Conversation Simulation
Orchestrated synthetic discussions with inspectable participation, source use, and final ballots.
AnthroSim runs one moderated conversation path. Choose one or more audience groups, with 1-50 generated participants per group, and set a simulated duration from 5 to 120 minutes. The same engine supports town halls, focus groups, panel discussions, jury deliberations, and market research interviews.
During a run:
- Private decision traces stay visible to the operator. Each trace records the simulated update, salient point, speaking gate, and evidence decision in fixed fields. These are operator-only simulation records.
- Participation is measured. The record shows who spoke, passed, or was called on. Each participant keeps a small set of discussion points that decays within the run.
- Source use is inspectable. Participants and the moderator can request a bounded lookup from enabled project material. The run stores the query, the exact text consumed, its provenance, and whether a source-backed belief update passed the evidence check.
- Final ballots are exact. Private mode hides earlier ballots from later participants. Public mode reveals them in sequence. Every ballot includes a position, rationale, and confidence.
Example: fictional civil jury
Topic: Ruiz v. Northstar Markets · Format: Jury Deliberation · Participants: 12 · Duration: 75 minutes
Timeline-focused juror The last signed inspection was at 5:35, and the fall was at 6:12. The blank 6:05 line matters because the policy calls for a check every 30 minutes.
Evidence-skeptical juror It matters, but Boyd changed his time after looking at a receipt and cannot identify the employee. I want to separate the inspection gap from proof that the liquid was already there.
Moderator The packet confirms the 37-minute inspection gap. It leaves the liquid’s source and arrival time disputed. Which liability element should the room address next?
Illustrative transcript based on AnthroSim’s fictional Ruiz v. Northstar Markets trial packet. It is not output from a real jury.
Project and web references
Upload a file, paste text, or import a page as project material. Participants and the moderator request short passages when a question arises; materials are not silently treated as facts.
When a web reference mode is selected and Brave Search is configured, a lookup can request current web context after project materials fail to answer. The run records the returned snippet and provenance. Important facts still need review against the original source.
The discussion research found a 45 percent error reduction for one private-thought configuration on one model family, while another model became worse. Evidence tests also found unsupported source claims in 19 of 34 source-claiming turns. Those limits are why model changes require qualification and source use remains visible in the product.
Next: Voting and results →
Request Early Access
We're onboarding a first cohort: academic and applied researchers, market research firms, enterprise product teams, law firms, policy organizations, and teams planning sensitive or regulated studies.
Early access participants receive preferred rates and priority onboarding.