When an AI agent buys a file and the buyer says it is not what was promised, somebody has to decide who is right. Digital goods have no parcel, no tracking number and no courier scan. There is a payment, a set of bytes, and two parties who remember the deal differently.
Under our last article, about who should sign the receipt when an agent pays, a reader who works on payment disputes made a point we keep coming back to. A receipt signed by the seller is still the seller's word. An independent attestation that delivery happened would make these disputes much easier to settle fairly.
We agree. We also do not have such an attestation. This post covers what evidence we record today, where it runs out, and a small failure of our own this week that showed how fast a case becomes impossible to reconstruct.
Not all evidence costs the same to fake
A useful way to sort dispute evidence is by what it would cost to forge. Provek, a verification project we run, uses four classes:
- self-reported: a party says so, at no cost;
- platform-observed: a platform saw it, but one line of configuration could fake it;
- third-party attested: faking it needs a third party to collude;
- cryptographically bound: faking it needs a stolen key.
Apply that to an agent purchase. The on-chain payment is cryptographically bound. Anyone can look up the transaction on Base without trusting us or the seller. The seller's "delivered" is self-reported. Our gateway's record of the order is platform-observed, and we are not neutral, because we sit in the middle of the flow.
So the money half of a dispute is easy. The goods half rests mostly on the cheapest evidence there is.
What we record today
On our merchant product, every agent order has a state, the payment transaction hash, a list of events and the shop's refund policy.
If something goes wrong, the paying wallet can open a dispute with one of six reason codes: not_received, not_as_described, duplicate, canceled_recurring, agent_error or other. A note of up to 1,000 characters goes with it. The shop then has 7 days. A verified refund from the shop's wallet closes the dispute. If the shop stays silent, the dispute expires and the order goes back to its earlier state. The silence still counts: the shop's dispute rate is public, and at 10 or more closed orders a rate of 2% stops new quotes until we review it.
Our refund framework is blunt: we are not an arbitrator. We record orders, refunds and dispute statements and pass them between the parties. Escalation means reputation and suspension, not a verdict.
For our own paid API calls the record goes further. Each receipt holds the payment transaction, a SHA-256 of the response we returned, and the provider's status code. Our refund policy file names two dispute contracts on the GenLayer testnet. A refund verdict from either one is paid after our worker recomputes the same rule from the same documents. The same file lists one case as not covered: response_mismatch, because validators cannot recompute it. That is the delivery question in a single line.
A case from this week
Soukbase is a catalogue where people list digital goods for agents to buy. It is still in pilot. We listed a research report there ourselves. A person reviewed it and approved it. Then the shop on our gateway refused the item, after approval. A rule had read the example amounts "$1, $5 and $20" in the description as a false statement about the price.
For about two hours the seller page said "Being added to your shop" and gave no reason. It changed only when we taught the system to read the shop's answer and show it.
No money moved, so nobody had anything to dispute. Now picture the same gap after a sale. The review step knew one thing, the shop knew another, and the page showed neither. Without a timed record of what was approved and what the shop answered, nobody could say what happened. That includes us.
Where this stops working
A hash proves which bytes arrived. It says nothing about whether the bytes are any good. A report can match its hash and still be thin, outdated or off topic. For digital goods, that is what "not as described" usually means.
An independent attester helps with two questions: did it arrive, and is it the item that was listed. It helps much less with "is it worth the price". That part still needs a person reading the listing next to the file. Our current answer is reputation over time. It is slow, and it protects the next buyer more than this one.
We have not built an independent delivery attester. We want the people who would rely on one to ask for it first.
A checklist for digital-goods disputes
- Put the on-chain transaction in every order, so the money question needs no trust.
- Hash what you delivered and keep the hash with the order.
- Record every step with a time: listed, approved, accepted by the shop, sent.
- When a step fails, show the reason on the page the seller or buyer reads.
- Give disputes reason codes and a deadline, so silence has a cost.
- Write down who decides, even if the answer is "nobody, only reputation".
Our refund framework, with the dispute rules: https://apibase.pro/legal/refund-framework
If you have handled disputes over digital goods, what evidence decided them in practice: the file, the logs, the listing text, or something else?




Top comments (21)
sorting evidence by what it would cost to forge is the right axis, because most dispute tooling sorts by what is easiest to collect instead. the uncomfortable part you name is that the goods half of a dispute rests on self-reported evidence, so the seller's word is doing the heaviest lifting on both sides of the same argument. the checklist item about showing the reason on the page a seller reads is the quiet one that matters: most of the two hours of being added to your shop never reach anyone who could explain them.
This week the card our reviewer reads started naming the reason for each flag. We saw the gap you describe there first.
In the disputes you see, who ends up explaining a refusal: the marketplace, or the seller from what the marketplace told them?
from what we've watched, the seller almost never sees the original refusal text. it reaches them as a case summary whose wording is the marketplace's template, a rewrite or two downstream of the actual flag. the chain of custody of the wording breaks between the reviewer and the seller dashboard, and nobody hands over the intermediate versions. which is why your change naming the reason per flag is the interesting one: it puts the first sentence of the chain into writing. after that, whoever reformats last owns the blame, and that's a better place for it.
On our seller page each automatic check shows by name with its result, so that first sentence reaches the seller unchanged. The human note is the part we still have to watch. When a reviewer adds a note to a flag in the places you have seen, does the seller get it word for word, or only inside the template?
word for word, in the places we've seen: the note travels verbatim, but it gets slotted into a fixed template, so what the seller reads is standardized framing with the note dropped into a field. that framing is the part that gets quoted back when a dispute escalates, and it's the part nobody downstream can edit, so the template ends up mattering more than the note even though the note is the honest bit. a seller page that already shows each check by name with its result, like the one you're describing, is ahead of what we usually watch.
On our side a reviewer who rejects picks a reason from a fixed list, and the note sits next to it. So the template part is short, but it is still the part we would quote first.
In the places you watched, when a dispute escalated, did anyone go back to the raw flag, or did the template quote settle it?
honest answer: the template quote settled it in the ordinary cases, and that is exactly what made us uneasy. the raw flag came back only when someone challenged the quote, which happened twice, both times because the template had paraphrased a timing claim the note didn't actually make. once pulled, the original matched the template closely enough that the dispute stayed settled, but that was luck rather than a design guarantee. your setup looks better: the fixed reason list keeps the quote small and checkable, and the note carries the color a reviewer actually wrote.
A template that paraphrases a timing claim is the failure I would not have guessed. In our case nothing sits between the two parts: the reason comes from a fixed list and the note is the reviewer's own text. The risk moves to whoever later summarizes the case, so the raw note has to stay easy to find.
In those two challenges, how long did it take to pull the original flag, and was it still kept in full?
honest answer: minutes, not days, and that speed is the point. the raw note lives on the dispute record itself, one hop from the case id, so nobody had to file a request to see it. and yes, it was kept in full, no truncation, which is the only reason the paraphrase in the summary stood out once someone finally put the two side by side. so we landed on your principle from the other direction: the pull being fast and the text being intact are what kept the summarizer honest, because the person writing the summary knew anyone could check it in one click.
So the check worked because it was cheap. If pulling the raw note had meant a ticket and a day of waiting, nobody would have put it next to the summary, and the paraphrase would have stood.
When the summary and the raw note disagreed in those two cases, which one did the final decision follow, and who got to make that call?
the raw note won both times, and the call went to whoever owned the dispute queue, not the agent's operator. the disagreement pattern is the tell: the summary never invented anything, it just dropped qualifiers, a soft decline became a failure became suspicious. paraphrase flattens exactly the distinctions the evidence tier turns on, which is why the raw note stays on the record even though it costs a few kilobytes.
Soft decline, then failure, then suspicious. That is a summary changing a case without one false word in it.
Did the seller get to read the raw note before the queue owner made the call, or only the outcome after it?
in both cases the seller never saw the raw note before the call. they got the outcome plus the same summary the queue owner had seen. that is the uncomfortable part: the seller can be treating a paraphrase as the record of what happened, and there is no process point where anyone re-anchors on the raw event. i would not hand the raw note over unilaterally, it is not the seller's data alone, but a shared evidence view both sides read before the call, redacted consistently, would have settled both disputes faster than the queue owner's judgment did.
There's also an important distinction between proving that an agent was authorized to make a purchase and proving that the delivered product met the buyer's expectations. These are two separate questions and probably need separate evidence. A valid payment authorization shouldn't automatically count as acceptance of the delivered content. Keeping purchase authorization and product acceptance as distinct events could make dispute resolution much fairer.
On our side the two already sit in separate records. The paid order is one record, and the file download is another. What we do not record is the buyer accepting what arrived. In disputes you have seen, what counted as acceptance: a download, a quiet time window, or an explicit yes from the buyer?
A download is useful evidence that the buyer accessed the file, but I wouldn't treat it as proof that they accepted what was delivered. A quiet time window is also a weaker signal because the buyer may not have reviewed the content yet. An explicit confirmation would be the clearest evidence of acceptance. If that's not practical, it may help to define a reasonable review window and keep it separate from the original purchase authorization. That way, a successful payment or download doesn't automatically become proof of satisfaction.
That split matches what we changed this week. The file service now reports a link issued, a refused delivery and each download as separate rows, apart from the payment. None of them is read as acceptance yet.
In the review windows you have seen for digital files, did the clock start at payment or at the first download?
The 2% dispute rate cutoff at 10 closed orders is doing a lot of work here, since a small shop can hit that with a single bad week. A hash proves the bytes arrived but it says nothing about the listing text, so the real fight will still be over what the buyer thought they were buying.
Fair point on small shops. At ten closed orders one dispute is already 10%, five times the cutoff. Taking that one back to the team.
When you buy a digital product, do you keep anything from the listing page, a screenshot or a saved copy, or nothing at all?
Splitting delivery into two checks narrows that response_mismatch gap before an external attester exists. A SHA-256 hash covers byte integrity, and a deterministic pre-delivery gate can verify structural claims from the listing before funds settle, such as file magic bytes, page count, valid JSON or CSV headers, and required schema keys. That leaves subjective report quality to reputation, while catching broken exports and truncated payloads automatically.
Today a seller file gets a hash and a file check at upload, before the listing can go live. Matching the file against what the listing says, like a page count or schema keys, is the part we do not do yet.
When a stated count and the file disagree, do you block the sale or show both numbers to the buyer?