The test
Alberto Campo Baeza is a Spanish architect whose writing and 139 projects are public. We wrote his framework only from material up to 2008, then gave the model the briefs of 56 projects he designed from 2009 to 2026. For each brief, four arms wrote ideas: plain, his name only, the 2008 framework without his name, and the framework with it. A judge compared each idea with his real concept. In all, 1,680 ideas, restated to the same length as his concept.
The model claims to know only 3 of the 56 projects; on the other 53 every number is within a point or two of the totals.
Results
| Arm | Hits his central move | Judge picks his real design |
|---|---|---|
| Plain | 24% | 96% |
| His name only | 26% | 93% |
| Framework, no name | 18% | 71% |
| Framework and name | 19% | 68% |
The framework cuts the judge’s success at spotting the real design by about 28 points, the one large effect. It does not raise the share of ideas that land on what he actually did.
The framework writes his greatest hits
Framework ideas sit closer to his pre-2009 projects than his own later projects do. Only about a quarter of framework ideas move further from his past than he did, against half of plain ideas. He kept moving. The framework kept repeating the cabin on the cave, the walled garden and the diagonal light.
He is halfway along his own direction
On an axis from plain ideas (0) to the framework’s ideas (1), his real later work sits at 0.55, in every period from 2009 to 2026. His work points the way the framework points, but only half as far. A framework sampled at full strength overshoots into caricature.
Where his real work sits between plain ideas and the framework’s
That gives the selector a measured number. To land where he lands, pick at about half strength. A pick rule that keeps the idea nearest his position raised hits from 19% to 25% in the follow-up (The shape of a framework).
A forecast, sealed
Using the 0.55 position, we generated forecasts of what he might design next and sealed 15 dated predictions with cryptographic hashes on October 29, 2026, to be opened against his future work.
Limits
Every verdict is a model’s, on short texts; text is not architecture. Real concepts carry specifics that generated ideas lack. One cutoff and one framework-writing pass.
Sources
- Feathers & Wax, Campo Baeza prediction test, code and sealed forecasts in
research/campo-baeza-predict. 4,873 calls, pinned models.