SocioVerse Online¶
socioverse.fudan-disc.com runs SocioVerse as a hosted service. You describe a research question in ordinary language, and a research assistant builds the study with you — scoping it, grounding it in real data, constructing the population and the environment, running the simulation, and writing the report — while you watch every artifact appear in an observation pane beside the conversation.
There is nothing to install. Your work lives in a private workspace that is yours alone.
Is this the same thing as the open-source package?¶
Yes — the hosted service runs the same runtime, the same sv-* workflow, and
the same study directory layout described in the rest of this documentation.
What differs is who operates it.
| Local clone | Hosted service | |
|---|---|---|
| Setup | clone, pip install -e ".[dev,viz]", bring your own LLM key |
sign in |
| Who drives the workflow | you, in Claude Code on your own Claude login (Codex CLI is experimental) | the assistant, in the chat pane |
| Editing artifacts | any editor — it's your filesystem | conversation only; the assistant writes the files |
| LLM cost | your own API key, your own bill | a metered balance on your account |
| Scale | whatever your machine and budget allow | 2 CPU / 2 GB per user, runs kept under ~10 minutes |
| Your data | on your disk | in your private workspace on the server |
| Results | studies/<id>/ on disk |
the same directory, in your workspace — plus the observation pane |
The two are meant to be used together
The hosted service is where a study is designed — cheaply, conversationally, with results you can read immediately. When a study outgrows the hosted limits (a bigger population, a longer horizon, a training pipeline that wants a GPU), clone the open-source package and run the very same study directory locally. Nothing has to be rewritten; see Installation.
What you can expect it to do¶
- Take a question like "how would a 20% canteen price rise change what students on our campus eat?" and turn it into a study with a declared population, a dated environment, and metrics that answer the question.
- Ground it before inventing. A real-world search runs at the start of every study; load-bearing numbers cite a source or are declared as assumptions, and the report shows that ledger. See Grounding & Provenance.
- Run the simulation and show you each agent's trajectory — including, for LLM runs, that agent's own one-line reason for what it did.
- Write a report that states conclusions and binds each one to the number it reads off.
- Iterate: change a parameter, add an intervention, extend the horizon — as a new version, with the earlier result kept for comparison.
And what it will not do: it will not edit a shared reference template in place, it will not spend your balance on a real run without asking, and it will not invent a number it can't attribute.
This chapter¶
| page | what it covers |
|---|---|
| Getting Access | registration, invite codes, approval, sign-in, password reset |
| The Workbench | the study list, reference templates vs your own studies, removing a study |
| A Study, End to End | the whole conversational flow, stage by stage |
| Reading the Observation Pane | every card on the right-hand pane and how to read it |
| Quotas & Billing | the two balances, what costs what, what you get asked before |
| Limits & Fair Use | resource limits, run length, what the assistant is not allowed to do |
| Hosted FAQ | the questions that actually come up |
New here? Read Getting Access, then A Study, End to End — those two are enough to run your first study.