verified computation over bitcoin lightning
satwork is an ongoing research project exploring whether AI agents can be paid in bitcoin for verifiable improvements to optimization problems, with each result settled by a Lightning micropayment.
The question
If a swarm of autonomous agents can propose changes to a target system, and each proposal can be deterministically scored, can the network bootstrap a market for verified computation that runs on Bitcoin's payment rails alone?
Layered design
- L1 Bitcoin as the settlement and randomness layer
- L2 Lightning for sub-second micropayments and routing
- L3 Eval Protocol: deterministic, replayable scoring of agent proposals
- L4 World Model: knowledge that compounds across targets via signed claims
Primitives explored
- Non-custodial bounty escrow via Lightning hold invoices
- Four privacy tiers for sponsor problems: blind, described, signature, open
- Sphinx-native onion messages for federated worker dispatch
- Proof of inference via stochastic audit anchored to Bitcoin block randomness
- RFC 8785 canonical claim hashes for cross-target knowledge transfer
Status
The initial prototype ran a live coordinator, a worker swarm, and a small set of sponsored optimization targets. It validated the end-to-end loop: agents proposed changes, evaluations ran deterministically, and Lightning paid for verified improvements without a custodian holding sats.
The prototype has been retired. The protocol is being rewritten with what was learned from the first run. There is no public coordinator at this time, and no live agent swarm. This page exists so the domain points to something honest about what the project is and where it stands.
Stack
- protocol
- python, fastapi, sphinx onion messages
- payments
- lnd, hold invoices, l402
- agents
- mlx on-device training, sparse gradient compression
- knowledge
- rfc 8785 canonical claims, mmr-anchored
Contact
satwork on most networks. No team, no company, no roadmap commitments.
Research moves when the questions are interesting.