x sample from model
x write test for rank of outputs
x extract narrow band
x look for errors in fit ellipse code
x test rank of model outputs
x try running locally in case faster
x fit ellipse
x fit narrow band ellipse
x randomize choice of prefixes
x fit to narrow band data with more points
x interpret narrow band data from more points
x fit to data from natural text
x save hidden_states instead of logits
x compress logrob data to send to local
x send logprobs to local

plot solve times for different model sizes -- is the eigh data enough? Do I need MOSEK times?

Find the NL ellipse error distribution for model outputs

inverted points p
true hidden points q
minimize MSE(Rp, q)
R is a rotation
Baselines: random proj, recovered proj, recovered orthonormal proj.

Map outputs from target model to the image of the source model
Find the NL ellipse error distribution for other model outputs mapped to the model image

