Skip to content

Releases: snath-ai/Lar-JEPA

v0.2.0 — The Nervous System: Universal Model Routing for LLMs, JEPAs, and Every Cognitive Architecture That Follows.

18 Apr 22:47
@axdithyaxo axdithyaxo
v0.2.0
5410133

Choose a tag to compare

View all tags
v0.2.0 — The Nervous System: Universal Model Routing for LLMs, JEPAs, and Every Cognitive Architecture That Follows. Latest
Latest

v0.2.0 — The Nervous System

Universal Model Routing for LLMs, JEPAs, and Every Cognitive Architecture That Follows

Released: 18 April 2026
Author: Aadithya Vishnu Sajeev / Snath AI
Repository: snath-ai/Lar-JEPA
Prior Art Preprint: DMN v3.0 "The Dream Loop" (Zenodo, April 2026)

What Changed

This release reframes and extends the Lár-JEPA repository from a JEPA-specific
orchestration showcase into its correct architectural identity: a universal
cognitive routing nervous system capable of routing any model type —
large language models (LLMs), JEPA world models, diffusion models, state-space
models (SSMs), graph neural networks (GNNs), and any architecture that follows
— as first-class, equally routable nodes within the same deterministic Lár
execution spine.

New: AbstractCognitiveNode — The Universal Node Interface

File: core/interfaces.py

Introduced AbstractCognitiveNode as the root interface for all cognitive
components in the Lár routing graph. Any model type implements three methods:

def encode(self, input_signal: Any) -> Any    # Signal → internal context
def forward(self, context: Any) -> Any        # Core inference pass
def decode(self, representation: Any) -> Any  # Output → GraphState signal

AbstractManifold (the JEPA-specific world model base) is now a subclass of
AbstractCognitiveNode, placing JEPAs and LLMs at the same level of the
type hierarchy. AbstractEntropicRouter is retained as the JEPA-specific
routing spine component.

New: AbstractContextBridge — Cross-Modal Signal Adapter

File: core/interfaces.py

Introduced AbstractContextBridge: a stateless signal conduit that adapts one
AbstractCognitiveNode's output SignalType to another node's expected
encode() input. This enables:

  • LLMs attending to JEPA latent predictions — bridge converts
    LATENT_EMBEDDING → TEXT or prefix embedding for LLM context window.
  • JEPAs conditioning on LLM semantic embeddings — bridge converts
    TEXT / GRAPH_STATE → manifold-compatible context vector.
  • Any future cross-modal composition — new bridge declares
    source_signal_type and target_signal_type, implements bridge().

Bridges hold no model weights. They are pure format adapters — the synaptic
connectors of the nervous system.

New: ModelType, SignalType, CompositionPattern Enums

File: core/types.py

ModelType

All model types are first-class citizens in the Lár routing graph:

Value Description
LLM Large Language Model (any provider via LiteLLM)
JEPA Joint-Embedding Predictive Architecture
DIFFUSION Diffusion / score-matching model
SSM State Space Model (Mamba, S4, RWKV, etc.)
GNN Graph Neural Network
CLASSICAL Deterministic non-neural function
HYBRID Heterogeneous composite
FUTURE Any architecture not yet named or invented

The FUTURE value is a formal architectural claim: any model type that does
not yet exist will implement AbstractCognitiveNode and become routable
without modifying the spine.

SignalType

Typed signal vocabulary for cross-node communication:
TEXT, LATENT_EMBEDDING, LGSL_SPEC, GRAPH_STATE, STRUCTURED_DATA,
TENSOR, IMAGE, AUDIO, GRAPH, DISTRIBUTION.

CompositionPattern

Eight named cross-modal composition design patterns:
LLM_ROUTES_JEPA, JEPA_INFORMS_LLM, PARALLEL_HOMOGENEOUS,
PARALLEL_HETEROGENEOUS, HIERARCHICAL, CROSS_ATTENTION,
SEQUENTIAL_PIPELINE, RECURSIVE_SELF_IMPROVEMENT.

New: JEPA_DMN_Consolidation_Node — Live JEPA→DMN Bridge

File: dmn_integration/consolidation_node.py

The consolidation stub has been replaced with a functional bridge connecting
the Lár-JEPA execution layer to the Default Mode Network (DMN) episodic memory
store (Hippocampus / ChromaDB).

write_trajectory_heuristic(trajectory_log, embedding)

  • Called after AbstractEntropicRouter returns COMMIT_TRAJECTORY
  • Converts the committed JEPA trajectory to a text summary + metadata record
  • Writes to ChromaDB long_term_memory collection with memory_type: episodic
    (DMN v3.0 memory typing for the Consolidation Loop)
  • Also writes to the DMN JSON journal for narrative history

recall_heuristics(query, max_results)

  • Retrieves semantically relevant past JEPA heuristics to warm the current
    planning cycle
  • Calls Hippocampus.recall() via ChromaDB vector similarity search

Both operations degrade gracefully if the DMN Hippocampus is unavailable.
The JEPA execution spine is never blocked by DMN availability.

This completes the first pass of the Consolidation Loop:

JEPA prediction
  → EntropicRouter: COMMIT_TRAJECTORY
  → JEPA_DMN_Consolidation_Node.write_trajectory_heuristic()
  → Hippocampus.save_memory()
  → ChromaDB long_term_memory + dreams.json
  → (idle cycle) Dreamer consolidates → semantic memory
  → (future) BrainNode LoRA trains on confirmed routing decisions

New: ARCHITECTURE.md — The Nervous System Design Document

File: ARCHITECTURE.md (top-level)

Formal specification of the Lár-JEPA framework as a universal cognitive
routing nervous system. Documents:

  • The three-layer architecture (Routing Spine / Signal Bridge / Cognitive Nodes)
  • All ModelType values with example architectures and output signal types
  • Six named composition patterns with ASCII dataflow diagrams
  • BatchNode concurrency for homogeneous and heterogeneous model ensembles
  • Forward compatibility guarantee via ModelType.FUTURE
  • The complete prior art chain (5 DOIs)

Updated: core/__init__.py

All new types and interfaces are now exported from the core package:

from core import (
    AbstractCognitiveNode, AbstractManifold,
    AbstractContextBridge, AbstractEntropicRouter,
    RouteDecision, ModelType, SignalType,
    CompositionPattern, StructuralImpasseError,
)

Updated: README.md — Reframed

Before After
Title: "Orchestrating World Models (A Post-LLM Architecture)" "The Universal Model Routing Nervous System"
Section: "The Post-LLM Paradigm (JEPA)" "The Universal Routing Problem"
Framing: JEPA replaces LLMs Framing: Lár routes LLMs AND JEPAs and everything after
Footer: "Conceptual Testbed and Showcase" Accurate description of implemented components

Prior Art Status

The architectural mechanisms introduced in this release are documented as
prospective prior art in:

Sajeev, A. V. (2026). DMN v3.0 — The Dream Loop: Memory Consolidation,
Familiarity-Weighted Retrieval, and the Learned Graph Executor in the Lár
Default Mode Network. Zenodo. (DOI pending publication 18 April 2026.)

The AbstractCognitiveNode universal routing interface, AbstractContextBridge
cross-modal composition layer, ModelType.FUTURE forward-compatibility
placeholder, and the six named CompositionPattern values were all designed
and committed by Aadithya Vishnu Sajeev under Snath AI prior to any employment,
corporate collaboration, or externally funded research. This release constitutes
the software implementation record complementing the above preprint.

Summary

The industry is building the Brain (LLMs, JEPAs).
We are building the Nervous System (Lár + DMN).

v0.2.0 is the release where that sentence became architecturally precise.

Assets 2
Loading