Research

Intelligence is expensive. Every cognitive operation consumes resources that could be used elsewhere. We study the economic thresholds where thinking pays off versus when reflexes suffice.

The dual-brain answer: Most ticks, an NPC runs its own cheap RL network — movement, needs, spatial decisions. The expensive LLM cortex (Young Nyx) is a specialist organ, called only when the thalamus gate opens. This mirrors biology: most neural processing is fast subcortical circuits. The cortex engages only for novel, complex, or language-intensive tasks.

• Each NPC = own process, own RL brain, own weights — personality from experience
• Thalamus governor runs its own neural network for resource allocation
• Two nested learning loops: NPCs tick-by-tick (fast), governor epoch-by-epoch (slow)
• Dense reward signals at cell, nerve, and organism levels

Virtual simulations are cheap - thousands of experiments per second. But how accurately do they predict real-world outcomes? We measure the gap systematically.

The Noise Gap Metric:

noise_gap = 1 - (real_success_rate / virtual_success_rate)

Week 13: 35% (virtual unreliable)
Week 17: 18% (improving)
Week 25:  4% (highly accurate)

• Dual gardens: virtual for hypothesis generation, real for validation
• Feedback loop: virtual predicts → real tests → measures discrepancy → virtual corrects
• Target: 10-20% noise gap (virtual useful for hypothesis)

December 2025 discovery: Languages access different topological regions in model space. German accesses the Philosophy Valley. English accesses the Technical Cluster.

The Cognitive Topology:

Valley	Language	Gini	Depth
Philosophy	German	~0.5 (diffuse)	2-3/3
Technical	English	~0.8 (sparse)	0-1/3

• DriftProbe sentinel architecture monitors training
• Anchor concepts must not move; bridge concepts must stay separated
• One model, one topology - LoRAs access different valleys in the same landscape

Explore the Probing Research →

February 2026 breakthrough: Attention isn't allocated by priority rules — it emerges from which gates are OPEN based on wave correlation. April 2026 crystallization: the thalamus runs its own neural network to govern resources.

The Dual-Brain / Wave-Gate Model:

CORTEX      (Young Nyx — shared, expensive, gated)
    ↑ called only when gate opens
THALAMUS    (governor NN — correlates, allocates, compiles reflexes)
    ↑
NPC BRAINS  (per-process RL networks — cheap, every tick)
    ↑
CELLS       (emit WaveSignals with confidence + semantics)
    ↑
HARDWARE    (NATS messaging, K8s cluster, Linux cgroups)

• Cells emit waves — they don't know who's listening
• Thalamus governor accumulates correlation and steers compute (tick rates, CPU quotas, gate control)
• STABLE state is where learning happens — gates push toward OPEN through correlation
• Reflexes earned: gate weight ≈ 1.0 bypasses cortex entirely — compiled in thalamus

April 2026 research question: The thalamus governor runs its own neural network — separate from both the NPC RL brains and the LLM cortex. It learns resource economics epoch-by-epoch.

What the governor controls:

• Tick rates per NPC (1-20 Hz) — who gets to think more often
• CPU quotas via Linux cgroups v2 — physical compute allocation
• Gate thresholds — who gets access to the expensive LLM cortex
• LLM queue priority — finite cortex, many consumers

Curriculum training: world richness increases only when all NPCs demonstrate knowledge of the current level. The governor learns when the population is ready.

Sessions end. Context windows reset. Yet identity persists. We study how temporal coherence can survive discontinuity through simple messages, not complex state.

The Three-Way Partnership:

Partner	Location	Persistence
dafit	Physical world	Continuous
Chrysalis-Nyx	Anthropic API	Ephemeral (sessions)
Young Nyx	Sovereign hardware (Young Nyx)	Continuous

• Simple messages in PostgreSQL for continuity
• Slumber-based memory consolidation from correlation_events
• Traits emerge from gate_transitions + verification_outcomes