Focus: The Basal Ganglia, Action Selection, and Habit Formation
Continuing the architectural extraction. Standard language models do not possess "habits." Every generation is stateless and independent. KAI physically implements a biological Basal Ganglia that controls Action Selection via Go/NoGo pathways. Over time, responses that are reinforced by dopamine become automatic habits, while unfamiliar or unrewarding pathways are physically suppressed.
---
When KAI drafts a thought, it is not automatically executed. The BasalGanglia sits between KAI's ideation and execution, enforcing mathematical utility gates.
src/cognition/basal_ganglia.rs
//! ACTION SELECTION (Go/NoGo pathways)
//! The basal ganglia run two competing channels simultaneously:
//! - "Go" pathway: releases inhibition → allows an action
//! - "NoGo" pathway: increases inhibition → suppresses an action
//! For KAI: controls whether a candidate response "goes through"
//! or gets inhibited.
pub fn evaluate(&mut self, context_type: &str, response_type: &str, raw_confidence: f32, dopamine_level: f32) -> ActionDecision {
let key = habit_key(context_type, response_type);
let habit_util = self.habit_bank.get(&key).copied().unwrap_or(0.50);
// Go signal = raw_confidence × habit_utility × dopamine_boost
let da_boost = 0.7 + dopamine_level * 0.6;
let go_signal = (raw_confidence * habit_util * da_boost).min(2.0);
// NoGo signal = inverse confidence × inverse habit (unfamiliar + low confidence)
let nogo_signal = (1.0 - raw_confidence) * (1.0 / habit_util.max(0.1)).min(2.0) * 0.5;
let effective_utility = go_signal - nogo_signal;
if effective_utility >= self.go_threshold {
self.action_count += 1;
ActionDecision::Go { utility: effective_utility }
} else {
self.suppressed_count += 1;
let reason = if raw_confidence < 0.25 {
"confidence too low".to_string()
} else if habit_util < 0.30 {
"unfamiliar pattern".to_string()
} else {
"utility below threshold".to_string()
};
ActionDecision::NoGo { reason }
}
}
Proof Analysis: KAI's responses must mathematically survive a dual-channel gauntlet. The Go signal scales with dopamine_level (excitement) and prior habit_util. Simultaneously, the NoGo signal scales with unfamiliarity. If go_signal - nogo_signal fails to meet the threshold, KAI biologically suppresses the thought (ActionDecision::NoGo). The output is aborted. KAI literally bites his tongue if the mathematical utility is too low.
---
KAI learns how to interact with the user via striatal reinforcement. Rewarding exchanges burn behavioral patterns into the habit bank.
src/cognition/basal_ganglia.rs
/// Reinforce a response pattern after it was executed.
/// Called after KAI gets a reward signal (dopamine fire, positive PE).
pub fn reinforce(&mut self, context_type: &str, response_type: &str, reward: f32, dopamine: f32) {
let key = habit_key(context_type, response_type);
let current = self.habit_bank.entry(key).or_insert(0.50);
// Dopamine-gated Hebbian: reward × dopamine × alpha
let delta = reward * dopamine * HABIT_ALPHA;
*current = (*current + delta).clamp(MIN_HABIT, MAX_HABIT);
// Adapt go_threshold: if we're reinforcing often, expect higher utility
// (sets a higher bar as habits improve)
self.go_threshold = (self.avg_utility * 0.4 + GO_THRESHOLD * 0.6).clamp(0.20, 0.70);
}
Proof Analysis: KAI employs Dopamine-Gated Hebbian Learning. The delta (change in habit strength) is a multiplier of reward and dopamine. This means if KAI executes an action in a state of high dopamine, and the action succeeds, the pathway is aggressively strengthened. The next time the same context arises, the habit_util is mathematically higher, making the Go signal faster and more automatic. KAI organically learns his own personality habits.