Two operations.
One settlement grammar.
A cross-chain transfer and an agent workflow have almost nothing in common. Walk them both through the settler and they take the same six verbs, the same escrow, the same outcome set and the same receipt. Only two things move between a transfer and an agent job: the evidence lane, and who grades it.
Start here, because this lane is decodable today. The failure it prices: It never arrives, and the refund sits in limbo while support absorbs the ticket.
Move 1,000 USDC to Base and pay out to a named recipient.
Ask what it would cost
Returns the derived terms: the bond floor, a premium quote, the referee lanes available, and how much cover capacity is left. It promises nothing forward and expires with the meter epoch.
The bond floor comes from a failure rate measured on this corridor, not assumed for it.
TIMEOUT is not UNVERIFIABLE. One is the worker failing to deliver; the other is the record failing to see. They pay opposite parties, and conflating them would make darkness profitable.
Now the same machine, pointed at work that never touches a chain. The failure it prices: The workflow reports success and the business result did not happen. Nobody independently checked, because until now nobody could. The node and the resolution engine are specced, not shipped. Everything else on this page is identical to the lane above.
Enrich 500 leads through a paid data API and write the results to a CRM.
Ask what it would cost
Returns the derived terms: the bond floor, a premium quote, the referee lanes available, and how much cover capacity is left. It promises nothing forward and expires with the meter epoch.
Measured per (kind, lane): an operator with settled history quotes better than one without.
| XFER | AGENT | |
|---|---|---|
| Evidence lane | CHAIN — the destination event, decoded against the anchored root | NODE — a signed envelope per execution, emitted inside the workflow |
| Who grades it | PROOF — checked against the root; no human judgment involved | AUTO for machine-checkable specs; a bonded evaluator only where judgment is required |
| What acceptance seals | The payout tuple: asset, amount, recipient, chain | The row schema and count, sealed before the run |
| Does the worker integrate? | No. The buyer can insure this unilaterally — the service need not know | Yes, one node inside the workflow. Consented, and minutes of work |
| Status today | Real — the chain lane is decodable now | Proposed — the node and the engine are specced, not shipped |
That is the whole claim, and it is falsifiable: if a new kind of operation needed a seventh verb, or a fifteenth outcome, or a different receipt, the grammar would not be a standard — it would be a bridge feature wearing a costume. A kind that needs an operation the table lacks is rejected as a spec bug, by design.