Every payment system needs someone to decide what counted

The referee and the bank

Every payment has two jobs hiding inside it. One job moves the money. The other decides whether the money should move, and how much of it. The second job is older than the first, and it is the one people argue about.

Call the first job the bank and the second job the referee. Banks have gotten extraordinarily good. Money crosses the world in seconds. The referee, by contrast, has never been standardized. Every era of commerce has rebuilt it, locally, inside whatever institution happened to own the transaction.

This essay walks through those rebuildings, because they make a new idea intuitive: the referee's decision, what counted and what was earned, might be separable from the institutions that have always contained it.

At the market stall, buyers inspected for themselves

The original verification layer was buyer inspection. You held the fruit, checked the cloth, bit the coin, and then paid. Inspection and payment happened at the same moment, in the same place, between people who might meet again next week. Reputation carried the rest.

This worked because fulfillment was observable at handoff. Its boundary was distance and time: the moment goods moved before payment, or payment moved before goods, someone was extending trust with no referee at all.

The taxi meter turned fulfillment into measurement

The taximeter is one of the first machines built specifically to decide what a service was worth as it happened. Instead of arguing about the fare at the end, the meter measured distance and time, and both parties accepted the instrument's reading as the earned amount.

Notice what the meter is: an agreed rule, applied by a device both sides trust, producing a number neither side authored. That is a verifier. Its boundary is narrowness. It measures exactly one thing, and it says nothing about whether the driver took the long way or whether the ride was any good.

Escrow and letters of credit pay against evidence

When trade stretched across oceans, inspection at handoff became impossible, so commerce invented paying against documents. A letter of credit lets a bank pay a distant seller when the seller presents agreed evidence, such as shipping documents, rather than when the buyer sees the goods. Escrow puts value with a neutral third party who applies pre-agreed release conditions.

These are referee systems, and they encode a deep idea Spoolis inherits: separate the agreement about what must be proven from the moment of proof, and name in advance who decides. Their boundaries are cost and rigidity. Letters of credit involve banks, fees, and paperwork sized for shipping containers, not for a $4 data purchase. And the classic escrow agent holds the money itself, which makes the referee a custodian, with everything that implies.

Marketplaces keep the referee inside the platform

Uber, Upwork, Amazon, and their siblings solved verification by owning the entire workflow. Uber can pay a driver quickly because Uber knows the ride happened. It dispatched the ride, tracked the GPS trace, and collected the rider's payment. The fulfillment signal is native to the platform. Upwork can run milestone payments because the work request, the delivery, and the client's approval all happen inside Upwork.

This is the strongest referee model ever built, and its boundary is exactly its strength: the fulfillment signal is trapped inside each marketplace. Every new marketplace rebuilds acceptance, dispute handling, and payout logic from scratch, and none of it transfers. Your five years of verified delivery on one platform mean nothing on the next one. The referee exists, but it is proprietary, and it only referees games played inside its own stadium.

Tests and CI made software acceptance explicit

Software engineering quietly built the most explicit referee culture in any industry. A pull request does not merge because the author says it works. It merges because a continuous integration system ran agreed checks, in a clean environment, and published a pass or fail that everyone can inspect. The criteria are written down before the work is judged. The judgment is reproducible. Disagreements are about the tests, not about opinions.

CI proved something important: acceptance criteria can be compiled in advance, executed by machines where machine judgment suffices, and escalated to named humans where it does not. A required reviewer approval is exactly that escalation. What CI never did is connect any of this to money. The pipeline decides what counted, and then nothing moves except a merge button.

Programmable payments: the bank got better, the referee did not

The newest payment systems are built for software as the customer. Stablecoin rails settle in seconds. Protocols like x402 let an agent pay for an API call inside the request itself, and machine payment protocols give agents budgets and payment sessions. Money movement between machines is close to solved.

But these rails move money on instructions, not on outcomes. A protocol can prove a payment happened. It cannot tell whether the dataset was complete, the extraction was accurate, or the milestone was actually finished. For atomic purchases where delivery and payment succeed or fail together, that is fine, and no extra layer is needed. The moment an agent buys an outcome that can be wrong in several ways, the old question returns with no one assigned to it: who decides what counted?

Spoolis separates the referee from the bank

Spoolis is a bet that the referee can be its own layer. The referee decides what counted. The bank moves the money. Spoolis is deliberately only the first thing.

A Spool writes down, before the work, what must be true, how each condition will be checked, and who has authority to judge the conditions machines cannot. Deterministic checks run like CI. Judgments that should stay human are recorded as explicit decisions by named parties. The output is a signed Outcome Receipt stating the result and the earned amount. Payment stays with the buyer's own payment authority, which can require a valid receipt before money moves. Spoolis determines what was earned; the buyer's payment authority moves the money.

Against the historical systems: it takes prior agreement and named authority from letters of credit, without becoming a custodian. It takes measurement from the taxi meter, applied per condition instead of per mile. It takes compiled, reproducible acceptance from CI, and connects it to earned value. And it asks whether the marketplace referee, the Uber model, can exist without the marketplace: a fulfillment signal that is portable across arbitrary economic work instead of being rebuilt inside every platform.

What a portable fulfillment signal could unlock

These are directions under exploration, not shipped promises. Each follows from the same primitive: a trustworthy, inspectable record of what was agreed and what was earned.

• Payment at the moment of verified completion: not when an invoice ages out, but when the agreed checks pass.
• Per-unit earned value: a thousand records requested, the verified ones counted, the earned amount following the count.
• Milestone payment without a platform: objective checks plus a named human approval, enforceable between any two parties.
• Verified-value streaming: earned value accruing as work accumulates, rather than in one terminal decision.
• Portable outcome receipts: a history of verified fulfillment that belongs to the worker or the agent, not to the stadium it happened in.
• Downstream uses of receipts: reputation, routing, and pricing that read verified history instead of star ratings. These are the most speculative, and they depend on receipts existing at volume first.

What stays uncertain

Honesty about the boundary matters more here than enthusiasm. Marketplace referees work partly because the platform subsidizes disputes and eats fraud. A neutral layer has to make its economics explicit instead. Some fulfillment is genuinely subjective, and no receipt makes a taste judgment objective; the honest move is naming whose judgment governs, which is what the systems that lasted have always done. And a receipt proves the agreed checks ran and what they returned. It does not prove the checks captured everything the buyer cared about. Writing good acceptance criteria is a skill, the same way writing good tests is.

The history cuts both ways. Every era did build its referee, which suggests the job is real and permanent. And every era built it inside an institution, which suggests separating it is hard. That is the experiment.