Lar Framework: Deterministic Metacognition and Joint-Embedding Predictive Architecture (JEPA) Integration

Read full text Summarise with VIRAT
Authors: Aadithya Vishnu Sajeev
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This preprint formally defines the theoretical foundation and architectural primitives of the Lár Engine, an open-source (Apache 2.0) deterministic graph-state orchestration framework designed to resolve the autoregressive unpredictability and catastrophic forgetting inherent in standard Large Language Model (LLM) agents. We formally classify the Lár Node Ecosystem, including generative (LLMNode), computational (ToolNode), conditional (RouterNode), scaling (BatchNode/ReduceNode), and regulatory (HumanJuryNode) primitives, which enforce absolute execution constraints via a proprietary JSON Blueprint grammar. Furthermore, this publication establishes explicit prior art for two novel orchestration paradigms: 1. Structured Metacognition & Fractal Agency: The theoretical formulation of self-evaluating autonomous loops (HallucinationDetectorNode, SelfAwarenessNode) operating identically to an external prefrontal cortex, scoring generative outputs against rigid environmental telemetry prior to state mutation. It additionally defines the mathematical basis for Fractal Agency (DynamicNode), enabling agents to recursively synthesize, validate, and spawn isolated sub-graph topologies at runtime. 2. Lár-JEPA (Prediction via Embedding Space): The theoretical adaptation of Yann LeCun's Joint-Embedding Predictive Architecture explicitly for deterministic graph orchestration. By shifting the computational burden from generative token prediction to the routing of high-dimensional latent state embeddings (System 2 routing), the framework executes safety vetoes and abstract action simulation before interacting with a physical environment. This document serves as permanent, cryptographic prior art establishing Aadithya Vishnu Sajeev as the sole inventor of these advanced architectural methodologies within the snath-ai/lar and snath-ai/lar-JEPA ecosystems.

Fields of Study & Subjects

AI-based Problem Solving and Planning Artificial Intelligence Applications Psychiatry Mental Health Neuroscience

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
27 March 2026
Language
EN
Document type
Preprint