Open-source cognitive architecture for transparent, self-improving AI systems.
We build the infrastructure layer — the execution engine, the routing primitives, the memory architecture, the compliance backbone — for AI systems that operate in environments where failure has real consequences.
Everything published here is Apache 2.0. The research is open. The architecture is open. The audit trails are the point.
Three preprints establishing the formal foundations of the Lár architecture:
| Paper | DOI | Published |
|---|---|---|
| Universal Cognitive Routing: A Sufficient and Extensible Cognitive Contract for Autonomous Multi-Model Systems | 10.5281/zenodo.20278775 | May 2026 |
| Divergence Is Not Noise: Multi-Stream Routing Without Modal Fusion and the Safety-Learning Equivalence | 10.5281/zenodo.20278781 | May 2026 |
| Architecture Is All You Need: Empirical Validation of the Divergence-Routing Self-Improving Loop | 10.5281/zenodo.20419182 | May 2026 |
The central result: the invariants that make a routing system safe are mathematically identical to the invariants that make inter-model disagreement a valid, label-free training curriculum. Safety and learning are not a trade-off. They are the same mechanism.
snath-ai/lar — The Glass-Box Agent Engine
Deterministic, define-by-run graph execution. 13 EU AI Act compliance primitives. Three HMAC-signed audit artefacts on every run. 21 CFR Part 11 aligned. The only open-source agent framework that produces forensic evidence a regulator can actually inspect.
snath-ai/Lar-JEPA — Ten-ABC Cognitive Architecture
Ten Abstract Base Classes. 33 formal invariants. AbstractDivergenceRouter
(V1–V6) routes by geometric divergence between independent latent streams —
content-blind by invariant. Domain isomorphism proven across crystal physics,
geophysics, computer networks, medical imaging, and quantitative finance.
One architecture. No domain-specific modification required.
snath-ai/DMN — Bicameral Memory Architecture
A biologically-inspired Default Mode Network for persistent AI memory. 3-tier architecture (Hot/Warm/Cold). Background consolidation during idle periods. The Dream Loop solves catastrophic forgetting architecturally, not with prompt tricks.
Route → Flag disagreement → Build D_hard curriculum → Train LoRA adapter → Improve routing → repeat
No human labels in the critical path. The routing invariants that prevent bad decisions are the same invariants that identify the most valuable training examples. The system improves itself using its own uncertainty as the curriculum.
- Never use an LLM to police another LLM. Use code.
- An approval is a cryptographic signature of a specific state, not a flag.
- Determinism is not a constraint. It is the product.
- Domain-agnostic by proof, not by claim.
- Issues — bugs and feature requests
- Discussions — architecture patterns and ideas
- Docs — documentation and guides
Apache 2.0 · snath.ai · docs.snath.ai