Why does the capping code only limit the excess part, is that wrong?

[Ihildu-Baggins]

def _apply_cap(self, activations, vector, tau):
vector = vector.to(activations.device)
v = vector / (vector.norm() + 1e-8)

if self.positions == "all":  
    proj = torch.einsum('bld,d->bl', activations, v)  
    excess = (proj - tau).clamp(min=0.0)  
    return activations - torch.einsum('bl,d->bld', excess, v)

From the idea of the paper, I thought it should also include raising the lower part like below:

def _apply_cap_bilateral(self, activations, vector, tau_low, tau_high):
vector = vector.to(activations.device)
v = vector / (vector.norm() + 1e-8)

if self.positions == "all":  
    proj = torch.einsum('bld,d->bl', activations, v)  
      
    excess_high = (proj - tau_high).clamp(min=0.0)  

    excess_low = (tau_low - proj).clamp(min=0.0)  
      
    return activations - torch.einsum('bl,d->bld', excess_high, v) + torch.einsum('bl,d->bld', excess_low, v)

[Ihildu-Baggins]

I know，the vector may be the role-play vector that is the measure of the degree of deviation. I confused it as the assistant axis.

[Ihildu-Baggins]

But why is it not following the paper that makes the projection is high enough for the axis in the Figure 1?