Where The Curve Turns Twice

You sweep the width multiplier $k$ of a ResNet on a CIFAR-10 training set carrying 15% label noise, holding the optimizer, epochs and data fixed, and record final errors:

width $k$	train error	test error
4	24.0%	33.1%
10	7.9%	27.4%
18	0.0%	36.2%
26	0.0%	29.0%
64	0.0%	21.3%

Where does the interpolation threshold sit, and what does the classical bias-variance reading get wrong here?

At $k = 26$ , the first width past the test-error peak, since the threshold is read off the test curve

At $k = 10$ , the last width before test error starts climbing; past that point extra capacity only fits noise, so the 21.3% at $k = 64$ has to be a seed artifact

At $k = 18$ , the smallest width that drives train error to zero; the classical reading stops at $k = 10$ and predicts every larger model is worse, missing the second descent from 36.2% to 21.3%

At $k = 64$ , since the widest model interpolates most easily; the 36.2% spike at $k = 18$ is run-to-run variance