softmax / log_softmax stable for logits of size 50: ok
discounted returns: ok
GAE: lambda=1 equals return minus value, lambda=0 equals one-step TD error: ok
group advantages: loo = G/(G-1) * mean, standardized has unit sd, constant group gives 0: ok
KL exact 0.11635   k1 0.11609   k2 0.09049   k3 0.11655   min k3 3.05e-15
clipped surrogate on both signs of the advantage: ok
pass@3 estimator mean 0.6590 vs exact 0.6570
bootstrap interval for the mean of N(1,1), n=400: (0.891, 1.089)
score function check, max |numeric - analytic| = 8.88e-11
all assertions passed
