/10 The takeaway. A reliability failure that made small reasoning models unusable on hard tasks turns out to be fixable at one token, in a few hours, without touching what the model knows. For anyone…