Interesting drift in numerical things (pre-optimization).

May 2021:

> loss(m_im0, p_m_im0_warped_norm)

20.39865222030749

October:

> loss(m, target_value)

20.39928388822998

now, attempting to exactly reproduce the October run:

> loss(m, target_value)

20.400127548040803

Not much difference, 1.0000413573248672 is the ratio of the last two numbers, but still... uhmmm... what did they change?