Hosted FAQ

Questions specific to socioverse.fudan-disc.com. For questions about the framework itself, see the general FAQ.

Getting started

What should my first message say?

State the question, not the mechanism. Say who (which population), what changes (the intervention or the shock), over how long, and what you'd want to measure. For example:

How does a 20% canteen price rise change what students at a Chinese university eat over one semester, and does it push them toward delivery or toward cooking?

That single sentence gives the routing, the horizon, the population, and the metrics something to hold on to. "Simulate a canteen" doesn't.

Your first message also fixes the conversation's language for everything that follows, including the report.

Why can't I edit a reference study directly?

Because it's the baseline every derived study is compared against, and because it's shared. Ask for a change and you'll get a fork — a new study with your modifications, carrying over the artifacts and the grounding ledger. The original stays intact, which also means you can compare your version against it.

Can I upload my own data?

Not through the interface today. If your study needs data the service can't reach, describe it to the assistant — grounding a value as a declared assumption with a stated rationale is a legitimate, documented option. For real file-based population data, run the study in a local clone, where provider_ref: file.personas reads a CSV or Parquet you supply.

During a study

Can I close the tab while a run is going?

Yes. The run continues on the server. Reopening the study replays the conversation and shows where things got to.

The assistant seems stuck. What happened?

Most often it is waiting on a question card you haven't answered — the flow cannot proceed past one. Scroll up in the conversation and look for an unanswered card.

If the service restarted while a question was pending, that card may be gone from the replayed history. Say "continue" and the assistant will ask again.

How do I stop something?

The Send button becomes a Stop button while the assistant is working. Pressing it interrupts the current turn. The transcript records that the turn was interrupted rather than completed.

Can I run two studies at once?

Each study has its own conversation and you can switch between them, but drive one at a time. Simulation runs in particular should not be started in parallel — they share your workspace's 2 CPUs and both will simply be slower.

Results

Where are my results, and can I get them out?

In your study's directory in your workspace, the same as a local clone: a DuckDB store with one panel row per agent per step, the aggregate metrics, the fired events, plus Parquet exports and the rendered report.md with its figures.

There is no download button yet. To take a study elsewhere, the reliable route is to clone the open-source package locally and rebuild it there — everything about the study directory format is identical, and the report contains the numbers and the grounding ledger.

Are results reproducible?

Rule-based and deterministic runs: exactly, seeded end to end. Live LLM runs: no — sampling differs run to run. But every recorded run is permanent in its store, and a branch replays its parent's stored actions exactly instead of re-asking the model. See the general FAQ.

I ran it again and got different numbers. Is something broken?

No — that's live LLM sampling. If you want a controlled comparison, iterate as a new version rather than re-running: the version gate keeps the earlier trajectory and the metric timeline can overlay the two.

What is the difference between a branch and a version? When can I only get a version?

A branch is the counterfactual case. The only thing it changes is the environment, and only at or after a fork step t*: one or more added interventions — a scheduled event or a broadcast — plus, optionally, a longer horizon. Everything before t* is replayed from the parent run's stored decisions, so it is the same people with the same history, no model calls are made for those steps, and only the steps from t* on are billed. That makes the branch and its parent a paired comparison on the same individuals, which the metric timeline labels treatment and control when you compare them.

Every other edit is a version: the current state is snapshotted, and the new version runs cold from step 0. You can only get a version when the edit changes the population, the behaviour function or the decision mode, the seed or the other run settings, or an intervention that fires before the fork step — each of those changes the whole trajectory, so no earlier step can be inherited. Two study shapes can never branch at all, because the engine cannot replay them: studies that wrap an existing simulator, and studies with more than one interaction round per step.

The assistant tells you which exits are available before it creates anything. If a later edit breaks the environment-only claim, the branch is downgraded to a version — it runs cold instead, the version box marks it ⚠ downgraded to version, and the assistant says which artifact broke it. Details in Iterate: versions and branches.

Why does my chart look empty?

A study plots only the metrics it declares for display. If a metric you care about isn't there, ask for it to be added — that's an iteration, and it takes a re-run to populate.

Money

What does it cost me?

Nothing personally. Your account carries two allowances — one for the assistant's own reasoning, one for the simulation and external data — of $10 each by default. See Quotas & Billing.

My balance dropped faster than I expected.

The two usual causes are a large population × horizon (every agent every step is a decision) and a population survey call, which runs one questionnaire per simulated respondent and therefore scales directly with the respondent count. Each charge is recorded as a line in the conversation — scroll back and read them.

I've run out. Now what?

Ask the team for a top-up; both planes can be raised independently. Your studies and results are untouched in the meantime.

Local vs hosted

When should I move to a local clone?

When you need scale (thousands of agents, long horizons), direct file access, your own model provider, or a study whose run belongs to an external pipeline. The study directory is the same in both places, so moving is a copy, not a rewrite. Start at Installation.

Is the hosted version behind the open-source one?

They run the same runtime and the same workflow. The hosted service adds the account, the isolated workspace, the metered balances, and the chat pane — and subtracts direct file editing and unbounded run sizes.