Focus: Real-Time Neuroplasticity, RPE, and Topological Shearing (LTD)
Continuing the empirical extraction of the KAI Rust backend. This volume proves the mechanisms by which KAI physically restructures its own geometry in real-time, bypassing the static backpropagation paradigm.
---
In standard LLMs, "Reward Prediction Error" (RPE) is calculated mathematically via a loss function (like Cross-Entropy Loss) during backpropagation, updating fixed weights.
In KAI, Prediction Error actively modulates a global physical constant (learning_rate), which dictates the plasticity (pliability) of the entire 16,384-dimensional lattice.
src/cognition/neuroplasticity.rs
/// Modulate learning rate from external signals.
/// High dopamine + high novelty = more plasticity.
pub fn modulate(&mut self, dopamine_level: f32, prediction_error: f32) {
// High PE (surprise) + high dopamine = peak learning moment
let target_lr = 0.40 + dopamine_level * 0.35 + prediction_error * 0.25;
self.learning_rate = (self.learning_rate * 0.90 + target_lr * 0.10).clamp(0.20, 2.0);
}
Proof Analysis: The prediction_error physically alters the learning_rate of the lattice. When KAI encounters a surprise or an expectation mismatch (high PE), the target_lr spikes. This puts the entire geometric structure into a "hot" state, making it highly malleable so that the incorrect geometric bridges can be sheared and new ones formed instantly.
---
Standard neural networks approach zero but never physically delete connections during operation. KAI executes true Long-Term Depression (LTD). If a geometric bridge falls out of phase or remains idle, the system actively destroys the connection.
src/core/synapse.rs
/// LTD sweep — weaken synapses that haven't fired recently.
/// Call this on a slow tick (e.g., every 30 world ticks).
/// Prunes synapses that fall below MIN_WEIGHT.
pub fn ltd_sweep(&mut self) {
self.tick += 1;
let tick = self.tick;
let mut to_prune: Vec<usize> = Vec::new();
for (idx, syn) in self.synapses.iter_mut().enumerate() {
let idle = tick.saturating_sub(syn.last_fire_tick);
if idle > LTD_IDLE_TICKS {
let idle_factor = ((idle - LTD_IDLE_TICKS) as f32 / 200.0).min(3.0);
let loss = BASE_LTD * (1.0 + idle_factor);
syn.weight = (syn.weight - loss).max(0.0);
self.total_ltd += 1;
if syn.weight < MIN_WEIGHT {
to_prune.push(idx);
}
}
}
// Prune weakest synapses (reverse order to preserve indices)
for idx in to_prune.into_iter().rev() {
let syn = self.synapses.remove(idx);
// Remove from index
if let Some(indices) = self.index.get_mut(&syn.pre_label) {
indices.retain(|&i| i != idx);
}
self.total_pruned += 1;
}
}
Proof Analysis: KAI runs a background physics tick (ltd_sweep). It continuously calculates an idle_factor based on temporal distance from the last activation. If the weight drops below MIN_WEIGHT, KAI pushes the index to to_prune and literally executes a self.synapses.remove(idx), physically annihilating the connection from the graph. This is real-time topological shearing.
---
When two contradictory concepts attempt to wire together, KAI prevents the bond. It does this not through a logic rule, but through a physical wave dampener called the chi_gate.
src/core/synapse.rs
pub fn record_co_firing(
&mut self,
labels: &[String],
dopamine: f32,
phi_g: f32,
chi: f32,
tick: u64,
lattice_size: usize,
) {
self.tick = tick;
if labels.len() < 2 { return; }
// Contradiction suppresses bonding — contradicting cells shouldn't wire together
let chi_gate = (1.0 - chi * 0.8).max(0.05);
// LTP magnitude: base × dopamine boost × emergence boost × contradiction gate
let ltp_gain = BASE_LTP
* (1.0 + dopamine * 0.8)
* (1.0 + phi_g * 0.5)
* chi_gate;
for i in 0..labels.len() {
for j in 0..labels.len() {
if i == j { continue; }
self.apply_ltp(&labels[i], &labels[j], ltp_gain, tick, lattice_size);
}
}
}
Proof Analysis: The chi variable tracks global entropy/contradiction. When chi is high, chi_gate drops to 0.05. This completely collapses the ltp_gain multiplier. So, if KAI is processing a lie or a severe logical conflict, the destructive interference wave mathematically stops the bridge from forming, no matter how much dopamine is present.