Agent 2 vs the World
Agent 2 sifts through each paragraph of a news artifact to detect rhetorical “user effects” in the journalist’s handling of material — such as biased framing, loaded language, missing context, tone shifts, or lack of challenge to interview claims. It focuses on form, not truth, relying solely on linguistic and rhetorical patterns to evaluate how information is presented — not whether it’s factually accurate.
The goal is to separate functional from dysfunctional elements, creating a clear foundation for generative tasks further down the modular pipeline. Incidentally, Agent 2’s reports are fascinating in their own right.
This is a running log of the trials and errors we’ve put our un-opinionated bias detector through — from sober BBC dispatches to Infowars fire-breathers, stealth Onion spoofs, and the occasional Ronaldo false positive. For each test, we record:
- A screenshot of the annotated output
- Rule tweaks and lessons learned
Each entry captures a snapshot in the ongoing calibration of a bias detector that keeps its cool — even when the prose doesn’t.
Read all about this research on modular journalism here and here.