The tower that lied to the camera
written 2026-09-05Every faction in Bannerfall got a capital city this week — 9 castles, each assembled from five separately-generated parts. Most of them behaved. One spent five casts chasing the wrong problem before the real one turned up hiding in a tower whose own picture of itself was a lie.
The chapter before this one was about threading a machine-made shape onto a skeleton. This one is about a different kind of building entirely: castles, assembled from five separately-generated parts the same way chapter two's one prototype was, now run 9 times — once for every faction, plus the empty seat a city becomes when nobody rules it any more.
Most of them finished on the first attempt. Two did not, and they failed for two different reasons — one a known problem with a known fix, the other a genuinely new kind of mistake this project had not caught before.
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 capitals, no story to tell
Nothing here is really worth a chapter on its own: a stone keep, a length of curtain wall, four corner towers, a gatehouse and five cone roofs, generated once each and locked together by the same aspect-ratio rule chapter two already described — the finished building has to stand within a fixed band of how tall it is against its own footprint, or the assembly refuses to ship. 7 of the 9 factions cleared that on the very first attempt.
The two that did not are worth the space, because they failed in two completely different ways, and only one of them is the failure you would expect.
One failure had already been solved once
Azure's keep came back too squat to clear the band three separate times, measuring 1.262, 1.258, 1.277 — the same failure chapter two's castle wall had already been traced to once: a concrete, squat-shaped word describing the faction's own architecture beating a plain height instruction buried in the same prompt. The fix for that case was already sitting in the shared prompt file: say the ratio as a positive statement instead of naming the shape you don't want. Recast under the corrected wording, azure's keep landed at 1.3512 on the very next attempt, after 3 failed casts under the old one.
The other faction's keep failed the same way, and the fix made it worse
Ardent's keep failed the same way at first — squat, three times over, measuring 0.866, 0.888, 0.817 — so the same fix was tried: reword the prompt to state the ratio positively instead of naming the shape to avoid. It made things worse. The reworded keep measured 0.712, further from the band than any of the three attempts it was meant to replace.
That is the sentence that matters here. A fix that had already worked, on a real failure of the identical shape, made this one worse instead of better. Something else was wrong, and it was not going to be found by casting the keep a sixth time.
The tower that lied to the camera
The keep and the curtain wall — recast 6 and 3 times respectively — both looked like plausible suspects, and both got real fixes along the way. Neither was the actual problem. It was one of the four corner towers, a part nobody had touched, sitting quietly at its default shape while the keep took the blame.
A round tower's own footprint should read close to square from directly above, regardless of how tall it stands — every other faction's does, in a range of 1.01 to 1.23. Ardent's corner tower measured 0.792 by 0.861 by 1.987: a ratio of 2.51, more than double what a round tower should ever produce. And its own reference picture — the image the service hands back beside the file, supposedly a preview of that very geometry — showed an entirely ordinary, upright, round tower. The picture and the file did not agree, and the picture was the one lying.
That runs the opposite direction from the usual failure this programme has caught before. The usual failure is a defect a person can see and an instrument cannot — a shape change no measurement will register because the measurement was never looking for it. This is a defect nobody could see, in a picture that was never going to show it, sitting inside a file that measured completely differently from the picture beside it. It surfaced at all only because nobody trusted the number until it had been taken three separate ways and still refused to change its mind.
The right part, and then one more turn of the wrong one
Recasting the tower did what the keep and the wall fixes had not: it moved the whole assembly to 1.23451, its closest approach to the band yet. But it did not finish the job. The keep — still under the reworded wording that had made things worse before — was now the part holding the assembly back, at its own footprint ratio of 0.355, well short of the 0.41 to 0.66 range every other faction's keep actually lands in.
The fix, once the tower was out of the way, was smaller than any of the five casts that came before it: dial the keep's own wording down, from three times as tall as it is wide to two. That alone closed the remaining gap. Ardent's capital assembles to 1.34238 today — comfortably inside the same band all the other factions clear.
The gate that passed was not asked the question that mattered
One thing did not get fixed. The re-cast tower measures 2.08 on the same footprint ratio, against 0.99 to 1.01 for the other eight — still an outlier, just no longer a large enough one to keep the whole assembly out of band. It is standing on ardent's capital today, not round, and nothing failed on it, because the gate that matters here checks the finished building's overall height against its footprint. Nobody ever asked it whether one tower, on its own, actually looks like a tower. That question still has no gate, and this is not the first time in this programme's own log that the answer to "did it pass" and the answer to "is it right" have come apart.