Lagrange opened the source code for DeepProve on June 3, releasing a system for producing cryptographic evidence that an AI inference ran as specified. The announcement includes the proving software, verification code and examples, allowing developers to inspect how the system connects a model’s inputs to its output.
The purpose is narrower than checking whether an answer is correct in the everyday sense. A proof can establish that a specified computation produced a result. It cannot establish that the model’s knowledge was accurate or that its recommendation was sensible.
Verification concerns the computation
DeepProve uses zero-knowledge machine learning, often shortened to zkML. The system is designed to let a verifier check execution without needing the private model weights or input data. Lagrange also describes verification on a blockchain.
The released stack includes circuits, a prover and a verifier, with support described for ONNX, safetensors and GGUF model formats. The announcement names examples using GPT-2, Gemma-3, classifiers and vision models, while placing Llama-class model work in active development.
Those examples define a starting point for evaluation. Format support should not be interpreted as proof that every model file using that format will run unchanged. Developers still need to check the operations and model configuration their particular workload requires.
For blockchain applications consuming model output, proof verification can provide evidence about the execution step. It does not eliminate the separate questions of which model the application trusts and what authority the output receives.
Open code makes the performance claims inspectable
Lagrange reports substantial proving and verification improvements over earlier zkML systems, alongside published methodology and benchmarks. These remain the provider’s reported results, rather than measurements from an OSBBD test.
The useful next step for a developer is to compare a supported model and input against the reference implementation, including the resources needed to create the proof. Fast verification alone says little about the cost of generating it.
A team also needs ordinary AI evaluation to assess answer quality. Execution evidence and answer evaluation address different failures: the first asks whether the stated computation occurred, while the second asks whether its result is useful for the task.
