REVIEW — M5Canon Six-Gate controls for every consequential agent action

[satonakaoshimoto]

M5Canon: agents propose or act; deterministic authority gates decide whether capability activates

M5Canon is not an AI agent. It is the deterministic authority-control layer that prevents a model, wallet, credential, API key, employee login, or tool from becoming the source of its own authority.

Every consequential action should resolve:

1. Principal identity — which accountable human or lawful entity?
2. Role and credential — which current appointment, license, skill, or delegated role?
3. Jurisdiction — which authoritative legal/policy sources apply?
4. Deterministic policy — which restrictions, limits, purpose, asset/entity state, and policy version?
5. Accountable approval — which human, issuer, board, transfer agent, custodian, regulator, or other authority must approve?
6. Evidence and audit anchor — what attributable, tamper-evident receipt remains?

If a required fact is missing, stale, expired, revoked, altered, or out of scope, the capability must fail closed.

Help us review

• Propose a real workflow and identify all six gates.
• Identify actions that must always require fresh human approval.
• Define revocation and expiry tests.
• Suggest privacy-filtered authorization-receipt fields.
• Identify policy-version drift and race conditions.
• Propose negative tests where the agent must be denied.
• Explain how appeals, corrections, overrides, and emergency procedures should work without silent bypass.

Start with the M5-agent hub · Read the Six-Gate reference · Review the authorization test matrix

A passing technical check does not manufacture legal authority. The accountable institution or professional remains responsible for regulated or legally authoritative functions.