A scripted GOD experiment on the PKU map.
Ordinary campus life, then a manually-injected public lecture event.
🌏 中文 · 📜 Operator Script / 操作脚本
⚠️ What-if framing. The experiment uses a hypothetical May 2026 China-visit framing only as broad background context. The campus lecture, route, delegation dialogue, student questions, and answers are AI experiment output. Do not publish them as a real itinerary, real speech, or real policy.
This GOD experiment is set on the PKU map package. It starts with ordinary campus life, then lets the operator manually inject a public notice: Donald Trump will visit Peking University and speak at Centennial Hall with a delegation that includes Elon Musk, Jensen Huang, and a coordinator.
GOD_EXPERIMENT=pku_trump_visit GOD_EXPERIMENT_RUN=1 ./scripts/god.sh new-runThe control room opens at:
http://127.0.0.1:5174/pixel-replay/pku_trump_visit/1
- Map package:
custom/maps/pku - Map id:
pku - The PKU map package is a local asset folder and is intentionally not
stored in this repository. If it is absent, GOD falls back to
the_ville. - Key locations:
west_gate,weiming_lake,boya_pagoda,library,centennial_hall,teaching_building,dormitory,canteen,gymnasium,lab_building,admin_building,campus_green.
The experiment defines 22 fixed-ID agents:
| Group | Count | Examples |
|---|---|---|
| 🎓 PKU campus | 18 | students, faculty, staff, media, club organizers, a visitor |
| 🛩️ Delegation | 4 | Donald Trump, Elon Musk, Jensen Huang, delegation coordinator |
The delegation agents are stylized fictional simulation personas. They do not represent real statements by those people.
The full step-by-step prompt sequence lives in
OPERATOR_SCRIPT.md. The high-level beat is:
- 🌱 Paste the daily-life director prompt, run the first 1–2 live steps to show ordinary campus life.
- 📣 Paste the visit-notice intervention from
OPERATOR_SCRIPT.md. - 💬 Run one or two steps to observe campus discussion.
- 🚪 Paste the delegation-arrival intervention, then gather everyone to
centennial_hall. - ❓ Use the Q&A blocks: each question is first delivered as an agent-to-agent
direct_message, then followed by a targetedAskfor a clean answer suitable for screen recording. - 📰 Finish with the aftermath and summary prompts.
The in-scene prompts are intentionally written in a natural campus tone. Keep publication disclaimers in captions, README text, or video descriptions — never have the characters repeat them in dialogue.
cd agentsociety
uv run python scripts/validate_map_package.py custom/maps/pkuStatic expectations:
init/init_config.jsonusesmap_id: pku.- Every
initial_locationsentry is one of the PKU map location ids. init/steps.yamlstarts at2026-05-15T08:20:00+08:00.