Medicine & health · the experiment
Audit the skull implants
Three public AI models, 239 damaged skulls, two criteria fixed in advance. Four questions, from beginner to expert, to understand why convincing shapes cannot be fitted.
1 · Set the baseline
Beginner
Why compare each model with "doing nothing", that is, handing back the remaining bone as it is?
2 · Read a negative score
On the 210 carved holes, AdaPoinTr's advantage over "doing nothing" is −0.036; that of the model trained for the task, +0.745.
Intermediate
What does the advantage of −0.036 mean?
3 · Make thickness readable
With 6,144 points on a skull, neighbouring points are 2.56 mm apart on average. To tell apart the two faces of a 3-mm plate, points must be about half its thickness apart, 1.5 mm. On a surface, spacing varies as 1/√N, where N is the number of points.
Expert
How many points would give a 1.5-mm spacing on the same skull?
4 · Change the written prompt
MSN-FM accepts a written prompt. But its text branch ends in a function that sets every negative value to zero: 97.5% of its text code is exactly zero.
Expert
What does changing the written prompt do to the output shape?
Audit complete
—