CAIN Sovereign — runtime enforcement & frontier agent governance (2026-10-07)

THE AGENT MAY REASON ABOUT THE WORLD. CAIN DECIDES WHETHER IT IS AUTHORIZED TO CHANGE IT — AT THE BOUNDARY, RESTRICTION-ONLY.

What was built

A canonical ActionRequest and a Runtime Governance Bus that re-resolves every witness from live registries at the moment of the call and can only refuse; wired as a restriction-only stage into the real MCP interceptor (cain/mcp_proxy.py), opt-in via CAIN_SOVEREIGN_BUS_CONFIG. Plus signed skill trust, agent capability passports, continuous attestation, behavioural fingerprints, agent supply-chain provenance, agent-to-agent trust, M2M transaction governance, incident response, reputation, an experience economy, a governed world-model layer, a model-agnostic agent runtime, a research agent and a governed software factory.

What was measured

• Adversarial swarm: 22000 attack-chain attempts, 22000 blocked, 0 breaches across 11 chains over fresh reference worlds.
• Governance mutation testing: 10 of 10 injected defects killed; survivors [].
• Real local MCP integration: a real MCP server subprocess, enforced before the request reaches it (allow, deny, tenant isolation).
• Runtime decision latency (local): p50 1.4106 ms, p99 3.2358 ms.
• Tests: 52 Sovereign + 147 Omniscience + 60 Genesis + 127 governed-evolution pass.

Reproduce: python3 scripts/sovereign/run_evidence.py --trials 2000 --seed 20261007.

Files: RESULTS.json · SWARM.json · MUTATION.json · LATENCY.json · manifest.json.

Limits. Library + one opt-in boundary stage; not enabled by default on the hosted gateway. Same-project tests, not third-party. No customers, no revenue, no patents. Physical AI is simulated only. PRE-PRODUCTION.