Show HN: Simulate 1M+ agents to predict pmf

Simulate how a population reacts to your product — and trace every number back to the turn that produced it.

Jahan (جہان) is Urdu for world. It builds a population of LLM personas, each grounded in a real survey respondent, places them on a social network, and lets them meet a product on an X-like feed, a Reddit-like forum and by word of mouth. They react, talk, change their minds, and answer survey waves about whether they would buy it.

📖 Documentation: https://fm882.github.io/jahan/

Recorded, not narrated. Every turn is written to an append-only trace, a tick at a time and whole. Any run can be replayed, resumed or audited.

Prompts are rebuilt, never stored. Any turn's exact prompt is reconstructed from the record and shown only if it matches the hash taken when the turn ran.

Intent is measured, not asked for as a number. Personas answer in their own words; answers are scored with semantic similarity rating against frozen anchor statements.

Channels like the real thing. The feed ranks like X (in-network first, then TwHIN-BERT similarity and age), the forum ranks like Reddit (hot), and word of mouth travels along the network's ties.

Honest about what it knows. Every report states its trust level. Today that level is uncalibrated, and the engine's own holdout found its filled-in attitudes lose to a plain demographic baseline (evaluation). The negative results are published with the positive ones.

What's inside

jahan/ The engine: schemas, brief, population, world, agent, inference, elicitation, runner, trace, analysis, report, web API

frontend/ A local web interface: define who you study, launch a study, watch its worlds replay, read the report and trace

docs/adr/ 48 architecture decision records: why the engine works the way it does

docs/prd/ The specification each module was built from

docs/evaluations/ Everything measured so far, including what did not work

CONTEXT.md The glossary: the words the engine and its docs use

Research use and data

Jahan is a research instrument (ADR 0016). The persona corpus it is built around is released for research only; it is not in this repository, and you download it and accept its terms yourself. Commercial studies need a population you have the rights to.

Copyright 2026 Faishal Manzar. Source-available under the Functional Source License 1.1, Apache-2.0 future license: use, modify and run it for any purpose — including your own commercial studies — except offering it, or something substantially like it, as a competing product or service. Each release becomes Apache-2.0 two years after it is published.