VERIFICATION REVIEW — Agent lifecycle, receipts, conformance, and recovery

[satonakaoshimoto]

How do we prove an M5Agent remained inside authority throughout its lifecycle?

Agent safety must be demonstrated across definition, activation, use, suspension, revocation, replacement, recovery, and retirement—not assumed from a model name or one successful test.

A reviewable lifecycle should cover:

defined → reviewed → credentialed/configured → delegated → activated
→ monitored → suspended/revoked/expired → recovered or retired

Recovery must restore only authority that remains current. It must not silently reactivate an expired role, revoked credential, compromised device, withdrawn consent, obsolete policy, or prior agent grant.

Help us review

• What identifies an agent build, policy version, tool set, model/runtime, principal, and delegation?
• Which receipts are required for proposals, approvals, denials, actions, disclosures, errors, and handoffs?
• Which public conformance evidence can be shared without leaking private content?
• What test vectors cover revocation races, stale offline grants, compromised agents, model substitution, graph drift, and recovery abuse?
• Who may suspend, revoke, replace, or retire an agent, and how is the member notified?
• How should independent reviewers reproduce tests and record limitations?
• What maturity labels distinguish concept, prototype, tested reference implementation, pilot, deployed service, and independently verified implementation?

Start with the M5-agent hub · Review the authorization test matrix · Read the M5POD activation state model

Automated tests support but do not replace domain, legal, security, privacy, accessibility, governance, or accountable human review.