The human as a compute layer
Models have outcome data in abundance and almost no process data: how a decision came about, what was discarded, what came from memory and what was computed fresh. People have the process and rarely see it.
Reading state over months, calibrated so that sessions and people stay comparable, yields a description of how a person computes: how much they hold at once, how deep they go, how fast a loop runs, how long attention holds under load. These are the limits of a mind, as numbers with units and an uncertainty.
Once a limit is a number, two things follow. It can be checked against an intervention, honestly, with windows free of intervention as the comparison and the person as the unit of analysis. And where the limit stays hard, the model can take that part, while where it gives, the gain is read off the same number that bounded it before. The human becomes a compute layer with known cost and known reliability. Not a competitor to the model, but an instance beside it.