How it works

Every private run goes through three steps.

1. 01

   Split

   Your words never travel whole. The model's activations are cut into two secret shares. Each node computes on its own share and never holds the real tensor.
2. 02

   Route

   Contributed GPUs are pooled into swarms that serve open models. Each node runs a block of layers and passes its share on to the next.
3. 03

   Measure

   One node's view is scored for how much of the prompt it can recover. A run only counts as private if that number sits at chance.

What a request looks like

A prompt enters the network at an entry node, passes through a chain of nodes that each run a contiguous block of layers, and leaves through an exit node that produces the answer. In the reference setup, four nodes each run 16 layers of a 64-layer model.

prompt → Entry (layers 0–15) → Node 2 (16–31) → Node 3 (32–47) → Exit (48–63) → answer

Without protection, every hop sees real activations. With PRINET, each hop holds one share, and alone a share is noise. See Secret-shared activations for the details.