# How web search works

Web search is optional grounding, not the whole product. New runs start with grounding **off**. Turn it **on** so the population sees what already exists.

## Grounding on (Scout)

When **Web grounding** is on, the lab searches the public web **once up front** (Scout) and compresses hits into a short evidence pack. Casting and divergence see that context. Later, **finalists** can get an assumption check against those sources. This pass is meant to be cheap: it does not grow linearly with population size, and it is not “deep research.”

Turn grounding **off** if you want a faster run, you are exploring a private or fictional brief, or you will research only the minds you click.

## Click a mind

After nodes exist, click one. That seat can run **web search** for everyone signed in (hosted trial included). It is a bounded lookup as that perspective — useful when you want sources for one idea, not for the whole field.

## Deep research

Deep research is a longer, more expensive search loop on a **clicked mind**. It stays behind a **saved Cerebras key** and an explicit opt-in on that seat. The hosted trial does not unlock it. Use it when one finalist needs a harder look at the web, not as the default for every node.

## What this is not

The lab does not browse the web as 500 independent agents. Scout is one shared pass. Per-mind search is something you start on purpose. Treat citations as leads to check, not as proof.