[Engineering]6 min read

How we verify outputs on untrusted GPUs

Anyone with a GPU can join the network, so no single node is trusted. Staking, sampled re-execution and signed receipts are how a response earns its place on your bill.

Sam WhitfieldProtocol Research

A decentralized network is only useful if you can trust what comes back. Topo nodes are run by independent operators, so we assume any one of them might be misconfigured, overloaded or dishonest, and we design so that none of those cases reach your bill.

Stake first, serve second

Every operator stakes TOPO before receiving traffic. The stake is collateral for every response the node signs. A mismatch found later slashes it, and the affected request is re-routed and served again at no cost to you.

Sampled re-execution

  • A sample of requests is re-run by independent attestors on different hardware.
  • Fixed decoding settings make outputs comparable token by token.
  • Three of three attestors must agree before a sampled response settles.
  • Sampling rates rise for new nodes and for nodes with recent misses.

“Verification is not a feature on top of the network. It is the reason a stranger’s GPU can serve your production traffic.”

receipt.json
json
{"request_id": "req_7Hq2…f91","node": "eu-west-2","attestors": "3/3","settled": "0.0041 USDT","tx": "0x51c…9ab"}

Every response carries a receipt like this one, signed by the serving node. You can inspect it in the dashboard or check it against the chain yourself.