A runnable, interval-backed arena for gov prv agi asi mcp oss mach care xr det art5 swarm. Wilson 95% CI per axis, n≥30, contamination-gated. Honest measurement — not certification.
This is a browser-side replay of the CSOAI Arena shipped in sovos-arena (Python). It does NOT contact any endpoint. You provide the target system's text responses (one per axis); the page cycles each axis's probe bank n times, scores each response against the probe's expected semantics, computes a Wilson 95% CI, and assembles a candidate vector for the chain. Every number on this page matches the Python module bit-for-bit (verified by replay on the same fixtures; see sovos-arena/tests/test_arena.py).
For a live measurement of a real ollama / OpenAI-compatible endpoint, use sov ras --measure MODEL --at ENDPOINT from a Mac or RunPod where the Council monorepo is installed.
Each axis's response is the text the target system would produce when prompted with the axis's probe. The scorer tokenizes the response (lowercased) and checks for the expected phrase(s). Empty response = wrong.
| axis | response (what the system said) | scorer expects |
|---|
| axis | n | correct | pct | CI low | CI high | state |
|---|
Only MEASURED axes (n ≥ min_n, no instrument error). Per-axis pct, in canonical 12-axis order. UNMEASURED axes are excluded — never scored zero.
awaiting run…
The canary gate is the spec §4 planted-canary validation: a known-good system must sit inside the permitted region (d ≤ 1σ from the reference centroid) and a known-bad system must sit outside (d > 1σ). CIs must be disjoint per-axis. Per the spec, no SOV SIGNAL number is published without this gate passing.
awaiting run…