Human, in theory.
Exploring the space between circumstance and response.
HES
Human Emotion Simulator
Circumstance becomes perspective.
Perspective becomes an evolving emotional state.
HES is an experimental affect-simulation framework built around a language model. It couples a fictional subject’s first-person response with a structured, six-dimensional emotional state—making the relationship between events, interpretation, and reaction visible over time.
Context02
Appraisal03
Transition04
Observation
Context-conditioned appraisal
Each interaction presents the model with the subject’s circumstances, conversation history, current emotional intensities, and response parameters. A circumstance update changes the fictional world; a spoken message addresses the subject within that world. The model generates both a first-person response and proposed intensities for every channel.
A multidimensional affect state
Curiosity, joy, trust, fear, sadness, and anger are represented as six independent values. Each spans 0–100, allowing simultaneous high-intensity states without forcing one emotion to displace another. This is a deliberately bounded representation for experimentation, rather than a complete taxonomy of human experience.
Stateful response dynamics
The first scenario establishes the emotional state. Subsequent interactions blend the previous state with the model’s proposed state. Sensitivity sets responsiveness; inertia preserves continuity. The displayed result therefore reflects both the new event and the state that preceded it.
Structured outputs, inspectable traces
The model returns a response, an observation, and six bounded scores through a structured output contract. Incomplete or invalid outputs are rejected. The Journal records successful interactions, proposed scores, resulting states, and parameters so that changes can be inspected rather than reduced to a single impression.
The experiment
HES provides a way to examine how an LLM portrays emotional continuity, conflicting signals, and changes in circumstance. Its architecture combines a language model with explicit state-transition rules; it is not a separately trained foundation model. Psychological fidelity and consistency remain questions to evaluate, not established properties.
A framework for simulated feeling.
Prometheus is an experimental Human Emotion Simulator (HES). It uses a language model to portray a fictional subject, exploring how circumstances and conversation shape an evolving emotional state.
Set the circumstances.
Describe who the subject is, what has happened, and what is at stake. The model responds from their point of view. Change the circumstances to advance the scene, or speak directly to the subject within it. Earlier circumstances remain in the session context. They describe the fictional subject, not the person using the simulator.
Six independent signals.
Curiosity, joy, trust, fear, sadness, and anger each use their own intensity scale from 0 to 100. Zero means absent; 25 mild; 50 moderate; 75 strong; and 100 extreme. Multiple emotions can be strong together. They do not share a fixed total.
The first scene establishes the state directly. Later responses evolve it through sensitivity and inertia: how strongly new events affect the subject, and how much of the previous state remains.
Beyond the six channels.
Experience labels interpret the displayed emotional state. Contentment and delight express aspects of happiness; dread and horror describe elevated fear; distress represents a mixture of fear, sadness, and anger. Three consecutive qualifying distress states produce the label “Sustained distress.”
These labels are interpretive rules, not additional model-scored emotions or evidence of felt pain or suffering. Each ring reports the intensity of its named core channel, rather than the intensity of the experience label.
Reading the experiment.
Session details include valence and arousal. Valence summarizes the balance of positive and negative channels; arousal summarizes activation from selected channels. Both are calculated heuristics derived from the displayed state, not independent psychological measurements.
The initial state is an illustrative baseline. Preview mode does not evaluate circumstances or generate responses. In live mode, the model supplies a response and proposed intensities; the transition rules determine the state shown on the interface.
Observe. Revisit. Question.
The Journal saves completed tests in your browser, including inputs, responses, and emotion scores. Session details expose the trajectory and model observations. You can export the active session for a separate record. Browser data is local; clearing it removes saved tests.
A simulation, openly.
These scores are model-generated estimates within a simplified framework. They are not validated psychological measurements, clinical assessments, or evidence that the model feels emotion. Experience labels and summary metrics are interpretive design choices.
Live sessions use your own OpenAI API key and send the fictional scenario and conversation directly from your browser to OpenAI. Prometheus holds your key only in the active tab’s memory: it is not saved in browser storage, Journal records, or exports, and it is not sent to the Prometheus server. Refreshing or disconnecting clears it. API usage is billed to your OpenAI account. The interface preview does not run inference.