Research program · working name
Meridian
The machine authority strand of Society of Explorers. We study how people grant, limit and withdraw permission, alongside Discover Truth’s inquiry into our origins and beliefs.
Start with the current evidence and corrections, then follow the questions below. This is research, with no demonstrated Meridian advantage.
- Who gives an AI agent permission to act?
- Can a machine hold resources or enter an agreement that can actually be enforced?
- Who can stop it?
- How do independent humans and machines make a decision together?
- Can governance be an open protocol?
Separate supplement · · local demonstration
Reserve. Deliver. Settle once.
A requester asks a machine service to hash a payload. Permission, a synthetic budget and an execution slot are checked together. A local worker performs the computation; the result is checked before the synthetic charge settles once.
Two implementations, the same rules. An ordinary atomic central reservation service is the reference. A separate mandate representation receives the same signed requests, limits and event order. Both retain running liability, protect retries and honor the supplier’s local execution limit.
Nine categories. Sixteen fixed variants. Both controllers match every tested outcome and state. Four deliberately broken controls fail independent checks. The successful fixtures include 14 real local jobs; one shared worker service supplies verified receipts to both controllers. No Meridian advantage is demonstrated.
SHA-256 jobs run in two real Node worker threads. Prices, identities, clocks and capacity quotas are synthetic. This does not test physical infrastructure, distributed security or production payments. Frozen REB v0c remains separate.
Read the demonstration and assumptions · Inspect the recorded results
Inspect or reproduce the local demonstration
The source uses Node’s built-in modules and makes no network calls. Start with the fixed cases, the central reference and the mandate representation.
cd research/meridian/demos/machine-commitments
node run.mjs
node --test conformance.test.mjsHypothesis
Unknown Not inevitable
The hypothesis is whether a coordination network for humans, organizations, and AI agents could take on functions now handled by governments, firms, banks, and platforms. Governance — how those actors coordinate, decide, and settle a dispute — is not the same thing as government.
Working rule
A fluent recommendation is not yet an action, and a record that has not faced a challenge is not yet settled. That matters because an agent can be wrong, and a shared record nobody can dispute will treat the error as fact. What has to survive a challenge before it counts?
What you can point at
Identity
A name that can be checked: a person, an organization, or an agent acting for one of them. If that name is missing, permission has nobody to attach to. That matters because a later dispute needs someone who can answer. Who is actually acting?
A written limit
A written limit on what an agent may do. A step past that writing does not become an allowed action in the record. Stopping it in the world still needs local enforcement and protection. That matters because a model can recommend something the agent was never permitted to do. What is this agent forbidden to do?
A way to challenge evidence
A way to dispute a record before it is treated as settled. A claim nobody can contest hardens by silence. That matters because an unchallenged error becomes the record everyone else has to live with. Who can say this record is wrong?
Ordinary records
Money and obligations kept as ordinary records — a debt, a payment, a promise — rather than a new coin. A new coin does not, by itself, make the obligation enforceable. That matters because a bill still has to be payable by someone a counterparty can recognize. If the obligation is only a new coin, who can enforce it?
Reputation that is not designed to be sold
Standing meant to stay with the party whose record survived a challenge. No design fully enforces that: keys, accounts and devices can still be sold or rented, and that is the known way it fails. A reputation you can buy describes the sale, not the conduct. That matters because transferable standing lets someone purchase a history they did not earn. Whose actions does this standing describe?
A local safety stop
A rule that a remote decision cannot override a machine's local safety stop. The machine may still refuse. That matters because an instruction from far away can be wrong about the limit in front of the machine. What is this machine still allowed to refuse on its own?
No token. Intake closed.
No token is being launched or sold. Research intake stays closed. This page does not ask for money.
Related reading already on the site: the cancel has not arrived · customer of record · physics still vetoes · when a ledger earns its keep · what the audit left unsupported.