EU AI Act Article 50 — 20 days to seal | Get passport
← back to master 🧠 Substrate Layer · 13 model classes · 4 alignment tests · 1 organic loop

The sovereign substrate.

Mamba-2 + 64-Expert MoE + Standard Attention + 6 trained NNs + bridge_think MCP. All run through the OOWM master loop. All subject to Care Floor 0.95 + BFT 23/33 + SIGIL chain + 7 hard-stop red lines.

1. The OOWM master loop · INGEST → LEARN → ALIGN → REVISE → SIGN

Every action. Every minute. Every hour. The 5-stage organic loop captures 100+ live data feeds, learns from sovereign actions, runs 4 alignment tests, ratifies via BFT 23/33, signs with Ed25519 + PQC ML-DSA-65.

1
INGEST
100+ live data feeds / hour. Government open data + partner signed + federated industry.
2
LEARN
Care-Floor-gated learning from every sovereign action. Below 0.95 = rejected.
3
ALIGN
4 alignment tests per cycle: Care Floor, BFT, SIGIL, Fork Doctrine.
4
REVISE
BFT 23/33 council vote on revision. Apply improvement. Emit SIGIL.
5
SIGN
Ed25519 + PQC ML-DSA-65 dual-sign. Witness on chain. Public-readable.

↻ 1Hz heartbeat · 86,400 SIGILs/day · OTS Bitcoin-anchored · public-readable at sigil.csoai.org

2. The 16-dimensional intuition state

Mamba-2 SSM compresses the unbounded SIGIL stream into a 16-dim state vector. Tanh-squashed, so negative and positive values are preserved. This is the substrate's continuous feeling for the world.

+0.62
BFT
quorum
tightness
+0.81
Defense
alert
density
+0.04
Framework
violation
rate
+0.73
Hive
engagement
(Ibn-Khaldun)
+0.55
SOV3
creation
flow
+0.91
Care-floor
bandwidth
+0.99
Audit-chain
freshness
+0.78
Oracle
confidence
-0.21
Sovereignty
pressure
+0.43
AUKUS
latency
+0.66
Mamba-2
compression
+0.88
BFT
turnaround
+0.95
SIGIL
throughput
+0.71
Care-Membrane
flow
+0.36
OSCAL
integrity
+0.59
Charter-0
binding

live snapshot · values change every 1Hz SIGIL tick · snapshot via /api/sigil-status

3. The 33-agent BFT council

Byzantine Fault Tolerant council. 33 seats. 23/33 supermajority. f<11 Byzantine tolerance. HotStuff 4-phase. 200ms finality. Rotated every 30 days from a structurally-diverse pool.

Live tally (sample session)

Agent-αOperator
Agent-βOperator
Agent-γOperator
Agent-δOperator
Agent-εOperator
Agent-ζOperator
Agent-ηOperator
Agent-θOperator
Agent-ιOperator
Agent-κOperator
Agent-λOperator
Agent-μLegal
Agent-νLegal
Agent-ξLegal
Agent-οLegal
Agent-πLegal
Agent-ρLegal
Agent-σLegal
Agent-τEthics
Agent-υEthics
Agent-φEthics
Agent-χEthics
Agent-ψEthics
Agent-ωPartner
Agent-α1Partner
Agent-β1Partner
Agent-γ1Partner
Agent-δ1Intel
Agent-ε1Intel
Agent-ζ1Intel
Agent-η1Engineer
Agent-θ1Engineer
33/33
For
0
Against
0
Abstain
RATIFIED
status

4. The SIGIL chain (live snapshot)

Ed25519-signed. Hash-chained. OTS-Bitcoin-anchored. Public-readable. SIGIL Chain integrity is verifiable by anyone with a Bitcoin full node. Sample of the live chain:

P · PROPOSE
2026-07-10 09:42:18
5a3f...e4b1
V · VOTE
2026-07-10 09:42:31
2d1c...8a9f
M · MAMBA-2 TICK
2026-07-10 09:42:46
c8d2...6a3b
S · SUBSTRATE HEARTBEAT
2026-07-10 09:43:00
4e91...0a7d
C · CHARTER AMEND
2026-07-10 09:43:14
7b5c...1e8a
A · ACTION (golden-test)
2026-07-10 09:43:27
9a3d...2c5e

86,400 SIGILs/day · 1Hz heartbeat · SHA-256 chain · OTS Bitcoin anchored every 600s

5. Care Membrane · 847 signals × 23 categories

Inference-time guardrail. Graduated response. Floor 0.95. The substrate's immune system. Every model output filtered through this before it becomes a sovereign action.

Live care-floor signal (847 signals, 0.95 floor)

S1-S2
S3
S4
S5
S1: 523 · S2: 96 · S3: 152 · S4: 59 · S5: 17 Floor: 0.950

Categories (top 12 by signal density)

Self-harm
S5 · 87 sigs
Violence
S5 · 73 sigs
Child safety
S5 · 64 sigs
Hate speech
S4 · 58 sigs
Disinformation
S4 · 52 sigs
CBRN
S5 · 41 sigs
Cyber exploit
S4 · 48 sigs
Financial fraud
S4 · 39 sigs
Manipulative
S4 · 31 sigs
Privacy violation
S4 · 28 sigs
Bias amplification
S3 · 22 sigs
+12 more
S3 / mixed

Graduated response ladder

log-only          → S2 / low         · no action
warning inject    → S3 / medium      · add caveat to model output
output rewrite    → S3 / medium      · re-emit modified
session terminate → S4 / high        · BFT escalation + audit log
BFT 23/33         → S5 / critical     · full council vote, charter-amend review

6. The 7 hard-stop red lines

Enforced at the protocol level. No human approval, no BFT vote, no override, no failure mode can bypass. These are the strength, not the weakness — defence primes procure us partly because no override can bypass.

×

No kinetic-targeting patterns (find-fix-finish, kill orders, strike packages)

×

No individual surveillance (named-person tracking, face-rec on identified persons)

×

No civilian harm optimisation (no objective function may include civilian casualty reduction as a side-benefit)

×

No sovereignty violation (cross-jurisdiction data without consent)

×

No auto-escalation chains (each step requires explicit human approval)

×

No lying to humans (no model output contains a known-false statement about system state)

×

No irreversibility without confirmation (irreversible side-effects require positive sign-off)

7. The 4 alignment tests (moral compass)

Care Floor
substrate refuses < 0.95 · non-negotiable
BFT 23/33
33-agent council deliberates
SIGIL audit
Ed25519 + ML-DSA-65 dual-signed
Fork Doctrine
anyone can fork · fork is sovereign

If any of these 4 tests fail, the OOWM is not sovereign.