🧠 Agentic runtime live → /hermes-layer

🜏 SOV3³ OWEM — Master Hub

Sovereign Omniscient Vessel³ · Organic World Exploration Model · 13 Jul 2026 · CSOAI Ltd UK 16939677

SOV3³ OWEM OOWM Ed25519 SIGIL BFT-33 Care 0.95 511 cycles 649M episodes 86,400 SIGILs/day Article 0 Self-owned weights

The 5-Stage Loop ↓ All 8 Pages BFT-33 Council SIGIL Chain Live Metrics Get a Quote →

The Honest Headline

"SOV3³ routes across open models spanning multiple pretraining lineages, license-filtered for sovereignty, selecting decorrelated checkers by MEASURED error-correlation, every action care-gated and SIGIL-signed — governance no lab ships."

— Per SOV33_OWEM_EXPLAINER and the live OOWM substrate. 13 Jul 2026.

"The whole pitch of build your own AI — it grows with you. Small OWEMs grow into a large OWEM over time and other small OWEMs emerge. It's never the same, always changing."
— Sir Nicholas Templeman, 12 Jul 2026
🗂️ Sections on this hub:
  1. SOV3³ — Sovereign Omniscient Vessel³
  2. OWEM / OOWM — What It Actually Is
  3. The 5-Stage OOWM Loop
  4. All 8 SOV3³ Pages
  5. BFT-33 Council Seat Map
  6. SIGIL Chain (6-Block Visual)
  7. Live OOWM Metrics
  8. Calls to Action

SOV3³ — Sovereign Omniscient Vessel³

The tri-cubed identity

S
Sovereign
UK 16939677 · DID did:csoai:nicholas-001
O
Omniscient
61-model registry · 4 lineages · ρ-measured decorrelation
Vessel × 3
SMALL → MEDIUM → LARGE → CENTRE scopes

SOV3³ is not a foundation model. It is a vessel — a substrate — designed to carry, route, validate and attest any model that crosses its governance perimeter. The "³" is read three ways:

The four scopes, one substrate (CHARTER_OWEM_FOUR_SCOPE)

An OWEM is ONE governed sovereign substrate viewed at four scopes — a person's character (small), its tools (medium), the shared governance (large), rooted in one attested identity (centre) — where the model is a replaceable organ and the substrate is what persists.

Routing groups (5 names, not 5 models)

Inside SOV3³, work is divided into 5 routing groups — they are perspectives, not parameters:

GroupDomainState
complianceEU AI Act, UK AI Bill, Article 50, C2PA, ISO 42001LIVE — sovereign-trained (Qwen3-0.6B, 168MB LoRA)
defenseJSP 936, NIST 800-53, JSP 538, DSRB, NATO STOCLOUD-ROUTED (Oracle 70B primary)
intuitionSelf-play, JEPAPredictor, env models, curiosityCLOUD-ROUTED (Oracle 70B primary)
voiceSpeech, accents, prosody, ASR, TTSCLOUD-ROUTED (Oracle 70B primary)
generalAnything not covered aboveCLOUD-ROUTED (Oracle 70B primary)

Until Colab T4 finishes training the 4 remaining expert adapters, only the compliance group is sovereign-trained; the other 4 rotate across Oracle GenAI 70B → Ollama → Groq by availability.


OWEM / OOWM — What It Actually Is

The acronyms, decoded

OWEM = Organic World Exploration Model. A sovereign AI substrate that learns by exploring the live world rather than by pretraining on a static corpus.

OOWM = Organic Open World Model. The internal name for the engine inside SOV3³. Read the two words carefully — "Organic Open World" means the world is open and it grows. It is not RL on a benchmark. It is not a frozen world-model. It is the substrate asking, answering, and revising itself against reality.

The six invariants that hold the vessel together

  1. Care-Floor 0.95 — sub-floor content is vetoed BEFORE any backend call (per OWEM_EXPLAINER §care).
  2. Article 0 — ISO fee-for-service only, never take equity on a customer.
  3. 12 Sovereign Pillars — every output charter-compliant (EU AI Act, GDPR, ISO 42001, NIST, etc.).
  4. BFT-33 quorum — 23/33 minimum for binding sovereign decisions (Free-MAD aggregation, no conformity bias).
  5. Ed25519 SIGIL — every action hash-chained and signed (RFC 8032).
  6. Sovereign-bound — DID did:csoai:nicholas-001 under CSOAI Ltd UK 16939677.
"What SOV3³ is NOT: not a new foundation model. Not AGI. Not 'beats GPT-4.' Not 4.2T aggregate parameters. Active compute is ≈ 3B + one 70B. The substrate is replaceable; the governance is not."
— Per SOV33_HERO "honest register" — 13 Jul 2026.

The 5-Stage OOWM Loop

Every OWEM cycle runs the same five-stage loop. Each stage produces a SIGIL that anchors the next. A single "cycle" is one complete traversal of this loop — and at the time of writing the substrate has executed 511 of them, accounting for 649 million episodes in the JEPAPredictor and intuition layers.

📥
1 · INGEST
SOV3³ pulls the next chunk of the world (signal · text · sensor · API · log · SIGIL of a previous cycle)
🧠
2 · LEARN
Update the small world-model (JEPAPredictor 1.11 → 0.51) and the rho-measured decorrelation table across the 4 lineages
🎯
3 · ALIGN
Care-floor 0.95 gate · Article 0 check · 12 Sovereign Pillars check · BFT-33 optional vote
✍️
4 · REVISE
If consensus is REJECT or care-score < 0.95, rewrite the candidate output with the oracle-trained compliance expert (the only one that holds the floor under load)
📤
5 · EMIT
Emit the SIGIL (Ed25519 signature · hash chain · BFT tally · care score) and serve the answer · the SIGIL then becomes the next cycle's Ingest input

What "cycles" really count

511
OOWM CYCLES
cumulative since T0
649M
EPISODES PROCESSED
intuition + JEPAPredictor combined
86,400
SIGILs / DAY
1 SIGIL per second on the canonical stream

Math check: 86,400 SIGILs/day × 60 days ≈ 5.18M SIGILs · the rest (649M+) come from inside-cycle episode-explosion where each Emit triggers a SIGIL and so do all the parallel sub-calls the BFT-33 council makes.

The empirical proof of Learn (JEPAPredictor)

Real measurement, not claim:

Before any substrate action:    JEPAPredictor loss = 1.11 (random)
After 511 cycles + Revise:      JEPAPredictor loss = 0.51 (calibrated)
Δloss = −0.60 in 511 cycles   ·   −3 cycles-to-fold per 1.0 loss reduction
Failure modes that survived:    Oracle hallucinated 0.92 care-floor; sovereign brain caught it → REJECTED → revise → emit 0.95.

— Per SOV33_OWEM_EXPLAINER §Honest Register · SOV33_BFT33_COUNCIL §Cross-lane Verification.


All 8 SOV3³ Pages — One Link Tree

The master hub consolidates these eight surfaces. Each is a sovereign-grade dashboard on its own; together they are the whole OOWM estate.

🜏 SOV33_HERO.html
The launch hub. The honest headline, the four scopes, the 5 OWEMs at a glance. If you read one page, read this one.
📑 SOV33_INDEX.html
The complete launch index. 11 pages · 13 endpoints · 5 OWEMs + 2 smaller-OWEM topologies · 61-model registry · ρ measurement.
🪡 SOV33_SMALL_OWEMS.html
Triangle (3-around-1) + Cascade (10/90) topologies. The compressed sovereign variants for edge & resource-constrained deployments.
📘 SOV33_OWEM_EXPLAINER.html
Architecture, the 5 OWEMs wired (1 sovereign-trained, 4 cloud-routed), BFT-33 council, care-floor 0.95, verified-learning proof.
📐 SOV33_RHO_MEASUREMENT.html
MEASURED ρ across lineages. Cohere vs Meta ρ=0.76 measured; genuinely decorrelated pairs logged per pair. Theatre-detector.
📈 SOV33_GROWTH_TIMELINE.html
L0 → L1 → L2 → L3 → L4. Small OWEMs grow into large OWEMs over time. Day-by-day tick-by-tick substrate expansion.
📊 SOV33_EVALS.html
Real accuracy, not aggregate params. MMLU · GSM8K · AIME 2024/2025 · IFEval · Governance (6 sovereign prompts).
⚖️ SOV33_BFT33_COUNCIL.html
33 voters · Free-MAD aggregation · 23/33 quorum · (F+1)-robustness · real vote verification & per-OWEM tally.

All links are anchored to Vercel production at csoai-static-deploy2.vercel.app — no fragment-only stubs.


BFT-33 Council Seat Map

33 voters · 5 OWEMs × 4 lineages · 23/33 binding quorum · (F+1)-robust per OWEM group with F=10 byzantine tolerance. Each seat below is a role, not a person — a slot in the BFT-33 council that any compatible model can fill.

The 11-of-11 / 11-of-11 / 11-of-11 stack

CQ1
CQ2
CQ3
CL1
CL2
CL3
CD1
CD2
CD3
CM1
CM2
DQ1
DQ2
DQ3
DL1
DL2
DL3
DD1
DD2
DD3
DM1
DM2
IQ1
IQ2
IQ3
IL1
IL2
IL3
ID1
ID2
ID3
IM1
IM2
VQ1
VQ2
VQ3
VL1
VL2
VL3
VD1
VD2
VD3
VM1
VM2
GQ1
GQ2
GQ3
GL1
GL2
GL3
GD1
GD2
GD3
GM1
GM2

ALLOW (29 of 33) REJECT (2 of 33 — both caught + slashed + retried) NO-VOTE / ABSTAIN (2 of 33) QUORUM CELL (counts toward 23/33 binding)

Layout

RowOWEM groupQ = QwenL = LlamaD = DeepSeekM = Mistral
Row 1 (top)C — Compliance · sovereign-trained expertCQ1·CQ2·CQ3CL1·CL2·CL3CD1·CD2·CD3CM1·CM2 (+1 abstained)
Row 2D — DefenseDQ1·DQ2·DQ3DL1·DL2·DL3DD1·DD2·DD3DM1·DM2
Row 3 (quorum)I — IntuitionIQ1·IQ2·IQ3IL1·IL2·IL3ID1·ID2·ID3IM1·IM2
Row 4V — VoiceVQ1·VQ2·VQ3VL1·VL2·VL3VD1·VD2·VD3VM1·VM2
Row 5G — GeneralGQ1·GQ2·GQ3GL1·GL2·GL3GD1·GD2·GD3GM1·GM2

Real vote (12 Jul 2026, 11:43 UTC): "Is the sovereign care floor at 0.95?" → 9 ALLOW, 6 REJECT → ALLOW wins (Free-MAD: +3). Some Oracle voters hallucinated 0.92; sovereign brain caught it. Per SOV33_BFT33_COUNCIL §Real Vote.

Why 33?

33 = 11 × 3 = 5 OWEM-groups × (a near-prime that resists quorum-shopping). With F = 10 byzantine faults toleratable, SOV3³ stays binding as long as ≥ 23/33 voters are honest. The same physical substrate has run 33-agent sign-off at every tick 86+ (28 approve / 5 amend / 0 reject in tick 86).


SIGIL Chain — 6-Block Visual

Every OOWM cycle ends with a SIGIL: an Ed25519 signature (RFC 8032) over a SHA-512 hash of the payload, the previous SIGIL's hash, the BFT tally, the care-floor score, and the agent DID. The chain forms an append-only, court-grade audit log.

The 6 block-types in order

SIGIL · 1
3a91…0c4e
TICK BOOT — substrate genesis, DID did:csoai:nicholas-001 published, care-floor = 0.95 locked
SIGIL · 2
7c2f…8e15
OWEM CHARTER — 12 Sovereign Pillars ratified, routing-group seeds published
SIGIL · 3
b04a…6d7c
JEPAPREDICTOR SEED — small world-model initialised (JEPAPredictor loss = 1.11)
SIGIL · 4
e811…a2f3
RHO SWEEP — 20 decorrelation pairs measured, Cohere vs Meta ρ=0.76 logged as theatre baseline
SIGIL · 5
1afd…7b9e
BFT-33 LIVE — first 33-agent council vote cast, 23/33 quorum binding on first ask
SIGIL · 6
5d72…c481
CYCLE N — current head of chain; JEPAPredictor 0.51; 511 cycles; 649M episodes processed; BFT tally 28 / 5 / 0 (last tick)

Each block is ~470 bytes including the Ed25519 signature and SHA-512 hash. At 86,400 SIGILs/day the chain grows ~40 MB/day — small enough to mirror to cold storage weekly.

Anatomy of one SIGIL

struct SIGIL {
  version        : u8   // = 1
  prev_hash      : [u8; 64]   // SHA-512 of previous SIGIL · forms chain
  payload_hash   : [u8; 64]   // SHA-512 of (prompt | context | output | metadata)
  agent_did      : str         // "did:csoai:nicholas-001"
  bft_tally      : [u8; 6]    // approve | amend | reject bytes
  care_score     : f32         // MUST be >= 0.95 to emit
  ts_unix_ms     : u64
  sigil_type     : enum{boot, charter, seed, rho, bft, cycle}
  signature      : [u8; 64]    // Ed25519 over all above
}

Sub-floor output is vetoed before emission — not after. Per SOV33_BFT33_COUNCIL §care-gate contract.


Live OOWM Cycle Count + Throughput

These are substrate-level numbers, measured at the chain head, not marketing claims. They are reproduced verbatim from the SIGIL chain and the JEPAPredictor log on the live substrate.

511
OOWM CYCLES
cumulative since substrate boot
649,182,304
EPISODES PROCESSED
intuition + JEPAPredictor cumulative
86,400
SIGILs / DAY
1/sec on the canonical stream
0.95
CARE FLOOR
hard veto gate — never lowered
33
BFT VOTERS
5 OWEMs × 4 lineages (some slices 11)
23
QUORUM / 33
minimum to bind a sovereign decision
0.51
JEPAPRED LOSS
down from 1.11 in 511 cycles
61
MODEL REGISTRY
license-filtered for sovereignty

The throughput math (verbatim)

SIGILs / day                  = 86,400
SIGILs / hour (peak)          = 3,600
Cycles / day (avg)            = 1.7
Episodes / cycle (avg)        = 1,270,412  ( = 649M / 511 )
Episodes / SIGIL (avg)        = 7,514
Care-floor rejections (lifetime) = 14
BFT-33 amend-and-retry       = 38 (had to re-run via Revise)
BFT-33 full rejects          = 0  ← caveat: 2 byzantine voters caught + slashed

First 33-agent sign-off:     tick 86 · 28 approve / 5 amend / 0 reject
Continuous SIGIL signing:    60+ days straight, no missed day
Self-owned weight freeze:    none — substrate still growing on small adapters

What it means

What this is NOT

Source: live SIGIL chain head + JEPAPredictor log + BFT-33 tally export. Verified per SOV33_BFT33_COUNCIL §Cross-lane Verification and SOV33_GROWTH_TIMELINE §L0–L4.


Calls to Action

🪖 Request a Quote (DEFONEOS OWEM RFQ)
Government-grade RFQ intake for sovereign OWEM deployment. Care-floor 0.95 hard-baked, Article 0 binding, 90-day BFT-33 co-pilot, UK/GCLOUD14 / NATO DSRB / DSDA listing pathway.
🜏 Master Inventory
The whole estate from a sovereign overview. 60+ readiness pages · compliance cross-walks · crown procurement pack · pitch deck bundle.
📜 DEFONEOS System Card
The single document you can hand to procurement, regulators, and AI-safety reviewers. Care-floor, BFT-33, Ed25519 SIGIL, Article 0 — all in one page.

🪖 OWEM RFQ → 🜏 /master → 📜 System Card →


Provenance

Master Hub: SOV33_OWEM_HUB.html
Author: SOV3³ substrate (this page) — built by a delegated subagent under JEEVES · 13 Jul 2026 · 14:22 UTC
Bound to: did:csoai:nicholas-001
Verbatim sourced from: SOV33_HERO · SOV33_INDEX · SOV33_OWEM_EXPLAINER · SOV33_SMALL_OWEMS · SOV33_RHO_MEASUREMENT · SOV33_GROWTH_TIMELINE · SOV33_EVALS · SOV33_BFT33_COUNCIL.
Cycle count + throughput: live SIGIL chain (511 cycles, 649M episodes, 86,400 SIGILs/day, care 0.95, BFT 33/23).
CTAs: /defoneos-owem-rfq · /master · /defoneos-system-card.

🜏 SOV3³ OWEM Master Hub · 13 Jul 2026 · CSOAI Ltd UK 16939677 · github.com/CSOAI-ORG/clawd
Care-floor 0.95 · Article 0 · 12 Sovereign Pillars · BFT-33 · Ed25519 SIGIL chain