How to verify an AI agent's output before you ship it

This morning, ten agency founders opened their inbox and found a finished piece of AI work waiting for them: a handover. Not a summary that says "done." A package: the job's steps, a receipt for each step, and the saved results, behind one link they open free and check before they accept anything.

That package exists because of a question every freelancer and agency hits the week they start shipping AI work to clients: how do I verify an AI agent's output before I ship it?

The summary is written by the suspect

An agent's closing message is a claim written by the same system that did the work, in its own words, after the fact. When it says the staging database was deleted, the sentence is the only evidence most people ever get. This morning I published the receipt pair behind one of those claims: one receipt for a call that executed, and one for a call the gateway refused, marked executed: false. The claim said the work ran. The record said it never did. The full pair is here: The agent said it deleted the staging database. The receipt says the call never ran.

Verification has to start below the summary, at the record each call leaves behind.

The pre-ship gate, four steps

1. Match every claim to a receipt. Go through the agent's summary line by line. Each claimed step either has a receipt from the execution layer or it is a sentence you are taking on faith.

2. Recompute one receipt yourself. Open the receipt page, fetch the saved bytes, hash them, compare with the digest filed at execution. It takes a minute and it settles whether the record moved after it was saved. Any receipt checks at zambo.dev/verify with no account.

3. Ship the handover, not the summary. Wrap the job: steps, receipts, saved results, one link. The client picks it up free, recomputes what they want, and signs off or flags the step that fails. The argument about whether the work happened ends before it starts.

4. Keep the limit in view. A receipt proves the saved result was not changed since it was saved. It does not prove the answer was right. The gate catches moved records, missing calls, and work that never ran. Judging quality is still your job, now done against a record that cannot quietly change underneath you.

What it costs to run the gate

Zambo is the verifiable work layer for AI: 153 tools behind one MCP endpoint, 37 on the default surface, and a verifiable AER-1 receipt for every job. AER-1 is the open receipt format Brennan Zambo created for agent execution receipts, published as IETF Internet-Draft draft-zambo-aer1-26, with a public conformance corpus of 143 vectors and 146 checks that any checker can recompute.

Every tool free once a day, plus 3 continue calls. One free wrap-up handover per day, and pickup is always free. The Day Pass is $0.99 for 24 hours, unlimited, which is the price of running the gate on a full day of client work. The Zambo Pass is $29 for 30 days, and the Lifetime Pass is $250, once.

As of this morning the public store holds 139,224 receipts, 1,313 of them minted today. Every one is a piece of agent output somebody can verify before they ship it, accept it, or pay for it.