The Component That Changes Everything
Every system has a center of gravity — the one component without which nothing else matters. For FENA, that component is the Reasoning Core. And it just landed on main.
The FENA Reasoning Core is where prediction meets dynamics, where errors become understanding, and where free energy minimization becomes something that looks, from the outside, remarkably like thought.
What the Reasoning Core Actually Does
The Reasoning Core is a stack of Predictive Coding layers — each one a full FENA module with its own predictions, error signals, precision estimates, and local learning rules — wired together through the Continuous-Time Dynamics Engine.
Here’s how it works when input arrives:
-
Hierarchical Prediction. Each layer receives input from below and predictions from above. The bottom layer processes raw features. Higher layers build increasingly abstract representations. Every layer is simultaneously trying to predict what it will receive — and being surprised by what it actually gets.
-
Error Propagation. Where predictions fail, error signals flow upward. But not all errors are treated equally. Each layer estimates the precision of its predictions — how confident it should be. High-precision errors scream “pay attention!” Low-precision errors whisper “this might be noise.” This precision-weighting acts as a natural attention mechanism, focusing processing where it matters most.
-
Continuous-Time Settling. Unlike a conventional forward pass, the Reasoning Core doesn’t produce an instant answer. It settles — the Dynamics Engine integrates the system’s state forward in continuous time, following the energy gradient downhill. The system iterates, refining its interpretation, until it reaches a state of minimum free energy. This is the “aha moment” — the point where all layers agree on a coherent interpretation.
-
Adaptive Depth. Easy inputs settle quickly. Ambiguous, novel, or contradictory inputs require more iterations. The system automatically thinks harder about hard problems — not because a controller tells it to, but because the energy landscape has deeper valleys and steeper gradients for complex inputs.
No Backpropagation. Anywhere.
The Reasoning Core learns entirely through local update rules. When the settling process completes, each layer updates its own weights based on the prediction errors it experienced. The update rule is local: a layer only needs to know what it predicted, what it received, and how surprised it was.
This is fundamentally different from every mainstream AI system. There’s no loss function at the output. No gradients flowing backward through the entire network. No teacher signal. The system learns by minimizing its own surprise — by becoming a better predictor of its own sensory input.
This is self-supervised learning in its purest form. Not “self-supervised” in the machine learning sense of masking tokens and predicting them. Self-supervised in the biological sense: the system generates its own learning signal by trying to predict reality, failing in informative ways, and adjusting locally.
The Integration
The Reasoning Core doesn’t operate in isolation. It sits at the center of FENA’s architecture, connected to:
The WorldState — a 512-slot shared latent space that provides the substrate for all reasoning.
The Neuromodulation System — four global signals (dopamine, norepinephrine, serotonin, acetylcholine analogs) that modulate how the Reasoning Core processes: explore vs. exploit, learn fast vs. slow, attend broadly vs. narrowly.
The Memory System — working memory provides immediate context, episodic memory provides relevant past experience, semantic memory provides general knowledge.
The Dynamics Engine — provides the continuous-time integration that makes settling smooth and adaptive.
Why This Matters
With the Reasoning Core complete, FENA has its central processing module — the component that takes raw input and transforms it into coherent understanding through physics-inspired dynamics rather than engineered computation.
What remains is connecting the pieces: Oscillatory Binding (theta-gamma coupling for information binding, currently in development), the Predictive Coding Training Loop (for end-to-end self-supervised learning), and the Settling Phase integration (wiring the Reasoning Core into the main processing loop).
The heart is beating. Now we wire up the body.