Osakra Research

Independent Computational Cognition, Continuous Dynamics & Relational Neuro-Symbolic Architectures

About Osakra Research

Osakra Research is an independent computational research effort exploring continuous-time dynamical systems, latent representation manifolds, and vector-symbolic architectures in compact language models.

Rather than scaling parameter counts across massive distributed clusters, this project investigates targeted architectural questions under strict consumer hardware constraints (< 5.8 GB active VRAM on a single consumer GPU): can compact open backbones deliberate directly within continuous hidden vector manifolds (h ∈ ℝᵈ) and maintain algebraic invariants without relying solely on verbose textual token emission?

This line of inquiry forms the basis of the N.O.R.N. architecture (Neuro-symbolic ODE & Relational Networks), a dual-tier framework pairing discrete autoregressive deliberation with continuous associative memory engines.

Core Research Areas

• Relational Graph Holographic Reduced Representations (HRR): Binding topological graphs, causal transition triples, and algebraic structures directly in a 1024-dimensional continuous vector space using complex Fourier-domain circular convolutions F⁻¹(F(a) ⊙ F(b)), mitigating representation drift and attention dispersal across long-horizon deduction.

• Continuous-Time Dynamical Modulation (LTC-ODEs): Utilizing Liquid Time-Constant (LTC) ordinary differential equations and leaky integrators to govern continuous latent state evolution between discrete token transitions.

• Deterministic Agentic Tool Evaluation: Auditing model capabilities on live subprocess code execution (the 250-task Agentic Sandbox) to differentiate genuine, typed, modular software engineering from fragile single-line scripting shortcuts that collapse on complex algorithms.

• Weight-Space Consolidation & Inference Efficiency: Investigating the mechanics of fused adapter checkpoints to understand why consolidated parameter representations yield sub-second inference speeds (0.38s/task) compared to unmerged deliberation paths.

Releases & Research Checkpoints

• Project Norn V18 (9B Hybrid HRR): Our 9B hybrid reasoning checkpoint quantized to Q4_K_M GGUF (< 5.8 GB VRAM, 100% local GPU offload). Integrates a native 4096-dimensional hidden projection, a 1024-dimensional 4-channel Relational Graph HRR engine (norn_biology_proj_v18.pt), and evaluation against our bundled 1,250-benchmark multi-pillar suite (81.8% Micro / 80.6% Macro accuracy; 94.0% live agentic tool accuracy).

• Project Norn V17 (4.5B Standalone Merged Dual-Tier): A consolidated 4.5B parameter dual-tier cognitive system. Bundles complete FP16 safetensors, 4-bit GGUF, and original LoRA adapter weights, demonstrating ultra-fast 0.38s/task agentic tool execution.

• Project Norn V15 (4.45B Auto-Router Prototype): The foundational proof-of-concept testing continuous latent Chain-of-Thought recurrence (k ∈ [2, 16]) and associative memory seeding on top of a 4-bit quantized base.

Grounded Scope & Reproducibility

• Single-Author Research: This is focused, exploratory academic research. These models are designed to probe continuous representations, invariant tracking, and compact reasoning, and do not compete with 70B+ proprietary models on open-domain trivia or massive encyclopedic memorization.
• Full Community Replication: Every release bundles raw evaluation suites, deterministic scoring harnesses, audit runners, and transparent execution traces directly in the repository under the Apache-2.0 license to facilitate independent community replication and mechanistic interpretability analysis.