GitHub ↩
Trusted compute · Rust

Built to understand, not to fork.

A Solana-style validator written from scratch in Rust: proof-of-history clock, Tower BFT consensus, Sealevel-style parallel execution, Turbine propagation with real Reed–Solomon erasure coding, CRDS gossip over UDP, a fee-priority mempool and an SBF-inspired program executor. Not a wrapper. Not a fork. Every line mine.

Source on GitHub ↩ Documentation
12
crates in the Cargo workspace
228
tests, all public
~1M
PoH hashes per second
MIT
licensed, fully public
What is inside

Eleven subsystems, all real.

Every cryptographic path is implemented, not stubbed: transactions are Ed25519-verified, votes are signed before they touch state, gossip messages that fail verification are counted and dropped without a reply.

Proof of History

A verifiable delay function: N sequential SHA-256 hashes per tick. Verification re-hashes N times and checks the result matches.

Tower BFT

Fork choice with lockouts — every vote on the same fork doubles the lockout, so switching forks gets expensive. Finality at two-thirds supermajority.

Sealevel-style scheduler

Conflict analysis over read/write account sets. Disjoint transactions batch together; write conflicts serialise. Slot-based with a compute budget.

Turbine + Reed–Solomon

Systematic RS over GF(256): any k of n shards reconstruct the block exactly, or the decoder returns None. Shreds are signed and verified on receipt.

CRDS gossip

Cluster Replicated Data Store with versioning, timeouts, push/pull and ping/pong — every outgoing message signed.

Accounts DB

Append-only store with a DashMap index for O(1) lookups and a Merkle tree over state for a verifiable state root.

Fee-priority mempool

Transaction fees are charged against the real accounts DB even when an instruction later fails — because that is Solana semantics.

Program executor

A unified native dispatcher with a System program (real SOL transfers, account creation) and an SPL-like Token program. No SBF runtime, so an unsupported program fails honestly instead of faking an Ok.

JSON-RPC 2.0

Account queries over a real server, backed by the same state the consensus path commits to.

Run it

Tests first, then the node.

228 tests across the workspace, cargo clippy -D warnings, cargo audit with zero advisories, and CI on Linux and macOS with --locked, plus CodeQL and the OpenSSF Scorecard.

cargo test --workspace
cargo bench -- --bench
What it is not

Research-grade, on purpose.

No SBF runtime, no QUIC (the transport is real UDP), no snapshot loading, and nothing tuned for production throughput. It has never run on mainnet and we do not claim it has — the README says plainly: do not use it for real transactions or staking. What it demonstrates is that the architecture can be held in one head, end to end.