The test
We took the same 2008 evidence on Campo Baeza and laid it out five ways: the current format, a generation-first rewrite, a library of 82 project cards, his own words only, and the current format plus his most similar past projects. Each shape answered the same 56 later briefs, scored as in the prediction test.
Results
| Document | Hit rate | Judge finds the real one | Position (his: 0.56) |
|---|---|---|---|
| His name only | 26% | 93% | 0.44 |
| Current format | 18% | 68% | 1.00 |
| Generation-first rewrite | 18% | 70% | 0.75 |
| Project cards only | 21% | 81% | 0.67 |
| His own words only | 12% | 81% | 0.79 |
What the numbers say
- Any condensed framework carries his voice. The judge’s success at spotting him drops from the 90s to about 70% whichever condensed shape is used. Raw cards or quotes alone are easier to see through.
- No layout predicts better. Every framework shape hits 12 to 21% of the time; his name alone hits 8 points more than the average framework document.
- The generation-first shape halves the overshoot at no cost in voice, landing at 0.75 against 1.0 for the current format.
- Ideas near his position hit more. Keeping, from each brief, the idea nearest his typical position hits 25% against 19% for the arm as a whole.
What we changed
The person’s distance from the plain answer now belongs in the framework as a measured number, used at selection. Our framework schema stores a profile, a document on how the person makes things, a calibration file with that number, and dated records of what they did.
Where each document lands, from plain ideas (0) to the current format (1)
Limits
One person, one cutoff, four ideas per brief per arm, and a model judge on short texts. The brief explains more of the variance in hits (59%) than the document does.
Sources
- Feathers & Wax, framework structure test, code in
research/framework-structure; schema inresearch/frameworks/SCHEMA.md. 3,612 calls.