The road that was underground
written 2026-09-09The roads on the board looked wrong, and the plan was to draw them a better surface. Then somebody measured how wide a road actually is on a phone, and how often the ground it lies on rises straight through it, and the answer to both was that the road should stop being a thing you can paint at all.
A road on this board joins the middle of one tile to the middle of the next. The way it was drawn for months was the obvious way: stretch a flat strip between the two centres, pin each end to the height of the ground beneath it, and paint a dirt texture along it.
That works perfectly if the ground between the two ends is a straight line. It is not. The terrain is built from noise on a scale much finer than a tile, so between any two tile centres the ground rises and falls several times over. A flat strip stretched between the ends of that has no idea about the middle.
Anything set like 0.0665 was measured by a machine and generated into this page from the committed results. Everything else was written by a person.
Most of it was underneath the ground
The strip was lifted very slightly — a hair above the terrain, enough to stop the two surfaces flickering against each other where they touched. That lift was the entire allowance for the ground doing anything unexpected in between.
Measured across every pair of neighbouring land tiles on a full-size board — 14,452 of them, sampled at 19 points along each — the ground climbs above that straight line on 69% of them. Not marginally: the typical overshoot is 1.8 times the lift the road was given, and the worst is 11 times it. For most of the network, most of the road was under the dirt, surfacing in patches wherever the ground happened to dip.
Nothing in the project noticed. Every test passed the entire time. A road that is buried still exists, still has the right length, still connects the right two places, and still reports every number a test could think to ask it for. It is only wrong to look at.
Four pixels wide
The plan at this point was to make a better dirt texture: generate a nicer surface, check that it tiles seamlessly, paint it along the strip. Real work had already gone into the machinery for that.
Then somebody measured the road on the screen it is actually played on. On a phone, a road on this board is about four pixels across. Four. There is no amount of gravel, rut or wheel track that survives being drawn four pixels wide — at that size the only things a road can communicate are its shape and how dark it is. The entire texture plan was answering a question nobody could see the answer to.
That is the useful kind of measurement: not one that shows the work was done badly, but one that shows the work was aimed at the wrong thing.
So the road stopped being an object
Instead of drawing a road, the board now asks every point of ground a question: how far are you from the nearest road? Close enough, and that patch of ground paints itself as road. Further away, and it stays grass. There is no road shape anywhere — no strip, no model, no texture, nothing to lift, nothing to bury.
Because the road is now made of the ground itself, it follows every rise and dip exactly, for the same reason your shadow follows the pavement. The burial problem does not get fixed so much as it stops being expressible.
It is also cheaper than what it replaced, which was not the point but is a pleasant surprise — the layer's own draw call and its triangles left the budget along with it. A quiet board draws 62 things per frame, in 16 distinct ways, and none of them is a road.
And the junctions solved themselves
The hardest part of drawing roads is not the roads. It is what happens where they meet. Two strips crossing at an angle leave a notch on the outside of the turn and a doubled-up wedge on the inside, and every road system ever built has some machinery for patching that — special corner pieces, mitre calculations, junction artwork.
Asking "how far is the nearest road" needs none of it. If a patch of ground is close to two roads at once, it takes whichever is nearer, and that answer is automatically the right one. Ground near a corner is near both arms, so it fills in. Ground where three roads meet is near all three, so the meeting point comes out as a neat rounded junction that nobody drew and nobody has to maintain. Add a fourth road and nothing needs to change.
It also cannot double up. The old strips were slightly see-through at their edges, and two see-through things stacked are darker than one — which is why every crossing had a smudge on it. Nearness has no such arithmetic. Being near two roads is not darker than being near one.
The wheel ruts come along for the ride. A rut is just "a bit off to the side of the middle", so the ruts bend around corners and merge at junctions on their own, because they were never drawn as lines in the first place.
The record
Every quantity on this page was read out of the committed measurements by a script. The terrain figures come from a probe that walks a real generated map and asks the game’s own height function what the ground does between each pair of neighbouring tiles — not a second copy of that function, which could drift, but the one the board itself renders from.
All three of the original complaints — that the roads were too wide, that the texture repeated in obvious squares, and that they left gaps when they turned — came from somebody opening the game and looking at it. None came from a test. Two of the three are now impossible rather than fixed: there is no texture to repeat, and there are no two pieces between which a gap could open.