We present VRAM Network, a decentralised protocol for coordinating distributed large-language-model (LLM) training that reduces inter-participant trust requirements through cryptographic verification of gradient quality inside Trusted Execution Environments (TEEs). GPU miners train independently on deterministically-assigned data shards and upload compressed gradients to decentralised object storage (Walrus on Sui). Validator nodes operating within AWS Nitro Enclaves evaluate each gradient by measuring loss improvement on independently-held held-out batches, sign the results with enclave-bound Ed25519 keys, and submit them to the Sui blockchain for on-chain verification and reward emission. We introduce Proof-of-Gradient-Quality (PoGQ), a primitive that converts per-window loss-delta measurements into stable Bayesian skill ratings via the Plackett-Luce OpenSkill model, which determine proportional token reward distribution. Miner storage credentials are protected from on-chain disclosure via the Seal framework (BLS12-381 IBE, t-of-n key-server committee), with the on-chain access policy deployed and client-side integration in progress. Coordination is orchestrator-free: all participants derive computation windows from the Sui blockchain Clock object. The protocol is deployed on Sui Testnet. We report end-to-end protocol verification across a 6-miner / 3-validator testnet run: gradient-submission and enclave-scoring mechanics function as specified, with approximately 200,000x gradient wire-format reduction (Top-K, K=32) and per-window cryptographic overhead under 8 seconds against a 600-second window budget. Quantitative convergence quality versus a centralised baseline is the subject of a follow-on study with the transformer adapter. We explicitly enumerate v0.6 trust boundaries (aggregator, Seal client, storage) and the concrete migrations targeting v0.7.
Sid Berraf (Sun,) studied this question.