Causal-native AI runtime in Rust - Pearl SCM · Active Inference · bio-inspired temporal stack
Numen is a research AI runtime that reasons causally, not statistically. Unlike LLMs, it builds and queries an explicit causal model of the world (Pearl SCM), combines it with Active Inference (Friston/FEP), and routes computation through a bio-inspired temporal hierarchy - from reflexive cache to deliberate causal reasoning.
Status: early research. Not ready for use.
- Causal engine first: explicit Pearl SCM (do-calculus, counterfactuals) - not correlation
- Active Inference: beliefs updated by minimizing variational free energy, not token prediction
- Temporal hierarchy: 4-layer stack (reflex → salience → contextual → deliberate) inspired by biological neural architecture
- Async consolidation: hippocampus-to-neocortex memory consolidation, solving catastrophic forgetting
- Rust-native: single binary, self-hosted, no cloud dependency
- Phase 1 - SCM minimal (Rust, petgraph)
- Phase 2 - LLM layer (mistral.rs) + causal conditioning
- Phase 3 - Active Inference loop + temporal stack
- Phase 4 - Benchmark (InterveneBench, ARC-AGI-3)
Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
Copyright (c) 2026 Rwx-G
Free to share and adapt with attribution. Commercial use requires explicit agreement. Contact: romain@rwx-g.fr - https://creativecommons.org/licenses/by-nc/4.0