Sieveworks
Verifiable distributed search compute — live at sievework.com.
▶ Watch the pitch — 1:46. What it is, why verification is the hard part, and a run on devnet end to end.
Fund a brute-force search job; contributors run chunks of the search space on their own hardware — in the browser via WebAssembly, a native CLI, or a native desktop app (CPU, with a GPU kernel for hash-grind) — and get paid per verified chunk on Solana (devnet). Every discovery is permanently attributed on-chain to whoever found it. Verification is deterministic and publicly reproducible at ~0.9% measured overhead (at a 10% audit rate), versus the ~200% of replication-based approaches.
New here? ARCHITECTURE.md is the one-page system map — components, the four verification surfaces, the on-chain money/trust flow, and the honest trust boundaries.
The class of problem served: hard to find, easy to check — compact inputs, deterministic evaluation, and a witness anyone can re-verify in microseconds. Launch vertical: Minecraft seedfinding (via cubiomes, MIT). The platform itself is game-agnostic.
How cheating is prevented without redundant execution
- Extremum reframing — every chunk answers "what is the highest-scoring seed in this range?" with a witness seed. There is no fakeable "found nothing."
- Merkle commit-and-challenge — workers commit to per-bucket results with a single root; the coordinator challenges random buckets and recomputes them.
- Honeypot seeds — the coordinator secretly knows answers scattered across the space; under-reporting is detectable and undetectable-by-workers.
- Stake and slash — enforcement is live. To lease paid work a worker must post a one-time global bond; the coordinator re-checks the bond at submit time; and withdrawing it requires the coordinator to co-sign (the on-chain unstake-lock), which it refuses while any lease or challenge is outstanding — so a caught cheat can't pull the bond ahead of a slash. A slashed bond is burned to the incinerator (never to the coordinator or the funder, so neither profits from slashing and a Sybil can't self-refund). With the audit rate at 10%, the burned bond puts fabrication below zero expected value. Free bounties need no bond, so browser onboarding stays frictionless. Honest remaining scope, stated plainly: the coordinator is still a trusted (not decentralized) verifier, and an unclaimed worker balance can delay a funder's escrow reclaim.
- Deterministic re-verification — every claimed record is fully recomputed before on-chain attribution.
The coordinator verifies with the identical WASM artifact workers run, pinned by content hash (worker_spec_hash), making worker/verifier drift impossible by construction.
Repository layout
| Path | Purpose |
|---|
apps/web | Next.js app — bounty board, live job detail, in-browser worker (Vercel) |
apps/coordinator | Fastify service — leasing, verification, SSE, voucher signing (Fly.io) |
apps/cli | Native CLI worker — lease → evaluate → merkle commit → sign → submit |
packages/protocol | The public interface: message schemas, canonical JSON, wallet signatures |
packages/merkle | Commitment tree shared verbatim by workers and the verifier |
packages/worker-core | C wrapper over vendored cubiomes; builds to native binary and WASM |
packages/wasm-runtime | Hash-verifying WASM loader used by browsers and the coordinator |
programs/sieveworks | Anchor program: find attribution, escrow, stake, claim vouchers |
supabase/migrations | Postgres schema with RLS (honeypots and rejection reasons are dark) |
Development
pnpm install
pnpm build # turbo builds packages in dependency order
pnpm --filter @sieveworks/worker-core test # native-vs-WASM determinism test
cd apps/web && pnpm dev # local web against the live coordinator
Built during the Crypto World's Fair (Sep 14 – Oct 12, 2026)
The platform above shipped solo in 7 days (Aug 24–31, tag pre-cwf marks that baseline). Everything after the tag is Crypto World's Fair work — tracked in CHANGELOG.md and specced in plans/cwf/:
- Verification modes — generalizing beyond extremum search:
output_hash recompute-audit (rendering) and checkpointed-trajectory training verification (evolutionary compute)
- Prize bounties — winner-takes-prize for the best verified candidate; an "AI learns to play" demo where the training swarm is paid per generation
- Desktop worker — full multi-core Tauri app; the website becomes onboarding
- GPU execution — wgpu kernel for hash-grind, verified by the unchanged CPU reference
Originally built for the Solana hackathon, August 2026.
License & public good
Sieveworks is open source under the MIT license, built in the open as a
public good. The value is the mechanism — a cheap, publicly reproducible way to
verify untrusted compute (~1% overhead instead of ~200% redundancy) — and it is most
useful if anyone can audit, fork, and build on it. Every verified record ships a
public audit recipe you can re-run against the pinned WASM artifact, so the
verification is checkable by anyone, not just taken on trust. Contributions welcome —
see CONTRIBUTING.md.
