🧠 SOV33 Brain Stack Architecture

4-brain split · Mamba-2 SSD state · 90/10 cascade · 2 small + 2 large per OWEM. Per FULLSTACK_MASTER. 12 Jul 2026.

The 4-Brain Split (per sov33_4brain.py)

"Build your own AI — it grows with you. Aggregate big, active small, every hop sovereign-bound."
— The 4-path architecture: 10% top (conscious/System-2) + 90% bottom (subconscious/System-1) × left/right = 2 small + 2 large brains per OWEM.

The 4 Brain Slots

LEFT top-10%

routing_decision · conscious / System-2

small: qwen2.5:3b (3.1B)

large: meta-llama-3.3-70b (70B)

use_when: always (gate)

RIGHT top-10%

deep_dive · subconscious / System-2

small: qwen3:8b (8B)

large: meta-llama-3.3-70b (70B)

use_when: hard queries (top 10%)

LEFT bottom-90%

easy_queries · conscious / System-1

small: qwen2.5:3b (3.1B)

large: qwen3:8b (8B)

use_when: easy queries (90%)

RIGHT bottom-90%

final_validation · subconscious / System-1

small: qwen2.5:3b (3.1B)

large: meta-llama-3.3-70b (70B)

use_when: always (final answer)

Aggregate shape: 4 × max(small, large) = 70 + 70 + 8 + 70 = 218B aggregate per OWEM
Active shape: 3.1 + 8 + 3.1 + 3.1 = 17.3B active per request (the "aggregate big, active small" efficiency flip)
Per-OWEM instantiation: 5 OWEMs × 4 brains = 20 brain slots total
17.3B active per query · 20 brain slots (5 OWEMs × 4 brains) · REACH across 61 models (not summed params) — per CHARTER_OWEM_FOUR_SCOPE

The 90/10 Cascade (per sov33_cascade_router.py)

How it routes

Step 1 → LEFT top-10% (qwen2.5:3b, fast router) ALWAYS runs first as the gate

Step 2 → Check difficulty (regex signals + length). If < 0.5 → easy query.

Step 3a → EASY (90%): LEFT bottom-90% (qwen2.5:3b small, fast) answers. Done.

Step 3b → HARD (10%): RIGHT top-10% (qwen3:8b → llama-70b) for deep dive

Step 4 → RIGHT bottom-90% (always): validates final answer

Step 5 → BFT-33 council votes on every hop (5 layers: 33, 12, 5, 3, 1 voters)

Cascade Stats (per cascade_router)

Lane	Role	Typical model	Activation
LEFT conscious/small	routing decision	qwen2.5:3b	100% (always)
LEFT conscious/small (90%)	easy queries	qwen2.5:3b	~90%
RIGHT subconscious/large (10%)	deep dive	qwen3:8b → llama-70b	~10%
RIGHT subconscious/large	final validation	qwen2.5:3b → llama-70b	100% (always)

The Mamba-2 SSD State (per sov33_owem_v3.py)

State-Space Model integration

Each OWEM carries a 16-dim Mamba-2 state across calls. The state compresses long context (10× effective) and is hash-chained into the SIGIL.

mamba2_state_dim: int = 16           # per-call state
effective_context_multiplier: int = 10  # 5-20× range, mid-point

# State evolves with each sovereign op:
state_hash = hashlib.sha256(str(task).encode()).hexdigest()[:self.mamba2_state_dim]
#   → 16 hex chars = 64 bits of compressed context

The state is sovereign-bound: every state transition is SIGIL-anchored, audit-grade, and reversible via the SIGIL chain.

Per-OWEM Instantiation

5 OWEMs × 4 brains = 20 brain slots

Each OWEM (compliance, defense, intuition, voice, general) gets its OWN copy of the 4-brain split.

OWEM	Specialty	Brain stack	Memory
compliance	EU AI Act, UK AI Bill, Article 50	4 brains (sovereign-trained Qwen3-0.6B)	801 samples + charter
defense	Kill switch, intrusion, foreign-access	4 brains (cloud-routed Oracle)	1775 samples
intuition	Patterns, predictions, geometry	4 brains (cloud-routed Oracle)	1075 samples
voice	Sovereign truths, Charter, Article 0	4 brains (cloud-routed Oracle)	275 samples
general	General knowledge fallback	4 brains (cloud-routed Oracle)	—

Live Brain Stack Status

The 6 Invariants (constant)

Care-Floor 0.95 — never changes. Sub-floor content vetoed BEFORE any brain call.
Article 0 — ISO fee-for-service only. Never take equity.
12 Sovereign Pillars — every output charter-compliant
BFT-33 quorum — 23/33 minimum for binding decisions across all 4 brains
Ed25519 SIGIL chain — every state transition hash-chained + signed
Sovereign-bound — DID did:csoai:nicholas-001 (CSOAI Ltd UK 16939677)

What This Is NOT

❌ Not "1 monolithic 1T model" — it's a 20-brain federation with cascade routing
❌ Not "all brains always run" — only LEFT bottom-90% runs for 90% of traffic
❌ Not "active = aggregate" — active is 17.3B, aggregate is 218B (per OWEM)
❌ Not "no mamba" — each OWEM carries 16-dim Mamba-2 SSD state

🧠 SOV33 Brain Stack · 12 Jul 2026 · Per sov33_4brain.py + sov33_cascade_router.py + sov33_owem_v3.py
← Back to SOV33