One question.
Three answers.
Where they
diverge.
Model choice, generation parameters, and steering — exactly as they were applied this turn.
Run a prompt above to see the live divergence.
Why the outputs
actually differ.
This isn't prompt engineering. The divergence is structural — mode selection, memory scoping, model routing, and persisted steering all happen before the first token is generated.
Mode selected
The user chooses a personality — or the system infers one from context and time of day. This determines which voice, which memory vault, and which model will respond.
Memory scoped
Each mode retrieves only its own memories plus shared facts. HONNE sees companion cues. YAMI sees late-night confessions. SHIN sees strategic goals. No cross-mode leakage.
Model routed
Different modes can use different LLMs. HONNE may run on GPT-4o for warmth. SHIN on Claude for precision. YAMI on Hermes for unfiltered output. Same character, different engine.
Voice adapted
Tone, length, formality, and directness are all mode-dependent. Steering values are persisted per mode and applied to the system prompt at generation time.