Files
consensus/pkg
zeekayandHanzo Dev c98bbf4e9a One proof implementation, over whichever message the caller has
A node-bound proof is one message among several: a committee entitles a
validator ON A CHAIN and signs the chain into its proof, so it has a
message of its own. It could only reach the same encoding and
possession clauses by restating them, and a second statement of what a
canonical key is is a second answer waiting to differ.

The general pair is the implementation now and the node-bound pair is
it with one message, so a registration and a committee are checked by
the same rules or by none. Admission is the same door for the same
reason.

Co-authored-by: Hanzo Dev <[email protected]>
2026-09-08 09:53:52 -07:00
..

Lux Consensus Source Code Organization

This directory contains implementations of the Lux consensus framework in multiple programming languages. Each language implementation provides the same consensus algorithms with language-specific optimizations and idioms.

Directory Structure

src/
├── c/          # C implementation (high-performance, minimal overhead)
├── cpp/        # C++ implementation with MLX extensions (GPU acceleration)
├── go/         # Go implementation (production blockchain integration)
├── python/     # Python implementation (research and prototyping)
└── rust/       # Rust implementation (memory-safe systems programming)

Language Implementations

Go (/src/go/)

  • Status: Production-ready
  • Use Case: Core blockchain node integration
  • Import Path: github.com/luxfi/consensus (via go.work at root)
  • Features: Full integration with Lux node, concurrent processing

C (/src/c/)

  • Status: Data structures only — not real consensus (per SDK audit)
  • Use Case: Embedded systems, performance-critical applications
  • Features: Minimal dependencies, ZeroMQ networking, SIMD optimizations

Rust (/src/rust/)

  • Status: FFI wrapper around the C library — inherits C limitations, not native consensus (per SDK audit)
  • Use Case: Memory-safe systems, async applications
  • Features: Zero-cost abstractions, async/await, compile-time guarantees

Python (/src/python/)

  • Status: Research/Development
  • Use Case: Prototyping, research, data analysis
  • Features: NumPy integration, ML frameworks, visualization

C++ (/src/cpp/)

  • Status: Development
  • Use Case: High-performance with GPU acceleration
  • Features: MLX extensions, template metaprogramming, SIMD/GPU support

Building

From the repository root:

# Build all implementations
make build

# Build specific language
make build-go
make build-c
make build-rust
make build-python
make build-cpp

Testing

From the repository root:

# Test all implementations
make test

# Test specific language
make test-go
make test-c
make test-rust
make test-python
make test-cpp

Consensus Engines

All implementations support these consensus engines (Quasar family):

  1. Wave - Threshold voting with fast probabilistic consensus (production)
  2. Nova - Linear chain consensus (wraps Ray sub-protocol)
  3. Nebula - DAG consensus with safe-prefix commit (wraps Field sub-protocol)
  4. PostQuantum - Quantum-resistant consensus with BLS + Pulsar + ML-DSA
  5. Chain - Linear chain consensus for ordered blocks
  6. DAG - Full directed acyclic graph consensus

Performance Benchmarks

Implementation Votes/Second Memory Usage Latency
C 14,000+ < 10 MB < 1ms
Rust 13,500+ < 15 MB < 1ms
Go 12,000+ < 20 MB < 2ms
C++ (w/ MLX) 15,000+ < 25 MB < 1ms
Python 5,000+ < 50 MB < 5ms

Note. These per-implementation votes/sec figures are unverified and, per the SDK audit, several measure data-structure insertion on stub/wrapper SDKs (C = data structures only, Rust = FFI wrapper over C, C++ = Wave-only stub); only the Python SDK implements real consensus. Treat as illustrative, not measured consensus throughput. See .github/workflows/README.md and the "Honest Assessment" block in LLM.md.

Protocol Compatibility

All implementations use the same binary protocol for network communication:

┌─────────────┬────────────┬──────────┬───────────┬──────────┐
│ Engine Type │ Node ID    │ Block ID │ Vote Type │ Reserved │
│ (1 byte)    │ (2 bytes)  │ (2 bytes)│ (1 byte)  │ (2 bytes)│
└─────────────┴────────────┴──────────┴───────────┴──────────┘

This ensures interoperability between different language implementations.

Development Guidelines

  1. Consistency: All implementations should provide the same consensus guarantees
  2. Testing: Each implementation must have comprehensive test coverage
  3. Documentation: Language-specific documentation in docs/{language}/
  4. Performance: Optimize for language strengths while maintaining correctness
  5. Compatibility: Maintain protocol compatibility across all implementations

Contributing

When adding features or fixes:

  1. Implement in the reference implementation (Go)
  2. Port to other languages maintaining consistency
  3. Add tests for all implementations
  4. Update documentation

See CONTRIBUTING.md for detailed guidelines.

License

Copyright (C) 2019-2025, Lux Industries Inc. All rights reserved.