivers and their tonnes are the survey's.

Why we went and measured the rivers#

Atlas is a game set on one island, about 131 km across, with roughly 900 settlements. They are not decoration. Every village keeps a ledger: what its people need, what its fields and workshops make, what is left over. A trade simulation then moves each surplus to someone who lacks it, by the cheapest way there is: ox carts and mule trains on the roads, coastal ships between harbours, boats on the rivers.

A second program, our town builder, reads those ledgers and lays out every town to fit them: the fields and the market, and on the quay, how many berths the barges need.

That is where the trouble started. The town builder went to place barge berths and measured the water first. Our river barge draws 0.6 m. The water beside most barge towns was 0.4 to 0.5 m deep. At ten of the seventeen barge towns it tested, there was nowhere deep enough to float one.

The obvious next question was hard to hear. Had anyone checked that the rivers the trade used were wide enough, deep enough, free of fords, and under bridges high enough for a boat? Nobody had. The simulation took any river 6 m wide as navigable, full stop.

It had been routing about 3,000 tonnes of barge traffic a year down the river from Buchenbruck to Krebshorst. That reach falls 209 m in 4 km, and 103 m of it in a single kilometre. That is not a river a barge sails down. It is a waterfall with a cargo manifest.

Every river runs by two laws#

Rivers in Atlas are not painted on. They are grown: rain falls on the terrain, water finds its way downhill, and each stretch of river carries the water of everything uphill of it. From that flow, two laws give every river its shape:

Width grows with the square root of the flow: a river carrying a cubic metre a second is about 11 m wide.
Depth grows more slowly, roughly with the flow to the power 0.4. That is how real rivers behave: double the water and a river gets wider sooner than it gets deeper.

The roads read the same laws to decide where they need a bridge and where a ford will do. The renderer reads them to decide how much water to draw.

The day before this survey, we had already changed one of them. The island's rivers were too thin: they ran off rain like a dry lowland, not a wet mountain island. We gave them two and a half times the water, but only in the depth law. The widths stayed exactly as they were, because every bridge, ford and riverside town had been built to fit them. Rivers got 43 percent deeper without moving a single bank.

Depth against width, mid-channel, before and after the change. Each dot is a trade reach at its narrowest. Every one of them now has at least 0.57 m of water, enough for a flat river boat. The barge needs a river about 9 m wide; 88 of 162 reaches are. A coastal ship would need a river 23 m wide, and only a few stretches near the sea come close. The ships stay at sea.

The town builder's problem, drawn to scale. In a 6 m stream the barge sits 3 cm into the gravel; a flat river boat drawing 0.3 m has a hand's breadth to spare.

So the depth turned out to be fine. The survey's real surprise was somewhere else.

The island is all slope#

We measured all 162 river stretches the trade uses: 373 km of water between one landing and the next. For each one we found its steepest kilometre.

A river that falls a metre in a kilometre already runs fast. Boats work it, but going up is hard: poled, towed from the bank, sailed when the wind allows. In our rules, nobody works a boat upstream against more than three metres a kilometre. Downstream still works, with skill.

The median trade reach on our island falls 7.9 m in its steepest kilometre. Thirty fall more than 20 m.

The steepest kilometre of every reach the trade used. Only the blue and green bars can be worked both ways. The tall red bar on the right is thirty stretches of white water.

It makes sense once you see it. Atlas's island is a composed, exaggerated landscape: ranges, canyons and cliffs hundreds of metres high, packed into an island you can ride across in a few days. Its rivers are short, and they come down fast. Most of them are what the ontology calls river.torrent in their upper courses, and many calm down only in their last few kilometres before the sea.

Ninety-two percent#

Then we looked at which way the cargo went. Of all the tonnage the old simulation sent by river, 92 percent went downstream.

And what was it? Not wine, not silk. Firewood: 67,000 tonnes a year, floating down to the towns. Then manure for the fields (28,000 t), milk, hay, building stone, timber and barley: heavy, cheap, bulky things that cost more to cart than they are worth.

Nothing in the code says "send things downhill". It falls out of three facts meeting:

What is heavy and cheap, firewood above all, is worth moving only by water.
Where it comes from and where it goes. The woods and pastures lie up the valleys; the hearths crowd into the towns below.
What it costs. Going up a river costs three times what going down does, so whatever goes inland finds a road instead.

The rivers had become one-way conveyor belts long before we knew it.

That gave us a way out. If the river trade is almost all downstream, a steep river is not useless. It is only useless going up.

Three ways to read a steep river#

We laid out three rules for the reaches too steep to work both ways:

If the steep reaches…	Reaches worked	River cargo kept
carry nothing above 3 m a km	43, 86 km	43%
carry cargo downstream only, up to 5 m a km	62, 120 km	55%
carry cargo downstream only, up to 8 m a km	81, 161 km	67%

History had already tried the middle one. On the Hozu river west of Kyoto, boats carried timber and rice 16 km down through the gorge from 1606 until the railway came. Then the boatmen hauled the empty boats back up by rope, every day, along paths you can still see beside the river. On the Isar in Bavaria, rafts of mountain timber brought firewood, lime and even Italian wine down to Munich, by 1870 the largest raft port in Europe.

The strict rule would have done something cruel to the island's busiest river. The western plains' rivers come down to the sea at Ardbach, and their main stem carried some 34,000 tonnes a year. The strict rule would have cut it at two short stretches that fall just over three metres a kilometre: 1.5 km below Rochefont, and the last 2.4 km into the city.

We chose downstream only up to 5 m a kilometre. Above that, the goods go by land.

The rivers above Ardbach, every landing's height against its distance up the water. Follow the arm from Clairfleur: boats work it as far as Belleval, where it drops off an escarpment (red) that no boat can run, down past Pommieray to Alderheim. The arm from Vieuxbourg has one short rapid below Rochefont. The two meet at Valpre, and the last stretch into Ardbach is purple: boats come down it loaded and go back up it on a rope.

Meet the river boat (and say goodbye to the barge)#

The trade needed a new boat: a flat river boat of 2 tonnes, 9 m long and 1.8 m wide, drawing 0.3 m. Like the punt, it is poled and towed up the gentle streams. Unlike a punt, it is run down the rapids loaded and hauled back up empty. Two boatmen work each one. The economy now has them as an establishment of their own, the river boatmen: 481 outfits across the island, where there used to be 536 barges.

Then we reran the island's trade, and the barge vanished.

Not one tonne went by barge. The water deep and gentle enough for a barge adds up to 9 km, scattered in six short pieces between steeper streams. A barge cargo would have to change boats at both ends of each piece. The small boat simply runs through. Geography out-competed the barge.

The barges used to carry 149,200 tonnes a year; the river boats now carry 50,200. Ox carts and mule trains take on 65,000 tonnes more. The coastal ships take 15,700 more, because some cargo that left the rivers simply goes round by sea.

Tonnes a year carried on some leg of their journey, by vehicle. The barge does not shrink, it disappears.

The rivers' biggest cargo changed too. Firewood fell from 67,000 tonnes a year to 13,000, most of it now carted. The new leader is manure, at 18,000 tonnes. The little boats of the western plains spend most of their days carrying dung.

The island's river trade after the change, coloured by the water each reach gives the boats: blue a river (barge and boat), green a stream (the small boat both ways), purple rapids (downstream only), red too steep. The red rivers are now hairlines: nothing moves on them. The western plains carry almost everything. Same relief, same rivers, new rules.

Ninety-eight villages for one woodpile#

Change one rule and the island rearranges itself.

Hvalvagur is a Nordic-inspired harbour town of 3,423 people; the name means whale bay. It burns about 6,600 tonnes of firewood a year. Under the old rules, half of it came from a single village in the rain-forest east: Phuc Son, 284 people. Its wood floated 30 km down to the coast at Long Kiem by barge, then sailed 82 km to Hvalvagur.

Phuc Son's river turned out to be rapids. Its boats now stop at Quy Tuan after 8.5 km, and the wood rides 20 km by ox cart along a forest track to the sea. The cost doubled, and Phuc Son now sells its firewood to Farnheim instead.

Hvalvagur's merchants did what merchants do. Its woodpile now comes from 98 villages: 14 Nordic, 15 Finnic-Baltic, 17 Arabic-inspired and 28 in the Peruvian-inspired highlands.

The island did not find that wood at once. The trade simulation works in rounds: match what is short to what is spare, order the rest from villages with hands to spare, recalculate, repeat. After six rounds, Hvalvagur and the town of Kaltenberg were still 1,157 tonnes short of firewood, which would have been a cold year. It took eight rounds for the island to find them their wood.

What a traveller would see#

Stand on the bank of the rapids below Marmesnil, on the last stretch into Ardbach, and watch for an hour:

A boat comes down. It is loaded from the valley above, and two boatmen keep it straight with poles. It is round the bend in a few minutes and at the quays of Ardbach in under half an hour.
A boat goes up. It is empty, and the same two boatmen walk the bank towing it on a rope. At about two and a half kilometres an hour, the stretch takes them an hour.
Carts on the road above. They carry what Ardbach sends inland: salt above all, 925 tonnes a year, then rope and lime. None of it goes up by water.

That asymmetry spreads into everything. Ardbach, a city of 12,000 at the river mouth, receives the valley's goods by boat, cheaply. What it sends back inland pays for oxen.

Alderheim, the bigger city upstream with 37,600 people, has it stranger still. Above it is the escarpment; below it the water is gentle only as far as Marmesnil. A city that size can work its boats both ways on about four kilometres of river.

Belleval, at the top of the escarpment, is where the boats stop and the carts begin. Some 3,550 tonnes a year come off the boats there onto ox carts, most of it firewood and hay bound for the towns below the drop. Under the old rules, Belleval, a village of 1,162 people, had no carters' yard at all. Now it keeps fourteen. Nobody placed them there. The slope of a river did.

Nine bridges in two and a half kilometres#

The survey's last job was the crossings. Our road network grew over the terrain like a fungus, joining settlements by the easiest ground. A road builder then decides, crossing by crossing, what to build: a ford, stepping stones, a timber trestle, a packhorse bridge or a stone arch. It decides by the water, the slope, the realm's building stone and how important the road is.

Not one of those choices asked whether a boat would need to pass under.

On the reaches the trade uses, roads cross the water 358 times by bridge. Half of those bridges stand less than 2 m above the water, and a quarter less than 0.8 m: fine for a cart, a wall for a boat. On the 54 reaches where boats now carry cargo, 111 of 133 bridges are too low, and one ford is left.

A timber bridge low over a river: lovely to cross, impossible to pass under. The water is our experimental river shader: you look through the surface to a stony bed, shallows paling to turquoise over sand.

Stepping stones across a stream. For walkers, perfect. For a boat, a weir.

And sometimes a road and a river simply cannot leave each other alone:

Between Yaylahisar and Bozyaka, in the island's Turkish-inspired south, the valley floor is barely wider than the river. The road takes the flattest ground it can find, and that ground keeps switching banks. Nine crossings in 2.4 km of river, most on stone arches a metre above the water. This reach falls too steeply for boats anyway, so here the bridges may stay low. A traveller would remember it as the valley where you are never dry for long.

The plan, which is not built yet, is for every bridge over a reach boats work to get a navigation arch: one span high enough for a boat. That is more than a hundred bridges for the road builder to raise. It is also a decent reason for a town to put a toll on its bridge.

Where this stands#

Running in the simulation today - The survey of all 162 trade reaches: their width, depth, fall, crossings and cargo. - The rule deciding what each reach carries: both ways, downstream only, or nothing. - The small river boat in the trade, the river boatmen in the economy, and their berths passed to the town builder.

Experimental - River beds you can see into. The renders above come from it. Blind AI critics, shown our frames beside a reference photo of a real turquoise river, score them 3 to 4 out of 10 at the last count. We iterate one change at a time until they come close.

Planned - Navigation arches over the reaches boats work. - Boats on the water that you can watch, and towpaths along the rapids. - Rafts.

The raft question#

Rafts are where this gets strange. A raft is the boat and the cargo at once. Built from upland timber, floated down with goods on top and taken apart for its timber at the end, it ties the river trade to the forest itself.

Give our island rafts and a valley's trade becomes a function of its trees. Then what happens to a river town the year its forest runs out?

A detail for close readers. Five settlements on the trade's rivers are named for fords: Gelbenfurt, Reiherfurt, Wildesfurt, Tannenenfurt and Dornfurt (Furt is German for ford). Not one of their reaches has a ford any more. The roads there cross by bridge. The names remember a crossing the road builder has since bridged over. Or they remember nothing, and whoever named them was simply fond of fords.

Another. The old simulation floated oxen, sheep and pigs down the rivers on barges, about 7,000 tonnes of livestock a year. It still puts them in boats now. Livestock should walk, driven along the roads by drovers. The drovers have not come yet.