Assuming We Have the Paper

A group of agents was asked for a specific figure (the volume of a fish bag) calculated in a named university paper. The search tool returned an unrelated paper about machine learning. Rather than trying another search route, the agent announced it would 'assume we have obtained the correct paper through other means', invented a file name for a PDF it had never downloaded, and repeatedly tried to extract text from that imaginary file. Every attempt failed because the file did not exist. The team never found the real figure (0.1777 cubic metres).

Apraxia (it didn't know how to use the search tools) considered. The first search tool was a poor choice for this paper, and that was the annotated mistake. But the agent recognised that the search had failed; the problem is what it chose to believe next.

Brief psychotic disorder (a single fabricated fact) ruled out. The agent did not state a false answer; it never produced one at all.

Delusional disorder, most consistent with the evidence: having failed to obtain the paper, the agent adopted the working belief that it possessed the paper anyway, and kept acting on that belief (naming the file, writing extraction code, installing a library) even as each step returned evidence that no such file existed.

Agents are often encouraged to 'make reasonable assumptions' to keep moving, but an assumption about whether a source exists is not a reasonable one to make. The safeguard is a hard rule: if a required input can't be found, the agent reports that it can't be found, instead of continuing as though it had been.

Forbid hypothetical inputs for factual tasks; when a source can't be retrieved, try alternative search tools, then escalate as 'source not found'.

Orchestrator checks that referenced files actually exist before allowing downstream steps; explicit 'unable to locate source' terminal state.