LAb notebook
Blogs
Updates from the lab as work happens — new results, deployment notes, and conference recaps.
Ledgermind lets a model provider prove that an output came from a claimed model and input, without disclosing the model weights themselves. We use succinct zero-knowledge proofs generated alongside inference, then verified by a lightweight on-chain contract in milliseconds.
The core challenge has been proof generation cost: naive zk-SNARK circuits for neural network inference are orders of magnitude slower than the inference itself. Our current prototype gets within 40x for a 7B-parameter model, down from over 4,000x eighteen months ago, using custom arithmetic circuits for attention layers.