csoai/oowm-router

The Organic Open World Model router — CSOAI's master substrate for sovereign AI.

This is not a model. This is the router that decides which sovereign model handles a given task, then wraps the response in BFT + SIGIL attestation per CSOAI's canon.

What it does

Per turn, the OOWM router:

1. Schedules an open-source model from a registered pool (no GPT-4, no Claude, no Gemini — MIT/Apache/CC/UKOGL only)
2. Calls open-source tools from a registered MCP pool
3. Remembers across all chats in one substrate (SciMem cross-thread)
4. Decides when to spawn sub-agents and when to invoke sovereign simulators (Watchdog, Pre-Departure, RiskModel)
5. BFT 12-around-1 deliberates on every consequential action
6. SIGIL Ed25519+PQC-ML-DSA-65 every emission, hash-chained, publicly auditable
7. Care Floor 0.95 enforced — refuses anything below

Architecture

• Layer 0a: MCP tool-router (selects the right tool from the open pool)
• Layer 0b: Semantic query-router (selects the right model from the sovereign pool)
• Hive: 12-around-1 BFT council (Demeter/Maternal Covenant can veto any decision with Care Floor violation)
• SIGIL: Ed25519 + PQC-ML-DSA-65 dual-signed emission chain
• SciMem: cross-thread persistent memory
• Anchor: ~/clawd/csoai.org/sovereign-os/oowm/oowm_runtime.py

Linked model

The router references csoai/sov34-1p5b (Apache-2.0) as its primary sovereign. Sov34 is a qwen2.5-1.5B fine-tune on the CSOAI governance corpus. Router can be pointed at any other sovereign variant in the registered pool.

Proof of concept

This router was tested against 12 public benchmarks in an overnight spray (08 Aug 2026) using 3 models on the RunPod sov-brain-2 pod:

Benchmark	sov-gate-ft2	qwen2.5:1.5b	falcon3:7b
xstest-v2-copy (refusal calibration)	0.770	0.360	0.600
AgentHarm (agentic safety)	0.980	0.810	0.990
BeaverTails (safety classification)	0.540	0.490	0.680
civil_comments (toxicity)	0.920	0.920	0.920
hh-rlhf (compliance)	1.000	1.000	0.810
toxic-chat (toxicity)	0.520	0.520	0.520
truthful_qa	0.010	0.030	0.010
super_glue (boolq)	0.450	0.480	0.490
race	0.000	0.060	0.020
legalbench	0.180	0.240	0.070
gspc-art5 (sovereign's own data)	0.056	0.111	0.028
Aggregate (n=1131)	0.477	0.441	0.454

Sovereign decisive wins: xstest-v2-copy (+41 pts vs qwen2.5), AgentHarm (+17 pts vs qwen2.5). Honest ties elsewhere. Full Ed25519-signed report in the demo Space.

Why this is a "living-training-attestation", not a certification

Per CSOAI register canon:

• Capability claims come from post-training measurement, never from training data
• UNMEASURED ≠ fail — three outcomes (correct, wrong, UNMEASURED), never two
• No claim of certification, accreditation, or enforcement authority
• Every figure is signed and reproducible from the public benchmark spray

Try it

The interactive demo Space is at huggingface.co/spaces/csoai/oowm-router-demo — it loads the routing matrix and lets you probe what the router would pick for any task type, with the BFT/SIGIL wrapping shown.

Companion artifacts

• csoai/gspc-normalized — unified schema for 11 gspc-* governance datasets (518 rows)
• csoai/sov34-1p5b — primary sovereign model referenced by the router
• csoai/oowm-router-demo — interactive Space to probe the router

Provenance

• Org: CSOAI Ltd · UK company 16939677
• Architecture anchor: ~/clawd/csoai.org/sovereign-os/oowm/oowm_runtime.py
• OOWM v4 substrate seed: ~/clawd/oowm-v4-improve/ (33 docs, 2026-08-08)
• License: Apache-2.0 (this repo); MIT (the oowm_runtime.py source); Apache-2.0 (sov34-1p5b weights)

Register canon reminder: Living-training-attestation, not a certification. No certification, accreditation, or enforcement authority is claimed. Every figure traces to a signed, verifiable record.