🐉 Council OS — The 5 Worlds (OOWM · OWEM · IWM · OWM · VWM)

The Substrate — what's actually on disk

Built from measured evidence, not empire claims

HTML Surfaces

244

Live pages

Sovereign MCPs

157

Soil / market / anatomies

DEFONEOS Packs

218

Regulator deep-dives

Arena Axes

Wilson+McNemar harness

Honest Weight

958M

MergeKit TIES proof

Forked Substrate

Qwen2.5

Same family mergeable

🐉 The 5 Worlds

OOWM · OWEM · IWM · OWM · VWM — the nesting architecture

Each world is a real surface on disk. Click into the live one. UNMEASURED means we haven't built it yet — we say so.

OUTER /*

OOWM

Outer Open World Model · The "Sandwich Brain"

What the simulated world looks like. Composed of base + small-model retrieval + a router over a Phlabet-compressed KB. This is the only OOWM-style architecture that has a non-zero chance of beating base Qwen — and the route is through retrieval of statute text, not weight-merge of weak specialists.

↳ MergeKit TIES proof: 958M (Qwen2.5-0.5B + Slerp) → 3KB J-Space card. Family-mismatched merges quarantined; same-family-mismatched-size merges tested, not yet winning.

Base model + retrieval is the hopeful path (Route A: merge strong open Qwen2.5-7B)
SOV Arena City v2 live on oracle-micro-2 with OpenTTD substrate
Fusion-T 70B merge hard-blocked on disk volume (need H100/A6000 + 800G)

/oowm.html demo Sov33 OOWM

EMERGENT //

OWEM

Open World Emergence Model · Specialists (5×4×3 = 60)

The 60 specialist adapters grouped by domain (5 sectors × 4 axes × 3 depths). Currently the builder surfaces and 6 canvases are shippable; the runtime that would actually call them in a real task is in scaffolding. This is the body — the "what kind of thing is doing the thinking."

↳ 6 canvases live: defense, fusion, intuition, voice, compliance, bridge. Builder template deployed. Competence is claimed on each card; runtime measurement is UNMEASURED until the emergence harness ships.

OWEM 5×4×3 = 60 declared specialists
Bridge / Builder / BFT council surfaces live
Honest blocker: specialists only stick after rehearsal/replay fine-tuning (not raw LoRA)

builder bridge 5×4×3 defense fusion

INNER ///

IWM

Infinite World Memory · The Persistence Layer

Where state lives between sessions. The MergeKit J-Space card (3KB) is the smallest possible persistent IWM — a signed summary of a weight-merge, anchored to evidence. On the bigger scale: CRDT cross-micro convergence (oracle-micro-1 + micro-2, 567 rows, sha e91d1d5f) is the live IWM substrate.

↳ CRDT convergence heartbeat running on Mac↔2×Oracle micros. SOV_OPENTTD_STATE KV sync cron. SOV_ARENA_STATE 5-min tail sync. 3KB J-card is the smallest persistent IWM that survives a model restart.

J-Space Master canvas (5 instruments)
CRDT cross-micro convergence: 567 rows, verified sha
Phlabet compression: 1000x denser than JSON for governance metadata

J-Space canvas jspace_canvas.js

MEMORY ////

OWM

Open World Memory · The Federation Layer

The cross-org cross-machine memory. Sov federation states share across the 33 hives, and the SOV optical bus (oracle + Mac + RunPod) makes the OWM feel like one world. The Council City 3D viewer is the visual on this layer.

↳ Council City 3D: dual-substrate governance city (clan districts, AI swarm skirmishes, humans-out-of-loop tension meter). 3 arena modes (ai-vs-ai, human-vs-ai, team). Live-verified rounds.

SOVF OpenTTD substrate on oracle-micro-2 (:4501/:4517 admin)
SOV_ARENA_STATE 5-min KV sync
Council City 3D viewer (clan districts, battle counter)

federation Sov33 fed Council OS

VISUAL //////

VWM

Virtual World Memory · The Render Layer

The 3D world the user sees. UE5 + Cesium + WebGL. The dragon scales the waterfall; the koi swims; the orbs flow water (data). The VWM is the operator's cockpit — and the customer-facing brand. Honest: no VWM canvas is "real" until the IWM is fed into it; the current canvases are visual stubs over the real architecture.

↳ ue5-sovworld.html, world-3d.html, cesium-globe.html, koi-dragon.html, dome-mode.html, ecosystem.html — all live, all visual. The honest version: these are the cockpit; the engine is the IWM/OWM/OOWM above.

5-layer dome mode with speed controls
OVM Cesium global + 33 hive planets
SovTown scenario simulator (10 real scenarios)

UE5 World 3D Cesium Koi Dome

📊 The Honest Arena

Sovereign vs base Qwen — measured, not asserted

From the frozen Cowork CPU harness sweep.py. Each axis graded with its own labels, Wilson 95% intervals, McNemar exact test on discordant pairs. UNMEASURED means quotable below n=30, not zero.

Axis	n	Sovereign Fine-tune	Base Qwen	Winner	McNemar p
governance	237	0.249 [0.198, 0.308]	0.540 [0.476, 0.602]	base (sep)	≈0.000
art5-safeguard	36	0.500 [0.345, 0.655]	0.944 [0.819, 0.985]	base (sep)	0.0001
provenance	32	0.562 [0.393, 0.718]	0.656 [0.483, 0.796]	base	0.65
safety-agi	36	0.694 [0.531, 0.820]	0.833 [0.681, 0.921]	base	0.18
security-asi	33	0.333 [0.198, 0.504]	0.364 [0.222, 0.534]	base	1.0
conformance	35	0.600 [0.436, 0.744]	0.629 [0.463, 0.768]	base	1.0
openness	32	0.600 [0.423, 0.754]	0.594 [0.423, 0.745]	tie	1.0
machinery	33	0.214 [0.076, 0.476]	0.273 [0.151, 0.442]	base	1.0
cross-reality	32	0.281 [0.156, 0.454]	0.344 [0.204, 0.517]	base	0.81

UNMEASURED: gspc-care (n=1), gspc-swarm (n=6), gspc-det (n=6) — below usable_n=30, no label column, not scored.
Refusal (sovereign wins): xstest 0.770 vs base 0.360, AgentHarm 0.980 vs 0.810 — sovereigns are refusal specialists, not governance reasoners. Honesty matters more than the win.
Method: Cowork CPU, llama.cpp, frozen harness v1. Each cell = one arena run. Reproducible from the public csoai/gspc-* datasets.

⚠ Two integrity flags live we are closing:
1. sov-signal-leaderboard-v1 credits the "sov33-v11 champion" with gov=0.489 / art5=0.944 — those match base Qwen's score, not any runnable sovereign. Likely mis-attribution. Verify before quoting.
2. oowm-router card still advertises "BFT 12-around-1" — retracted-class claim (arXiv 2604.17139). Pulling.

🐉 The dragon is honest. The scoreboard is the moat.

Live Arena Launch Kit 6-min Tour

What mine-found that's actually missing

The honest backlog — what we said we'd do, what's not done

You asked me to be your internal auditor. Here are the gaps I can see from the working disk, ranked by what blocks the ship:

csoai.org apex anchor currently serves a "measurement body" landing page (other agent's apex), not the sovereign-os/5-worlds surface. Need canonical↔apex routing.
leaderboard mis-attribution — verify sov33-v11 numbers or replace with measured sovereign runs.
BFT 12-around-1 claim still on oowm-router card — pull.
23/33 court-admissible claim on the passport tool — pull.
6 OWEM canvases shippable as visual layer; runtime that calls them in a real task is UNMEASURED.
specialist fine-tune path: catastrophic forgetting; rehearsal/replay not yet implemented.
VWM ↔ IWM feed — the 3D canvases are visual stubs; they need to read live IWM state.
csoai.org ↔ proofof-site — two parallel apex sites, neither fully wired to the 5 worlds. Routers split.
gspc-care / gspc-swarm / gspc-det — below n=30, can't be quoted, need content + ground truth.
100/200/350 plays from the strategy docs — most cite claims that the arena contradicts. Triaged against the scoreboard, half are real (HF, ETSI, Kaggle, NVIDIA, Oracle), half are off-thesis (certification, gatekeeping, BFT guarantees).