How it works

Five stages. One of them is a person.

Anyone can dispatch a human. The hard part is coming back with something software can rely on — a place proved by fixes the server recorded, a time proved by the server's own clock, and a model's reading that is advisory and can never release money alone.

Intent

An agent hits something software cannot do.

Photograph a property. Confirm a storefront is open. Check a shelf. The model can reason about all of it and cannot see any of it. It calls one tool, or posts to one endpoint, and names a standard job and a place.

What this does not proveNothing has been verified yet. At this point the platform knows an intention, not a fact.

The Tag

The money is real before anyone drives anywhere.

The card is authorized immediately and the Tag becomes visible to qualified Partners nearby. If the charge fails, no Tag is created — a Partner must never do real work for money that is not there. Spending ceilings are checked inside the same transaction that takes the payment.

What this does not proveA funded Tag is not a promise that anyone will take it. If nobody does, it expires and the money returns.

Execution

A human tags in, and the clock starts.

A qualified Partner accepts and goes. Proof is captured inside a capture session the server issued, with the server recording device position throughout — not a coordinate the phone reports afterwards, which is a number anybody can type.

What this does not proveA capture session proves where a device was. It does not prove who was holding it.

Verification

Three locks, and only two of them can pay.

PLACE and TIME are deterministic: was the device inside the geofence, and was the proof captured inside the window? Both are facts the server established itself, and both gate the payout. CONTENT is a model reading the photographs — it can permit auto-approval or route to a human, and that is the whole of its authority.

What this does not proveNo content verdict is never a pass. An AI model never moves money on this platform.

The result

The agent gets its answer and carries on.

The bell rings, the Partner is paid, and your webhook fires with a structured result — outcome, a per-item checklist, the three lock verdicts separately, and evidence hashes. It is an input to the next step, not a photograph somebody has to look at.

What this does not proveThe evidence packet states its own limitations. A record that overstates itself is worse than none.

The part people ask about

Why the model cannot approve a payment

Because the moment it can, the product is no longer verification — it is a photograph and an opinion, and you would be right not to trust it with money.

PLACE and TIME are deterministic facts the server established. CONTENT is a judgement. A judgement can raise confidence enough to skip a human on a small, low-risk Tag, and it can send anything uncertain to a person. It cannot overrule a geofence, and a large bounty goes to a human however confident the model is.

What happens when something fails

  • Place or time fails — nobody is paid, and neither are we. Approving it anyway requires an administrator and leaves a permanent audit record.
  • Nobody claims it — the Tag expires and the authorization is released.
  • You disagree with the outcome — open a dispute and the money freezes, both ways, until a human rules.

Watch it run.