MURM simulates public opinion locally

The sharp part here is not the “agents debate a document” gimmick; it is the attempt to make crowd simulation reproducible, inspectable, and private enough to use on sensitive material. Local NetworkX and ChromaDB keep the workflow off third-party memory services, which matters for policy, internal strategy, or confidential docs. The demographic archetypes, quota sampling, and multi-seed branching directly target the usual swarm-agent failure mode: everyone converging to bland neutrality. Live Wikipedia grounding plus GraphRAG make the output easier to audit, though the quality of the seed context still matters a lot. Entropy, polarization, and opinion velocity are the right kinds of metrics if the goal is research or decision support rather than a flashy demo. The big question is calibration: if the reports stay stable across seeds and actually predict real-world reactions, this could be more than an interesting local toy.