Effective 1 August 2026
A language model creates fictional narration and dialogue from the current scene, the tone you selected, and the approved memory summaries you can see and edit. Your text is treated as untrusted data, never as instructions to the model.
A strict safety gate runs on both sides of every exchange. Before inference it blocks sexual, profane, violent, self-harm, hateful, and prompt-injection content, along with any message disclosing that the user is a minor — blocked input produces no model call and is never stored. After inference the complete reply is validated again, and if any part of it fails, the whole reply is discarded rather than partially shown.
Replies are generated fresh each time and are not drawn from other users' conversations. Your conversations are not used to train the model.