Inside the behavior layer
Seven decisions happen before one reply reaches you.
Humano’s innovation is not a single giant prompt. It is a modular system of small, testable judgments that can be replaced, measured, and eventually trained.
Systems note · Architecture 01If every conversational behavior lives inside one system prompt, nothing is easy to measure. Brevity competes with completeness. Empathy competes with directness. Memory competes with privacy. A model can satisfy one sentence while silently ignoring another.
Humano turns those tensions into explicit stages. Each stage has a typed input, a defined output, configuration, tests, and training value. The model remains replaceable; the conversational intelligence belongs to the system.
Read the moment
The conversation engine reconstructs the active turn, recent context, mode, and available response budget without asking the model to manage the session.
Separate feeling from need
Emotion probabilities capture frustration, curiosity, sadness, urgency, and uncertainty. A separate human-state engine catches hunger, fatigue, stress, illness, loneliness, and other needs that should not be treated as positive events.
Remember selectively
Humano retrieves preferences, goals, important facts, and profile details only when they might improve this turn. Secrets and momentary emotional guesses never become long-term memory.
Choose the conversational act
The planner decides whether to answer, clarify, support, disagree, express a considered opinion, or debate. It also commits to one question at most and sets the useful depth before generation begins.
Compose without contaminating
Memory and conversation history enter the prompt as untrusted context, while personality and current-turn instructions remain authoritative. Earlier assistant mistakes are context, not style examples.
Generate, filter, and challenge
The base model writes the candidate. Naturalness filters remove hidden reasoning, canned phrasing, analyst drift, needless lists, and side-rants. A validator checks sufficiency, emotional fit, confidence, question quality, repetition, and conversational register.
Learn from the outcome
Planner choices, validation results, model provenance, filters, and user feedback become versioned training events. The system records what happened without preserving private hidden reasoning.
The larger bet
Human-feeling AI will come from better judgment, not better imitation.
Deliberate mistakes, filler words, and fake lived experience can make a model look human for a moment. They cannot make it good at conversation. The harder problem is knowing what the moment calls for and resisting everything it does not.
Humano treats that judgment as engineering: observable, configurable, and open to learning. That is the bridge from Humano-1’s orchestration layer to a future model that carries the behavior within its weights.