🧬 The 10 Components
1. OWEM Core Layers (5-layer SOV33 v3 architecture)
The 5 core layers of the SOV33 OWEM v3: Sovereign Binding → BFT-33 Council → 4-anchor MoE → Sovereign-Merge Brain → SIGIL Chain.
Status: ✅ BUILT · Ref: sov33_owem_v3.py
- ✓ L1 Sovereign Binding (Care-Floor 0.95 + Article 0 + Sovereign Mist 12 pillars)
- ✓ L2 12-around-1 BFT-33 Council (23/33 quorum, f=10 BFT, 4 mandatory co-routers)
- ✓ L3 4-anchor × 5-elders MoE (COMPLIANCE / DEFENSE / INTUITION / VOICE)
- ⚠ L4 Sovereign-Merge Brain (qwen3:30b-a3b + QLoRA + Mamba-2 SSD) — STUB, model not pulled
- ✓ L5 Sovereign SIGIL Chain (Ed25519 + OpenTimestamps + Sigstore-cosign)
2. Fluid Pyramid Architecture
N-layer stacked residual OWEMs with per-layer mixing ratio (topology proof, not LLM capability).
Status: ✅ BUILT · Ref: sov33_fluid_pyramid.py
- ✓ N-layer residual learning (each layer learns the residual of the layer below)
- ✓ Per-layer mixing ratio (nu) configurable
- ✓ Fluid depth (grow/shrink layers)
- ✓ Measured winner: 8-flat for 1-brain (Claude measured 8-loss=0.0566 vs 12-loss=0.0593)
3. 4-Brain Hybrid Cascade
2 small + 2 large models, split across left/right (conscious/subconscious).
Status: ⚠️ STUB · Ref: sov33_4brain.py (large models not pulled)
- ⚠ Left top-10% (conscious/routing) — large: meta-llama-3.3-70b NOT PULLED
- ⚠ Left bottom-90% (conscious/easy) — large: qwen3:8b NOT PULLED
- ⚠ Right top-10% (subconscious/deep) — both B and D NOT PULLED
- ⚠ Right bottom-90% (subconscious/final) — large: meta-llama-3.3-70b NOT PULLED
- ✓ Hybrid merge measured winner: 0.1/0.9 split, nu_small=1.0, nu_large=1.0
4. SSD Expert-Streaming Pipeline
6-lever speedup stack: SSD stream, peer prediction, Venturi batch, SIGIL inline, LRU cache, async prefetch.
Status: ✅ PROXY-MEASURED · Ref: ssd_venturi_speedup.py (25.2× combined)
- ✓ SSD stream only active experts (44.6× proxy speedup)
- ✓ Peer prediction (small brain predicts routing)
- ✓ Venturi batch (amortise SHA + care)
- ✓ SIGIL inline (SHA free on critical path)
- ✓ Expert LRU cache (50-75% hit rate)
- ✓ Async prefetch (hide SSD latency)
- ✓ Combined pipeline: 25.2× proxy speedup vs baseline 232 ms/tok
5. Truthful Parameter Accounting
Refuses stack-summing; enforces honest T-base citation. The substrate protects itself from its own overclaims.
Status: ✅ ENFORCED · Ref: ledgerboard_v2.py
- ✓ Refuses parameter stack-summing (45B / 30+7+4+3+1 = RETRACTED)
- ✓ DeepSeek V4-Pro: 1.6T BASE / 862B DEPLOYED (verified)
- ✓ GLM-5.2: 744B total (headline only, architecture retracted)
- ✓ Architecture disagreement logged (proxy vs measurement)
6. Governed Robustness Benchmark
Measures accuracy under adversary. Median SOV33 wins 0.67 vs single-model collapse 0.00. Adopts HarmBench / StrongREJECT / BeaverTails.
Status: ✅ BUILT · Ref: GOVERNED_ROBUSTNESS_LEADERBOARD_2026-07-14.md
- ✓ Measures accuracy under adversary (not keyword refusal)
- ✓ Median SOV33 wins 0.67 through 2/5 adversarial members
- ⚠ Standard datasets (HarmBench / StrongREJECT / BeaverTails) STAGED — not yet run
7. Real Open-Source Base Models
Named, licensed, downloadable models for each tier (T-base, soupable experts). Apache-2.0 commercial-safe Qwen3.
Status: ⚠️ 3/17 PULLED · Ref: absorb_open_source.py
- ⚠ DeepSeek V4-Pro-Base 1.6T (MIT) — NOT PULLED (320 GB)
- ⚠ DeepSeek V4-Flash 158B/13B (MIT) — NOT PULLED (80 GB)
- ⚠ Qwen3-4B / Qwen3-8B (Apache-2.0) — NOT PULLED (7.5 GB)
- ⚠ Qwen3-32B / 30B-A3B (Apache-2.0) — NOT PULLED (~18 GB)
- ⚠ Qwen3.6 (Apache-2.0) — NOT PULLED
- ✓ Qwen Research License (Qwen2.5-3B) — RETRACTED for commercial build
8. Owner-Gated Launch Sequence
4 irreversible actions to transition from sovereign-by-design to sovereign-by-evidence.
Status: ⚠️ BLOCKED · Ref: DISPATCH.md (4 actions, 16 minutes)
- ⚠ Stripe live + £999 Payment Link (5 min)
- ⚠ GitHub repo SOVEREIGN-LAYER-ZERO-CHARTER (60 s)
- ⚠ Push 27 files (30 s)
- ⚠ Send 3 cold emails (10 min)
9. Self-Healing + Autonomous Operations
Cron loops, L7/L8 intuition, BFT-33 voting, Care Floor 0.95. The substrate runs without intervention.
Status: ✅ RUNNING · Ref: PROGRESS_2026-07-14.md
- ✓ 9 cron loops (full-auto, daily-batch, auto-heal, L7, L8, etc.)
- ✓ L7/L8 intuition snapshots every 30 min
- ✓ BFT-33 vote on every sovereign action
- ✓ Care Floor 0.95 enforced at every mint
- ✓ Substrate receipts: 2,052+ → 200/day target
10. Continuous Honesty + Auditing
Fabrication retraction, proxy vs measurement, chain-anchored corrections. The substrate protects itself from its own overclaims.
Status: ✅ LIVE · Ref: ledgerboard_v2.py (LEDGERBOARD-V2 chain)
- ✓ Three-times honesty pattern (Claude's own count) mirrored
- ✓ Retractions logged on chain, not buried
- ✓ Architecture disagreement logged (proxy vs measurement)