Leviathan

Leviathan is a Solana-coordinated training network for open language models. Anyone with a GPU joins by posting a bond, earns Proof of Gradient for contributions that survive verification and loses the bond if they lie: the chain carries commitments, audits and money, the peer to peer mesh carries compressed tensors, and the model belongs to the network that trained it. Its network substrate, leviathan-net, forks Nous Psyche and implements the bond, audit, slash, committee and appeals layers the upstream left unimplemented. The protocol is live on Solana devnet; bonds with real value and a mainnet deployment wait on an external audit.

What it does

• Bonded participation: a deposit, a withdrawal challenge window and slashed collateral forfeited to the run vault
• Audit lottery: verifier assignments derived from the round seed, deterministic and replayable
• Replay verifier with a calibrated tolerance band and sha256 fraud proofs matching the on-chain commitment
• Committee vote and appeals court: a two thirds quorum slashes, and an overturned conviction costs the verifiers their bonds
• Robust aggregation by centered clipping with far outlier excision
• Coordinator, authorizer and treasurer programs live on Solana devnet, exercised end to end