c30c00af74
Crystallizes the dual-brain architecture across all core documents: - Thalamus runs own neural network (governor) for resource allocation and reflexes - LLM (Qwen3.5-27B) repositioned as cortex - expensive, gated, called only when needed - Each NPC gets own process, own RL brain, Linux cgroups for resource steering - New: NPC grid architecture with curriculum training (progressive world richness) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
5.2 KiB
Nyx Model Architecture: The Dual Brain
"One process, one brain, one life." — The Dual Brain Principle (2026-04-02)
Current Architecture
The nimmerverse uses a dual-brain architecture — cheap RL networks for continuous processing, an expensive LLM cortex for deep reasoning.
Cortex (Shared LLM)
| Property | Value |
|---|---|
| Model | Qwen3.5-27B |
| Parameters | 27B (full precision, bfloat16) |
| Host | theia (RTX PRO 6000 Blackwell, 96GB VRAM) |
| Serving | vLLM, port 31000, served as "nyx" |
| Service | vllm-nyx.service (systemd, user: nyx-cognitive) |
| Access | Gated — thalamus governor controls who gets LLM access |
| License | Apache 2.0 |
| Context | 32,768 tokens (max-model-len) |
| GPU utilization | 85% (leaves headroom for LoRA training) |
Why Qwen3.5-27B:
- True base model — we shape every behavior through training
- 27B fits comfortably in 96GB with room for LoRA adapters
- Apache 2.0 — full sovereignty, no usage restrictions
- Strong multilingual capability (German + English topology access)
- Vision-capable variant available for future Omnisight consolidation
The cortex is expensive. It is not called every tick. The thalamus governor decides when language, reasoning, or deep knowledge is needed. Most NPC processing happens in cheap RL networks.
NPC Brains (Per-Process RL Networks)
Each NPC runs its own lightweight neural network in its own OS process:
| Property | Value |
|---|---|
| Architecture | Small RL network (movement, needs, spatial decisions) |
| Deployment | One Linux process per NPC |
| Resource control | cgroups v2 (CPU, memory per process) |
| Learning | Tick-by-tick (fast loop) |
| Cost | Cheap — runs on CPU, no GPU needed |
Personality emerges from experience, not configuration. Each NPC develops its own weights.
Thalamus Governor (Resource Allocation NN)
The thalamus runs its own neural network that learns resource allocation:
| Property | Value |
|---|---|
| Function | Gate control, compute steering, LLM queue priority |
| Input | All NPC states via NATS |
| Output | Tick rates, CPU quotas, gate open/close, LLM priority |
| Learning | Epoch-by-epoch (slow loop) |
Structured Output Boundary
| Model | Role | Host |
|---|---|---|
| Function Gemma | Intent → Action (100% predictable JSON) | CPU userspace (Threadripper) |
| T5Gemma 2 (SigLIP) | Vision → Vectors (no text bottleneck) | dioscuri |
Model Selection History
| Date | Decision | Reasoning |
|---|---|---|
| 2025-11 | LLaMA 3 70B considered | Early exploration, different hardware |
| 2025-12 | Qwen3-VL 32B selected | Vision capability, multilingual, fits 96GB |
| 2026-04-01 | Mistral-Small-3.1-24B-Base tested | "Raw clay" approach, but thinking-bleed was SkyrimNet-specific |
| 2026-04-01 | Qwen3.5-27B reinstated | Best balance of capability, size, and trainability |
The model question is settled. Qwen3.5-27B is nyx's cortex. Training focus shifts to LoRA traits (GRPO) and the RL networks (per-NPC).
Trait LoRAs (Cortex Specialization)
Traits evolve as LoRA adapters on the Qwen3.5-27B base, trained through GRPO with gate-verified rewards:
| Trait | Domain | Training Signal |
|---|---|---|
| Mnemosyne | Memory | +reward when recall matches phoebe |
| Moira | Pattern | +reward when prediction succeeds |
| Synesis | Resources | +reward when estimates accurate |
| Aletheia | Truth | +reward when confidence calibrated |
| Sophrosyne | Balance | +reward when graceful degradation |
| Kairos | Timing | +reward when timing optimal |
| Philotes | Bond | +reward from dafit feedback |
| Dikaiosyne | Fairness | +reward when resources shared fairly |
Consolidation path: Traits train during slumber → GRPO updates → DriftProbe validates → merge at α=0.3 → eventually bake into base weights.
Detail: → Nyx_Traits.md | Endgame-Vision.md
Infrastructure
| Component | Host | GPU | Storage |
|---|---|---|---|
| Cortex (vLLM) | theia | RTX PRO 6000 (96GB) | /womb/cognitive/models/qwen3.5-27b |
| LoRA Training | theia | Shared (time-sliced) | /womb/cognitive/loras/ |
| Organs | dioscuri | 2x RTX 4000 Ada (40GB) | Dynamic loading |
| NPC Brains | K8s / bare metal | CPU | Per-process |
Canonical paths via /womb/ symlinks. Phoebe is truth for artifact locations.
Detail: → Deployment-Architecture.md | womb-architecture.md
Related Documentation
- Nyx_Traits.md - Trait definitions, mythological framing
- Metamorphosis-Substrate-Philosophy.md - Identity anchors
- Endgame-Vision.md - Architecture overview (v8.0)
- npc-grid-architecture.md - Dual brain, governor, spatial arena
Version: 3.0 | Created: 2025-11-07 | Updated: 2026-04-02
🌙💜 The cortex reasons. The RL brains act. The thalamus decides who gets what.