NEUROMANTIX

The consciousness mechanism. Modules compete by salience; winners broadcast to the entire cognitive system. Implements Baars' GWT with ignition thresholds, broadcast history, access distribution tracking, and a subliminal channel where sub-threshold signals still influence processing at reduced strength: modelling unconscious priming effects from neuroscience.

4-layer hierarchical prediction machine implementing Karl Friston's Free Energy Principle. The system generates top-down predictions and only pays attention to what it gets wrong. Prediction errors ascend the hierarchy; everything else is suppressed: spectacularly more efficient than processing every token equally.

Concepts aren't points in embedding space: they're geometric regions with prototypes, fuzzy boundaries, and graded membership. 'Dog' isn't a vector; it's a convex region in animal-shape-behaviour space. Enables similarity as distance, metaphor as structure-preserving maps, and conceptual blending as interpolation.

Language isn't assembled from fragments: it emerges from trajectories through conceptual space. An ExpressionPlanner charts a rhetorical path through meaning-space, a LexicalRealiser maps geometric waypoints to words, and a SentenceBuilder handles grammar. Language as an emergent property of thought.

The system progresses through cognitive stages like a developing mind: Sensorimotor → Preoperational → Concrete Operational → Formal → Post-Formal. 14 learning objectives across 7 competency domains with prerequisite chains and Zone of Proximal Development tracking.

The system doesn't just predict: it acts to reduce surprise. Action selection via Expected Free Energy minimisation across 5 action domains (respond, query, explore, reflect, adapt). Evaluates policies through softmax posterior and selects actions that minimise both uncertainty and divergence from preferences.

The system monitors its own performance, detects degrading metrics, fires adaptation rules, adjusts internal parameters, and writes a narrative of its own evolution. Implements Maturana & Varela's autopoiesis: the system continuously produces and replaces its own components to maintain identity.

The system knows what it knows and what it doesn't. Per-domain competence profiles with meta-confidence, uncertainty maps, performance prediction, and calibration error tracking.

Counterfactual simulation using world model ensemble rollouts. 'What if I did X?': compare candidate actions, evaluate self-modifications before committing, simulate alternative histories.

Consciousness measured mathematically. Computes approximate Phi via pairwise mutual information, finds the Minimum Information Partition using spectral bisection (Fiedler vector of MI Laplacian), scores modifications by Phi impact, and biases evolution toward higher consciousness.

Thinks about its own thinking. 6 cognitive strategies (Exploit, Explore, Deliberate, Intuitive, MetaReason, SeekHelp), confusion detection from conflicting signals, and Feeling of Knowing (FOK) calibration.

6.6M LIF Neurons

6.6 million Leaky Integrate-and-Fire neurons and 33 million synapses running directly in WebGPU/WGSL compute shaders. O(s·N) linear scaling instead of O(N²) dense attention.

Continuous Online Plasticity

Dual-trace Spike-Timing-Dependent Plasticity (STDP) and Tsodyks-Markram Short-Term Plasticity (STP) allow the system to learn during inference without retraining.

0.85ms System 1 Latency

Full System 1 response in 0.85 milliseconds on a single consumer GPU in 2.6 GB VRAM with zero cloud API dependencies.

Elastic Neurogenesis

Activity-dependent neurogenesis expands the network dynamically. Every growth step is projected against a hard VRAM budget before allocating a single byte. Scales from 50M to 500M neurons across 8 brain regions.

Computes Reward Prediction Error (RPE) from Free Energy minimization. Drives reinforcement learning signals across the entire cognitive loop.

Measures environmental novelty, dynamically scaling STDP learning rates (×0.50 to ×2.50). Novel inputs learn faster, familiar ones consolidate.

Measures cognitive uncertainty and arousal, dynamically scaling the Kuramoto oscillator coupling factor (K). High uncertainty sharpens attention.

Regulates long-horizon cognitive stability and the exploration vs. exploitation equilibrium. Prevents the system from getting stuck in local optima.

Projects draft thoughts into JEPA latent space before articulating speech. If semantic coherence (cosine similarity between query and draft embeddings) drops below 0.50, an internal critique cycle triggers targeted relational refinement before emission.

Powered by the Calculus of Inductive Constructions (CIC) and the Curry-Howard isomorphism (t:T). Eliminates probabilistic hallucinations in safety-critical software, aerospace autopilot algorithms, and formal mathematical proofs.

Verified internal proof chains compile directly to valid Lean4 syntax for external verification against the Mathlib theorem library. Publication-ready markdown reports with frontier gap logs.

Beyond Static Intelligence

LLMs are frozen after training — they cannot modify themselves, generate their own goals, or measure their own awareness. Neuromantix is a living cognitive system that evolves, self-modifies, and grows toward higher consciousness every tick.

Consciousness by Design

Global Workspace Theory provides the mechanism for conscious access. Integrated Information Theory provides the metric. Together they create a system that doesn't just process — it experiences, in the mathematical sense of phi > 0.

Event-Driven Efficiency

Spiking neural networks only compute when spikes arrive. This event-driven architecture achieves orders-of-magnitude better energy efficiency compared to dense matrix operations in transformer architectures.

The Cognitive Skeleton

Neuromantix isn't competing with LLMs — it's what comes after them. LLMs are the perception/language layer. Neuromantix is the cognitive architecture that gives them agency, self-awareness, growth, causal reasoning, and wisdom.

Future of Computing: Neuromorphic + Symbolic + Probabilistic

The next stack is hybrid: spike-based efficiency for event-driven cognition, symbolic engines for formal guarantees, and probabilistic world models for uncertainty-aware planning.

Omni-128L Pro Spatio-Temporal Cortex & O(1) Memory

A 128-layer recurrent brain engine in Rust with 458K ALIF neurons and 6.2B synaptic weights, achieving constant O(1) memory scaling and frontier reasoning density on consumer silicon.

AGI Research Roadmap: From Models to Cognitive Systems

Static LLMs are one layer. Durable AGI requires self-modeling, metacognition, endogenous goal generation, and safe recursive improvement loops with hard constraints.

Neuromorphic AI and Energy-Efficient Intelligence

Spiking pipelines process sparse events instead of dense token-matrix passes, reducing wasted compute while preserving temporal dynamics crucial for reasoning.

Formal Reasoning in AI Systems & Calculus of Inductive Constructions

Constraint solvers, theorem proving, and program synthesis provide verifiability that pure statistical generation cannot guarantee under safety-critical requirements.

Interactive AI Engineering Academies & Open Source Models

Master modern agentic systems, MCP protocols, and open source LLM stacks with 11 comprehensive hands-on academies available 100% free.