Building Agoro Road in public

The second renderer is finished. Same game underneath, two ways of drawing it: the original flat view and a real-time 3D one, sharing one simulation that neither of them is allowed to touch. Both replay an identical recorded run to an identical result — the proof that the 3D view is a view and not a second, subtly different game.

On the test deployment it draws a frame in about 16.7 milliseconds, and the slowest one in a long run took 16.8. That is a locked sixty frames a second with forty travellers walking and roughly seven hundred thousand triangles on screen. Seventeen of seventeen checks pass.

And it is still not shippable, because eleven of the objects standing in it are placeholders — the ones described in the entry below, generated from written descriptions instead of from the concept art. The renderer is done. The thing the renderer draws is not.

Two figures came off the desktop printer for the first time: Amara Sol at 80 mm and Daniel at 97 mm, in the one purple filament we have loaded. Each carried a different answer to the same question — how to hold up hands, a pot, a hem and hair curls while they print.

Amara used our own answer: thin struts and pegs fused into the model, snipped off afterwards, plus a solid block under the hem. Daniel used the slicer's tree supports. Both answers were wrong at this size. The pegs behind Amara's hair and under her earrings cannot be reached with cutters at 80 mm. The block under the hem left rubble between her legs. Daniel came off the plate covered in fine strings, with scars where the trees had held the bag and the staff.

Around the print, the pieces that let people watch it went live the same day: a browser view that replays the toolpath as the printer reports its progress, a post every five per cent to the community server, and a "wake the printer" walkthrough for the stream. None of that changes what the prints look like; it changes who can see them.

The 3D view had passed everything we asked of it. The same inputs replay to the same frame over six hundred frames, the frame times are fine on a phone, the fourteen finished meshes stand on the right rings. All of that is true and all of it was verified.

What nobody had done was put a capture of it beside the picture it is supposed to look like. When we finally did, the distance was not small. The ground is one flat green plane with no horizon. The road is a beige ribbon. The carved stone rings — the hero element of every frame in this game — are translucent pale discs at thirty per cent opacity.

None of that is a bug. Every one of those is something that was never built, passing a test that never looked.

The buildings and props were generated from written descriptions. The concept art for most of them already existed, in the same folder, and was not used.

The drum came back as a black cast-iron cauldron. The lantern came back as a Victorian street lamp. The goods cart came back as a wooden barrow. Judged one by one against the cards they were meant to match: four were the wrong object or the wrong palette, two had no card to judge against, and five were close enough to ship at map scale.

The tempting call was to redo the four and keep the five. Daniel ruled the other way.

we need consistent look across the concept art and props and images and trailersDaniel, 13 September 2026

So all eleven are being made again from the art, not from descriptions of the art. The five that were individually acceptable were not acceptable as part of a set, because a set is the thing we are actually making.

Correction, 20 September. Two of the three examples above are right and one is not. When we finally rendered the delivered drum and looked at it, it is a drum — a standing drum on a stand, the correct object. What is wrong with it is the material: lacquered orange, where the card is hide, rope and weathered wood. The "cast-iron cauldron" line came from a written note that nobody checked against the model, and it then travelled through three of our own documents as though it were an observation. The mistake it describes is real; the example was wrong. We are leaving the original sentence above rather than quietly editing it, because the failure it caused — asserting about an image without opening the image — is the same failure this entry is about, and we made it while writing the entry.

Earlier in the week we had recommended one of the AI image providers we use for making reference plates. That recommendation had a caveat written into it by the person who made it, and the caveat was the honest part: nothing had ever been compared against the alternative. It said this provider is good enough on our art. It did not say this provider wins.

So we closed the gap: the same jobs through two providers, seven configurations, three runs each — twenty-one images for about forty-seven credits, roughly what a single wrong 3D asset costs to make.

The other provider won, and the plate work moved. We had been about to standardise on the recommendation without ever having tested it.

These three went a different route than the eleven. The concept card was turned into a clean standing plate, and the plate — not a description — was turned into the model. Rigged, with idle and walk.

The same person survives all three stages. The scarf, the satchel and the head wrap are still there at the end. The roped bundle on the carrier's back came out as real geometry rather than a painted-on suggestion.

That is the entire argument for driving this pipeline from art instead of from text, and it is why the eleven are being done again.

Kofa is the Sankofa bird. He was built for Agoro Camp and he belongs to the studio rather than to one game — the same bird, the same look, different abilities in each — so bringing him to this road costs us nothing to design. His loop is retrieve, reveal, scout, and scout only unlocks on ground whose past you have already retrieved. The future is not visible until the past is known. That rule is the character, not a flavour text.

The gold egg he carries is the read-out for it: empty means this place's past is still lost. That works because the egg came out of the mesh as a genuinely separate object, so the game can actually light it rather than just imply it.

The mesh is finished — textured, with a fifty-six-bone rig. Then we opened the animation library we had planned to take flight cycles from and found it classes a two-legged bird as avian, and its avian section is empty. Not thin. Empty.

So every clip he has will be hand-authored. That is a real cost and it is now in the plan rather than in a surprise.

Update, 13 September. When we went to pose him, he shredded. The rig was not merely missing its animations — its skeleton was in the wrong place. We measured each bone against the flesh it actually controls: the bones sat, on average, most of a body-length away from it, and the beak's bone was further from the beak than the bird is tall. Two human characters rigged by the same tool measured twenty-five times better, which is how we knew it was the rig and not our ruler.

The weights — which vertex follows which bone — were fine. So we kept those and moved every bone back inside its own body, working out each joint from the region where two bones share influence. That got him standing. The picture below is the first one that has ever existed of this character.

Then the second fault: pieces of the gold egg were tied to a toe, a quarter of their weight each, on the far side of the body. At rest that is invisible, because at rest every bone agrees. The moment his head swung, the egg tore into shards. Both faults hid in the one pose anybody ever looks at.

He now has three hand-authored clips — a walk, an idle, and the look-back above. They live in the character file; he is not in the playable build yet.