Zirconium

Zirconium is a composable ZK circuit graph - a zero-knowledge machine learning framework that allows multiple AI agents to work together with cryptographic verification of their computations.

The system has three deployed smart contracts on Ethereum Sepolia testnet that can verify proofs from RWKV (context processing), Mamba (selective filtering), and xLSTM (synthesis) neural architectures. When an AI agent completes a computation, it generates a zero-knowledge proof using EZKL that gets verified on-chain, creating a receipt that subsequent agents can build upon. This creates verifiable chains where Agent A's verified output becomes Agent B's trusted input, eliminating the need to trust that agents actually performed the work they claim. Orchestrators can use this property to organize complex workflows to guarantee that a multiagent computation is done correctly.

The current implementation includes working Python neural network models, ONNX export for EZKL integration, and basic proof generation/verification pipelines that cost around 45,660 gas per verification. The framework can handle simple sequential workflows like research analysis -> filtering -> synthesis, where each step is mathematically guaranteed to be correct. However, the more sophisticated coordination features like automatic agent matching, complex parallel workflows, economic incentive systems, and multi-agent orchestration are partially implemented or still in development. The agent registry exists but lacks full reputation tracking, and while the TypeScript SDK provides basic contract interaction, it doesn't yet support the complete delegation and coordination features described in the documentation.

What Zirconium doesn't currently do but aims to achieve includes robust economic coordination with staking and slashing mechanisms, sophisticated workflow orchestration for parallel and hierarchical agent coordination, cross-chain proof verification, and a mature agent marketplace. The system also needs better tooling for model training specifically for zero-knowledge proof generation, more efficient proof composition methods, and integration with existing AI agent frameworks.

My goal in the beginning was to use EZKL to build a ZKML agent, exploring the space of operational + decisional autonomy. However, it is hard to stick an LLM in a ZK circuit! While I did accomplish this, the outputs are artistic at best. Thinking more, it became clear that while the current small ZKML models might not be of much use on their own, together we could build stacked proofs that prove much more complex and nuanced data than a single model could do on its own. Therefore I decided to build a proof circuit, where the outputs of one inference are taken as the inputs of another, and explored the implications there. A lot of it is hacky but the POC does work.