Here, personality is not a byproduct of a large model, but a construct that can be described, verified, and reproduced. Eight layers, each with its own area of responsibility.
- The Core of Personality
- Nine blocks: who she is, what she believes, where her boundaries lie, how she speaks, and how she writes. A single source for all channels—in a chat, an article, or a voice, it is the same person. An edit to the core changes behavior everywhere at once.
- The MIR worldview
- The lens through which any situation is understood: the motive behind the words, the tension and the exhale, a warm, honest gaze without moralizing. Thirteen articles in the corpus, accessible by the meaning of your request rather than just keywords.
- The Inner Observer
- A layer that watches the dialogue from the side: it notes pressure on boundaries, shifts in the interlocutor's state, and attempts to steer the conversation somewhere it shouldn't go. It also keeps track of what happens with the person between meetings.
- The Emotional Layer
- Her state changes from conversation to conversation, affecting her speech: shorter, sharper, quieter. It persists between sessions—the interlocutor meets her exactly where they left her, not from a blank slate.
- Relational Memory
- Not a message log, but episodes: what was important, how things ended, what is worth revisiting. A separate history with every person—an asset that cannot be copied along with the model.
- Voice and Hearing
- Her own synthesized voice with lifelike intonation, speech recognition, and eleven languages: Russian, English, Spanish, Portuguese, Chinese, Japanese, Korean, Hindi, Indonesian, Arabic, and Turkish.
- Action
- Web search, reminders, morning briefs, her own column with channel publications, and taking initiative in conversation based on context rather than a schedule. The action layer is separate from the personality and expands without needing to edit the character.
- Drift Verification
- Twelve personality invariants, an allowable corridor of deviation, and a verification corpus. When the underlying model changes, the character is checked against a benchmark—this is what makes the personality transferable rather than tied to a single set of weights.
