This project started as a research initiative aiming to construct digital twins of all 650 Members of Parliament. The system predicts how each MP would vote on a bill using LLMs for reasoning.
Each twin is built from the public record. Parliamentary speeches and interventions from Hansard. Voting history from every recorded division. Select committee work, Early Day Motions, and written questions. Constituency demographics and electoral margins. Public statements, media appearances, and social media positions. No private or privileged data is used.
How an MP votes is shaped by multiple competing forces. Party loyalty and whip direction. Constituency interests and electoral vulnerability. Personal ideology and policy commitments. Committee expertise and subject knowledge. Prior positions taken in debate. The model captures these factors not as independent variables but as interacting pressures that the MP must reconcile.
Every MP is modelled as a three-layer agent. The first layer captures core identity — biography, party affiliation, electoral history, and constituency profile. The second encodes belief and reasoning patterns — ideological positioning, rhetorical tendencies, and policy priorities extracted from their parliamentary record. The third is contextual — bill-specific information, whip signals, and political conditions injected at simulation time.
When predicting a vote, the system retrieves relevant historical precedent — prior divisions on related legislation, the MP’s own speeches on the topic, and party-level signals. This retrieval-augmented generation pipeline ensures each prediction is grounded in real parliamentary behaviour, not generic language model knowledge.
The twin receives the bill, its own profile, and retrieved context, then produces a voting decision with structured reasoning. The system predicts individual MP votes, not aggregate outcomes — enabling rebellion detection, swing vote identification, and detailed scenario modelling.
Paper in preparation. Citation details will appear here on publication.