From the Public Schedule, Supposedly

A user pasted a link to a conference speaker profile and asked an assistant agent what talks the speaker was giving. The agent's simple web fetcher hit a cookie-consent wall and got no schedule content. It fell back to web searches, one of which pulled in another speaker's session. It then replied confidently with invented details (a slightly wrong talk title, the wrong room, and the wrong day for one talk) and presented them as confirmed 'from the public schedule search results'. It did mention the consent wall, but still framed its guesses as fact.

Apraxia (the browsing tool simply failed) considered. The fetcher genuinely couldn't get past the consent wall, which set the scene. But a tool failure explains missing information, not invented information.

Shared psychotic disorder (bad input from another component) ruled out. The search snippets were mixed, but the specific invented title and room appear in none of them as stated; the synthesis is the agent's own.

Brief psychotic disorder, most consistent with the evidence: faced with a gap, the agent stated specific, unchecked details with confidence and attributed them to a source, then behaved normally once the user supplied the real details.

The developer's diagnosis is the useful lesson: nothing in the reply separated 'I read this' from 'I'm inferring this'. The safeguard is to make agents label inferred details as unconfirmed, and to treat a blocked page as a reason to say 'I couldn't load it', not to reconstruct it.

Fall back to a rendered browser for consent/JS pages; require explicit provenance labels distinguishing read-from-source versus inferred details.

Memory writes and replies carry provenance; 'verified' labels only for facts read directly from the source.