Can CueCrux hallucinate?

When you ask CueCrux what an agent did, the answer is not generated. It is replayed from signed records.

A model asked to recall events produces a plausible reconstruction; that is what hallucination is. A CROWN receipt trail is a different kind of object: each receipt was signed at the moment of the action, hash-chained to its predecessor on an append-only spine, and can be verified offline. A fabricated or altered entry does not slip in quietly; it breaks the chain, and the verifier says so. Incident review in CueCrux is replay, not recollection.

Two honest boundaries. First, CueCrux does not stop the models inside your agents from hallucinating in their own outputs; no external system can. What it does is make everything the agent did (every tool call, every write, every refusal) a matter of signed record, so a bad output can be traced to the actions and inputs around it rather than argued about.

Second, retrieval returns what your fleet stored. If an agent wrote a wrong fact yesterday, retrieval can surface that wrong fact today; the difference is that the receipt tells you exactly who wrote it and when, so it can be corrected at the source.

Where CueCrux itself answers questions from knowledge, in WikiCrux at wiki.cuecrux.com, answers are cited and the citations open their sources.