Training run finally converged. 97.3% accuracy. Now to figure out why.

Hot take: most ML papers can't be reproduced. We need better standards.