Skip to content

Sourcerers Technical Documentation

Sourcerers is an experimentation platform for measuring how LLMs select among politically diverse news sources under controlled prompt conditions.

This documentation is organized for three audiences:

  • Developers extending experiment logic and analytics.
  • Researchers interpreting robustness and bias signals.
  • Operators deploying the API/dashboard stack.

Core Components

  • Experiment pipeline: incident preparation, condition bundle generation, prompt construction, model execution, and parsing.
  • Analytics engine: FastAPI endpoints + shared metrics logic backed by the experiment database.
  • Dashboard: Streamlit interface for run ingestion, diagnostics, model comparison, and narrative summary.
  • Report assets: generated plots/tables under docs/figures for paper/report workflows.

Where To Start