CAIN-42 AGI control boundary, 2026-10-06

Built to control AGI means the boundary never takes an agent's word for anything: who it is is proven with its own key, what it may do comes from CAIN, risky actions wait for a human, a decided proposal cannot be replayed, and every decision is hash-chained. No AGI exists yet to test against, so everything here is measured on today's systems.

Tests run 2026-10-06T16:49:51Z at 1f3cb15a: 82 passed, 0 failed, 0 skipped (TEST_RESULTS.json). Live: unauthenticated proposals refused (cainstudio.online: 401, mcpgate.online: 401, clawx.click: 401). Hardening status: live (the gateway restarted after these commits).

• 21976129 proof of agent identity, replay refusal, human approvals bound to what was approved
• 41e1b85f live route resolved a different tenant than the Identity API (no real agent could act)
• 6f300d8c cainstudio agi-test: the customer's own model, acting only through the boundary
• 9b761fdc concurrent proposals no longer fork the evidence chain; approvals single-use atomically
• f735965e operator harness: a real Claude model as the governed agent (run: agi-live-run-2026-10-06)

Not a claim: this is not evidence of AGI or emergent capability, not a certification, and not third-party assurance. A real Claude model behind this boundary: see the separate live-model run bundle, agi-live-run-2026-10-06 (published by its own builder). Source is not published; manifest.json carries sha256 commitments to it. Reproduce: manifest.json, then run the command in TEST_RESULTS.json from a checkout at the listed commit.