A neural network can't fit in a block.
Millions of float weights would need more than 15 million gas for one inference. So today's AI stays off-chain, and contracts trust a signed message that arrives a block too late.
ZK-ML proves it, slowly.
The model still runs off-chain. A prover spends 10 to 300 seconds, then the chain pays around 350,000 gas to verify. Correct, but never inside the attack's own transaction.
Werracle stores a coordinate, not a model.
Weights are recomputed on demand from the escape dynamics of z → z² + c, in Q16.16 integers. One SLOAD reads the whole brain for 100 gas. Nothing leaves the chain.