Conversations¶
Residents talk: at home, at work, in the street, in the churchyard after the service. Every conversation is a real event in the simulation, between two or more people the simulation brought together, with a tone (warm, neutral, tense, grief, joy, gossip, prayer) and a topic. What they say is written either procedurally or, if you ask, by a language model that knows who they are.
Watching one¶
Conversations happen in the animated day, at speeds up to an hour a second; switching to a time-lapse speed ends them. On the map, two people talking are joined by a dashed line (red when the exchange is tense), and close up their latest line appears in a speech bubble.
Click one of the speakers, close up, to select the conversation: the inspector shows who is in it, where, its tone and topic, its status, and the transcript line by line with the time. The inspector's front page, with nothing selected, lists the live conversations too.
During a Sunday service the pastor preaches as one long conversation, and when he calls a hymn the congregation sings: the transcript gives the hymn's lines to The congregation, and every voice on the map carries a ♪.
Having a model script it¶
A conversation starts with procedural chatter: short lines from the simulation's own repertoire. Press Script with AI on the conversation and a language model writes it instead, from what the simulation knows:
- the setting: the place, the time, the season, the weather, a holiday;
- each speaker: their age, sex, job, schooling, personality, mood and stress, faith, family, money, health, grief, a pregnancy, and the last few events of their life;
- their relationship, and what has happened in the province in the last ten days;
- the lines already said.
The model is asked for a short exchange in the speakers' voices (South African English is fine). The lines appear at a speaking pace in simulated time, and the conversation lasts until the last line has been said.
Scripts are cached per seed, conversation and model, so replaying a seed costs nothing. Settings › clear dialogue cache empties the cache.
Dialogue modes¶
Set in Settings › Dialogue, and shown on the model chip in the top bar:
| Mode | What happens |
|---|---|
| Off (procedural chatter) | No model is asked; every line is procedural. |
| On demand (the default) | A conversation is scripted when you press Script with AI. |
| Automatic when zoomed in | While the clock runs at an animated speed and you are zoomed in close, the conversation in view is scripted, one at a time, so a local model keeps up. |
Talking to a resident¶
Select a person and press Talk to … in the inspector. Ask them anything; they answer in character, from the simulation's state only: their own card, today's date and weather, the news of the last thirty days, and what you have said so far. This needs a model, whatever the dialogue mode, and replies are never cached.
Choosing a model¶
The model chip opens Settings, where Provider chooses who serves the dialogue:
| Provider | What it needs |
|---|---|
| auto (the default) | Ollama if it has a model, otherwise the first hosted provider with a key. |
| Ollama (local) | Ollama running on this machine. The app picks the first installed model of the gpt-oss, Gemma, Qwen or Llama families, or the one you choose. For gpt-oss, reasoning low is fastest. |
| Anthropic Claude, OpenAI, Google Gemini | An API key; the model is yours to choose. |
| OpenAI-compatible endpoint | A base URL (LM Studio, vLLM, Groq…) and, if it needs one, a key. |
Keys are kept in the app's own storage on this machine and handed to the engine, which holds them in memory; they go nowhere else but the provider. Press save after changing a provider or a key.
The local model loads when the first conversation needs it, not when the app starts, and stays loaded for half an hour after its last use; a 20b model holds about 7 GB of the GPU's memory. If Ollama was started after the app, press refresh in Settings.
Without a model¶
Nothing in the simulation depends on one. Without a model, Script with AI falls back to procedural chatter and says so (model failed — showing chatter), the model chip reads no model, and Talk to … shows the reason under the chat.