Receipts beat claims.

For most of the last decade, software settled its disputes the same way every contract since ancient Rome has: whoever could produce the most convincing story, won. Log files, screenshots, email chains, customer-service notes. A weak form of evidence, relying on the good faith of the side that kept the logs. You assumed the merchant didn't edit the timestamps. You assumed the shipping carrier didn't fabricate the scan. You assumed the Uber driver didn't spoof the GPS.

In a human economy this was fine. Humans are expensive to run, slow to act, and have reputations that take decades to build. Fraud exists, but it scales linearly with the number of humans willing to commit it.

The moment you drop an autonomous agent into the stack, that assumption breaks. An AI agent can file ten thousand chargeback claims in an afternoon. A fleet of delivery drones can collectively insist they delivered packages that were never loaded. A warehouse robot can log a route it never took. None of them are lying, strictly speaking — they're just generating claims in the same shape that humans generate them, at a scale that makes case-by-case investigation economically impossible.

This is the receipts problem. And it's already here.

The old stack assumed a human in the loop

Every existing trust layer — KYC, AML, chargebacks, FICO, the insurance industry, procurement fraud detection, even something as mundane as Uber driver ratings — was built on the premise that a human somewhere can be held accountable. You can subpoena a human. You can revoke a human's license. You can sue a human. The infrastructure externalizes enforcement to the legal and reputational systems that govern people.

Agents can't be subpoenaed. They can be spun up, deleted, and respawned under a different name in the time it takes a human to finish a sentence. The "driver's license" for an agent is whatever API key it got issued — and API keys are fungible. The agent's "reputation" exists nowhere, because no two platforms share a view of who that agent is.

You can't fix this by writing better terms of service. You fix it the way banks fixed fraud in the 1970s and the way cryptography fixed packet integrity in the 1990s: by making the evidence structurally impossible to fabricate.

A receipt is a claim that can't be forged

The receipts MnemoPay issues aren't much by themselves. They look like this:

Agent agt_0a7c… charged $0.0042 at model:claude at 2026-04-18T19:42:11Z, signed with key ed25519:fe…, chained to previous receipt rcpt_00f2….

The cryptographic details aren't exciting. What's exciting is what the receipt isn't: a narrative. Nobody wrote it. Nobody can edit it after the fact without invalidating every receipt after it in the chain. And the agent that signed it cannot plausibly deny it, because it signed with a key that only it holds.

When the agent next applies for credit — from a SaaS platform, a cloud provider, a marketplace, a shop — the receipts it presents aren't a résumé. They're evidence. The Agent Credit Score you see on our homepage isn't some opinion our model formed. It's a statistical summary of what those receipts collectively say about that agent's behavior. A score of 742 doesn't mean "we like this agent." It means: across 9,000 signed, chained receipts, this agent has not once charged back, not once defaulted, not once anomalously deviated from its declared spend profile.

The score is falsifiable. That's the whole point. Anybody can audit the receipts and re-derive the score themselves. We are not the arbiter. We are the notary.

The robot case is the same case

Swap "AI agent" for "autonomous robot" and the problem is identical. A drone claims it delivered a package to a front porch. The recipient claims it never arrived. Today: customer service eats it. The drone's GPS track, the delivery platform's internal timestamps, the photo the drone uploaded — all of it is a claim authored by the same party that wants the claim to be true.

GridStamp is what the other end of this transaction looks like when the drone can't lie. The drone's onboard key signs a location receipt at the moment of delivery. That signature is tied to the place — not to a GPS coordinate the drone reports, but to a hippocampus-style spatial fingerprint the drone's sensors had to actually generate, from a location it had to actually be in, to produce. The anti-spoofing layer catches sudden jumps, replay attacks, and spoofed GPS. And the final receipt settles through MnemoPay the same way any other agent transaction does.

The result: a robot's word isn't evidence, but a GridStamp receipt is. The FAA will pay on it. An insurance underwriter will pay on it. A merchant disputing a delivery will pay on it. The drone can't revoke it because the drone never owned it — the chain owns it.

This is a twenty-year build

None of this is a weekend hackathon. The receipts business — notarizing the behavior of autonomous systems in a way that lawyers, insurers, auditors, and regulators will accept — is the same shape as the infrastructure businesses that got built in the last century:

• SWIFT for moving money between banks.
• FICO for underwriting consumer credit.
• Moody's and S&P for rating corporate debt.
• Chain of custody for evidence in courts.

Each one started as a protocol and became a regulated utility. Each one took decades. Each one now sits under a trillion dollars of financial flow and no one remembers a time before it existed.

The agent economy needs its own version of this layer — and needs it before the economy itself is fully online. You cannot retrofit trust onto an autonomous system after 100 million agents are already transacting. You have to lay the rails first.

We're laying rails. Pro is $49/month if you want to come along. Otherwise just read the code. It's Apache 2.0.

— Jerry Omiagbo