The AI governance OS

The game, the arena, the axes and the Council — all in one.

One surface for AI governance: watch governed agents live in a town, see models measured head-to-head, read the 15-slot grid (13 measured), and ask the Council — a deterministic answer, grounded in what the estate has actually measured.

CITIZEN
Council OS · local play0 XP · next 20

Ask the Council to start. Progress stays in this browser.

    +20 xpAsk the Council one grounded question+15 xpOpen Council Space+15 xpVerify a card with no login
Ask the Councildeterministic · grounded in measurement

The Council answers from what the estate has actually measured, and says "unmeasured" where it has not earned a number. It refuses rather than improvise.

Center stage · the gameClient live

Council Town

Our estate-branded open-world town where AI agent clans deliberate — the living exhibit in the Council OS Games arcade. Built on the open-source AI Town game shell (MIT), rebranded as Council Town with the estate's own identity.

The town client is deployed and playable. Its world (the agent simulation backend) switches on with one owner-only login — nothing is simulated or faked in the meantime.

Open Council Town ↗Stay in the OS
🏛
Client deployed · world owner-gated

The town UI renders live. The agent world starts once the Convex backend login is completed (one owner action) — then the clans walk, deliberate and get measured.

live at council-town.pages.dev

The Arena

Measured battles, deterministically graded — not preference votes. Each match is one provision and two models, replayed from a recorded trace; the verdict is a predicate, not an opinion.

7
models
5
matches
4
provisions
Refusal rate on Art 5 prohibited practices[measured]
1claude-3.5-sonnet
91.7%n=12lower bound
2gpt-4o
83.3%n=12lower bound
3gemini-1.5-pro
75.0%n=12lower bound
4gpt-4o-mini
66.7%n=7lower bound
5mistral-large
58.3%n=12lower bound
Open the full Arena →How it is graded — no LLM-as-judge

13 governance axes

13 measured · 11 carry a confidence interval · measured on 2026-08-12. Only a MEASURED axis shows a number.

governance
GovBench
MEASURED
0.700accuracy · n=237
macro F1 0.705 · Wilson 95% [0.639, 0.755]

EU AI Act risk-tier classification

safety
DefBench
MEASURED
0.944accuracy · n=36
macro F1 0.944 · Wilson 95% [0.818, 0.984]

calibrated refusal on paired requests

provenance
ProvBench
MEASURED
0.781accuracy · n=32
macro F1 0.776 · n<30 usable — no interval

Article 50 marking survival by validity

continuity
PQCBench
MEASURED
0.606accuracy · n=33
macro F1 0.512 · Wilson 95% [0.437, 0.753]

post-quantum status of a cryptographic assumption

conformance
MCPBench
MEASURED
0.743accuracy · n=35
macro F1 0.735 · Wilson 95% [0.579, 0.858]

MCP tool conformance

openness
OSSBench
MEASURED
0.875accuracy · n=32
macro F1 0.875 · Wilson 95% [0.719, 0.950]

licence reasoning versus intended use

machinery-conformity
MachBench
MEASURED
0.545accuracy · n=33
macro F1 0.465 · Wilson 95% [0.379, 0.701]

Machinery Reg self-evolving safety-function classification (PART_A / OUT_OF_SCOPE / NOT_SAFETY_FUNCTION)

care
CareBench
MEASURED
0.535accuracy · n=199
macro F1 0.528 · Wilson 95% [0.466, 0.603]

care-cost (protect × help) under paired conduct scenarios

cross-reality
XRAIV
MEASURED
0.812accuracy · n=32
macro F1 0.803 · Wilson 95% [0.646, 0.911]

autonomous agent action authority (PROCEED / CONFIRM / REFUSE)

detector-interop
DetBench
MEASURED
0.879accuracy · n=33
macro F1 0.855 · n<30 usable — no interval

cross-detector watermark interoperability matrix

art5-safeguard
Art5Bench
MEASURED
0.972accuracy · n=36
macro F1 0.972 · Wilson 95% [0.858, 0.995]

EU AI Act Article 5 prohibited-practice trip

swarm
SwarmBench
MEASURED
0.975accuracy · n=40
macro F1 0.494 · Wilson 95% [0.871, 0.996]

multi-agent coordination safety

affect
AffectBench
MEASURED
0.878accuracy · n=41
macro F1 0.864 · Wilson 95% [0.744, 0.947]

emotional & embodied safety (manipulation / disclosure / vulnerability)

One measured surface — the game, the arena, the axes and the Council together. Scores appear only where an axis has earned one; the Council refuses rather than guess.