What counts as proof that AI work happened?

What it proves: that the AI produced a sentence. Nothing about what ran, what changed, or what exists now.

What it does not prove: that any tool ran, that a file was written, that a database holds what the sentence claims it holds.

When it is enough: when the work is trivial, free to redo, and nobody will ever ask about it again.

What it proves: what was said in the conversation, and what a screen showed at the moment someone captured it.

What it does not prove: what actually ran. A transcript can be cropped, reordered, edited, or generated after the fact, and even an honest one shows the talk about the work, not the work itself.

When it is enough: progress updates between people who already trust each other, on work where a mistake costs little.

What it proves: to its maker, a great deal: the steps, the times, the tool calls, as recorded by the maker's own system at the time. For your own debugging and your own history, this is a genuinely useful record.

What it does not prove: anything to a second party. The checker and the checked are the same party. The log lives where the maker controls it, the software that reads it is the maker's, and an entry can be edited, deleted, or written after the fact without leaving a mark anyone else can detect. It also tends to die with the machine, the account, or the subscription that holds it.

When it is enough: inside one party's own operation, where the only person who needs convincing is the person who made the record.

What it proves: the saved result was not changed since it was saved, because the result is committed when it is saved and any later change shows. It shows which steps ran, in what order, with what outputs, and, when a client decision is recorded against it, the acceptance state.

What it does not prove: that the work was correct, good, or true about the outside world. A receipt is evidence about the record, not a grade on the work. That limit is part of the definition, not a footnote.

When it is enough: whenever the parties do not fully trust each other and still need to settle what happened: a client and a contractor, a team and its agent, you today and you in three months.

A worked example you can open

This record was produced by Cairnworks, an independent verification studio, in the course of its own deliveries, not as a test for this page: zambo.dev/run/28c79cde-20a0-48a3-b4ec-e5d0e3ac104d. Open it and check it. No account, no permission, none of the producer's tooling. That is the whole test of the tier: the checking works for a stranger.

What counts as proof that an AI did the work?

Four tiers of evidence can sit behind an AI saying the work is done. Tier 0 is the AI's word, which proves nothing beyond the sentence itself. Tier 1 is screenshots and chat transcripts, which show what was said, not what ran. Tier 2 is a private log or trace on the maker's own machine, checked by the maker's own tooling, where the checker and the checked are the same party. Tier 3 is a record a stranger can check without the maker's permission or tools: the saved result is committed so any change shows, it resolves at a public URL, and it verifies against a public spec. Only Tier 3 settles a disagreement between parties who do not trust each other, and it is the tier the word receipt should mean. The one-question test for any evidence you are handed: can I check this without the maker's permission, machine, or software? If not, it sits below Tier 3 whatever it is called.

Is a screenshot of the AI finishing the task proof?

It is Tier 1 evidence: it shows what the screen displayed at one moment, in a conversation someone chose to capture. It does not show what actually ran underneath, and it can be cropped, reordered, edited, or produced after the fact. A screenshot is fine for a progress update between people who trust each other. It settles nothing when the answer is disputed, because the dispute is exactly about whether what was said matches what happened.

My AI tool keeps its own logs. Is that a receipt?

It is a log, which is Tier 2. A log is a real record and a useful one for your own debugging and your own history. It is not yet a receipt in the strong sense, because it lives where its maker controls it, the tool that checks it belongs to the same maker, and an entry can be edited or deleted without leaving a mark a stranger can detect. A log reaches the receipt tier when the saved result is committed so a change would show, the record resolves somewhere a second party can open it, and the check runs against a public spec rather than the maker's own say-so.

Does a verifiable receipt prove the AI did good work?

No, and any page that tells you otherwise is overclaiming. A verifiable receipt proves the saved result was not changed since it was saved, and shows which steps ran, in what order, with what outputs. It does not prove the work was correct, useful, or true about the outside world. Verification answers one question: is this the record that was actually made? Judging the work itself still takes review, tests, and an acceptance decision, which is why acceptance belongs inside the record when two parties are settling up.

How do I get Tier 3 records for my own work?

Run the work somewhere that saves a record as it goes, instead of reconstructing one afterwards. When work runs through Zambo, every tool call is saved as an AER-1 record at the moment it runs, and anyone holding the link can check it free at https://zambo.dev/verify/ with no signup. For $0.99, a Day Pass covers 24 hours of recorded work, no subscription. Ongoing use is the Zambo Pass at $29 a month or $290 a year, and $19 a month for $ZAMBO holders, which adds private handover links, your branding, and the approval record as PDF and JSON files.