AI ATTESTATION
Check the
computation.
motiivi checks a model’s computation with a smaller verifier derived from its weights.
Model computationmotiivi verifier
Linear / ReLU prototype · schematic
How it works
Two ways to compute
the same value.
The model calculates each full layer output. motiivi uses reduced weights to calculate a single value from the same input. That value can be compared with a projection of the layer’s output.
The current prototype uses linear and ReLU layers. These checks concern the computation; the answer itself still needs evaluation.
From the model’s output
(hW + b)r
=
From the reduced weights
h(Wr) + br
h is the layer input. W and b are its weights and bias. The vector r reduces the output to one value.
The small branch is derived from the model’s weights, without training a second model.