Question about res function

[HuangruiChu]

I think in the paper we define the auditing algorithm A : (X × [−1, 1])m → [−1, 1]X

That is because
f_0 ->[0,1]
y->[0 or 1]

f-y -> [-1,1] if f-y is more closer to zero, that means the prediction is more precise. 

however, in res, the mini value start from 1.
def res(p, y):
return y * ((p>=0.1)/(p + 1e-20) + (p<0.1) * (20 - 100 * p)) +
(1-y) * ((p < 0.9)/(1 - p + 1e-20) + (p>=0.9) * (100 * p - 80))

delta = res(probs,y_val) has the range of [1,20] according to the code you provide.
clf = Ridge(alpha=1)
clf.fit(latent_val[idxs1][samples1],delta[idxs1][samples1])

Thus, the clf will fit the value from [1,20]

It is different than the claim "auditing algorithm A : (X × [−1, 1])m → [−1, 1]X"

Can you explain why you choose to use delta rather then directly use the residual (probs - y_val)?