Still no GP for age, but in case you’re curious, here’s the model I promised above:
See the pdf for a figure describing the structure. There’s a somewhat unconventional thing whereby instead of one big multivariate normal for all the subject coefficients, I instead break them into k sets of 3-variable mvns, which I’m hoping achieves a good/principled balance between the desire to let the three parameter classes (influences on errors, influences on log-RT location, influences on log-RT scale) mutually inform without the bias-to-zero on this mutual-informativity that the full mvn would impose.