True-Virtual-Biological-Plexus

TVBP-Omni (formerly TVBP-AIO): True Virtual Biological Plexus

TVBP-AIO is a paradigm-shifting 475-million parameter neural network designed not as a linguistic text generator, but as a fully simulated biological brain for virtual entities (e.g., NPCs, game AI, virtual organisms). By fusing State Space Models (Mamba), Sparse Mixture of Experts (MoE), and biologically accurate endocrine/epigenetic systems, TVBP allows virtual creatures to feel, adapt, and evolve rather than just following hardcoded behavioral trees.

---

🧠 Core Architectural Innovations

TVBP-AIO departs from traditional Reinforcement Learning (RL) agents by simulating real anatomical and neurochemical pathways:

• State Space Models (SSM - Mamba): Replaces traditional LSTMs for infinite-horizon memory, representing the Hippocampus and working memory. Memory retention is dynamically gated by the organism's Encephalization Quotient (EQ).
• Affective Mixture of Experts (Sparse MoE): Acts as the Amygdala and emotional center. Instead of standard tokens, this MoE routes sensory data to emotional "Experts" (Fear, Aggression, Curiosity, Mating).
• Frontoparietal Attention Network: Simulates logical reasoning and spatial attention. This network is heavily modulated (damped) in lower-IQ organisms.
• Hormonal & Visceral System: The brain takes real-time inputs of Cortisol, Adrenaline, Melatonin, and Glucose. Hormone levels chemically alter the neural weights (FiLM modulation) and trigger epigenetic trauma responses.
• Glymphatic Clearance & Sleep: A simulated sleep cycle physically zeroes out weak synaptic connections (plasticity < 0.05), simulating biological sleep-dependent memory consolidation and synaptic pruning.

---

🧬 Phylogenetic DNA & Taxonomy

Every entity powered by TVBP-AIO is instantiated with a 16-dimensional Phylogenetic DNA vector (identity_input), eliminating the need for different models per species. The DNA tensor dictates the physical and cognitive limits of the brain at inference time:

• Dim 0 (Vertebrate vs Invertebrate): Determines the presence of a centralized spinal cord vs decentralized ganglia.
• Dim 1 (Thermal Dependency): Endothermic (warm-blooded) vs Ectothermic (cold-blooded). Cold-blooded organisms experience direct motor suppression relative to environmental temperature drops.
• Dim 2 (Taxonomic Class): A continuous coordinate mapping (0.0=Mammal, 0.2=Avian, 0.4=Reptile, 0.6=Amphibian, 0.8=Fish, 1.0=Invertebrate).
• Dim 3 (Diet): Carnivore (high aggression, predatory visual focus) vs Herbivore (high flight response, peripheral threat detection).
• Dim 4 (Cognitive Bottleneck / EQ): Encephalization Quotient (0.0 - 1.0). Controls the severity of Amnesia (hippocampal decay) and Synaptic Noise (instinct overriding logic).

The EQ Bottleneck (Encephalization Quotient)

To prevent a 475M-parameter brain from making a spider as intelligent as a human, TVBP applies physical limits based on Dim 4 of the DNA. Lower EQ causes immediate memory decay (Amnesia) and introduces massive Stochastic Synaptic Noise into the reasoning layers, forcing the organism to rely entirely on basal instincts (MoE) rather than logical planning.

---

🧩 The Plugin Ecosystem (Evolutionary Brain Extensions)

To maintain a 12 GB VRAM limit while supporting immense biodiversity, TVBP-AIO dynamically routes sensory data through evolutionary plugins based on the Phylogenetic DNA:

1. MammalianPlugin (~44M params): Features Advanced Mirror Neurons for complex social imitation and Vocal Pack modules for pack-hunting coordination.
2. AvianPlugin (~49M params): Features Aero-Spatial cortices and a Flight MoE for 3D aerial navigation and flocking behavior.
3. ReptilianPlugin (~45M params): Driven by an Ectothermic Basal Ganglia. Heavy thermal dependency; optimal performance requires heat.
4. AmphibianPlugin (~45M params): A hybrid aquatic-terrestrial spatial mapping network.
5. AquaticPlugin (~49M params): Features Hydro-spatial mapping and Electro-Sonar networks for pitch-black deep-sea navigation.
6. InvertebratePlugin (~48M params): Replaces centralized processing with Decentralized Ganglia, Swarm Logic, and UV-spectrum visual processing (for arthropods and mollusks).

---

🔄 Inputs & Outputs

The brain acts as a continuous function mapper for the organism's body:

INPUTS:

• Visual (384): Foveal and peripheral vision, color spectra.
• Audio (16): Frequency bands, threat directionality.
• Sensory (16): Olfactory (smell), tactile, vibration (lateral line).
• Homeostatic (8): Energy, hydration, body temperature, circadian rhythm.
• Hormonal (6): Adrenaline, Cortisol, Oxytocin, Testosterone, Serotonin, Dopamine.
• Social (4): Proximity of pack members, predators, and pheromones.
• Proprioception (26): Continuous feedback loop of the organism's own motor state.

OUTPUTS:

• Motor Action (26): High-dimensional skeletal control, locomotion, vocalization.
• Visceral Response (4): Autonomic nervous system commands (heart rate adjustment, hormone secretion triggers, adrenaline dumping).
• Epigenetic Gamete (22): A mutation tensor passed down to offspring during mating, containing survived trauma markers.

---

🎓 Training Methodology

TVBP-AIO is trained in two distinct phases:

1. Phase 1: Standardized Pre-training (Childhood): The model undergoes millions of simulated biological epochs across diverse phylogenetic DNA. It learns basic causality (e.g., fire = pain = cortisol increase = move away).
2. Phase 2: Human-in-the-Loop & Opponent Processing: Reinforcement learning based on survival metrics (Homeostatic stability). The Opponent Process theory ensures that prolonged Fear leads to compensatory mechanisms (e.g., seeking safety triggers endorphins).

---

🚀 THE FUTURE: Turning LLMs into Virtual Organisms

While TVBP-AIO currently processes raw environmental tensors, the ultimate endgame is to transform this architecture into a Next-Generation Multi-Modal Transformer / LLM Framework.

In the future, TVBP's biological foundations (Hormones, Sleep, Epigenetics) can be fused with a Large Language Model's (LLM) reasoning capabilities. Instead of a chatbot that blindly generates text, we envision an Embodied Linguistic Organism:

• An LLM that actually feels "tired" (Context window decay via Melatonin).
• An LLM that panics (Top-K sampling narrows, Temperature spikes due to Adrenaline).

TVBP is not just a model; it is the bridge between Silicon and Carbon.

---

📚 Project Terminology

The TVBP ecosystem defines several project-specific terms used throughout the architecture.

---

📚 Deep Dive: How It Actually Works

For a deep mathematical and biological breakdown of the internal mechanisms (including Opponent Processing, Glymphatic Synaptic Pruning, Corpus Callosum simulation, and the exact Training Pipeline), please read our Architecture & Mechanisms Whitepaper.

---

🔬TVBP-Omni Additions & Expansions

The architecture has recently been upgraded from the ~330M parameter AIO version to the ~475M parameter TVBP-Omni version. The following major updates are now actively functioning in the PyTorch pipeline:

• Sensory Re-Routing & Contextual Awareness: The environment b_scenario (weather, terrain, time of day) is now strictly encoded directly into the olfactory/sensory input tensor. Organisms "smell" the environment, allowing the affective MoE to select appropriate instinctual responses (e.g., triggering hibernation behaviors when temperature drops).
• Proprioceptive Loop: A direct feedback tensor (v_proprioception) tracks the organism's muscular state from the previous frame, ensuring continuous, smooth animations in game engines and preventing erratic spasming.
• Router Entropy Loss for MoE Collapse Mitigation: The Amygdala (MoE) now utilizes a forced entropy loss floor (1.5 / ln(8)), mathematically preventing "expert collapse". An organism is forced by the gradient descent to explore its entire affective spectrum (seeking, playing, foraging) rather than getting stuck in a pathological state of permanent fear.
• The NeuroplasticityAnnealingEngine: A revolutionary LR scheduler that applies Cosine Annealing to simulate circadian sleep/wake cycles, while instantly injecting massive Learning Rate spikes (Panic Learning) whenever the organism experiences a near-death trauma simulation, permanently rewiring the Glymphatic plasticity_state tensor in FP64 precision.