Sources
The roster is the instrument.
Every claim Prism makes is relative to the sources it reads. “The coverage agrees” means these outlets agree; “absent from the coverage” means absent from these pages. That makes the roster itself the most consequential editorial decision in the system, and it is governed by two rules.
Diversity without deletion. A repeated wire dispatch is not twelve independent confirmations, but every relevant article remains in the editorial corpus. Prism does not thin a large cluster to a supposedly representative subset before the editor reads it. The report may identify shared origin, common carriage or near-duplicate accounts as part of its judgment, while the source list preserves the full roster it actually received.
Balance by addition or disclosure, never by subtraction. When the roster leans, the correction is adding voices or disclosing the lean, never quietly removing coverage. The current configured roster of 57 sources spans the spectrum as follows.
CONFIGURED IS NOT OBSERVED
The roster above is the set Prism attempts to read. It is not a claim that every source published on every story or was successfully retrieved that day. Feeds fail, paywalls and page structures change, full text can be incomplete, and conservative clustering can leave a relevant article outside a story. Each published report therefore discloses the smaller realized corpus: every article and outlet that actually entered its synthesis.
Prism does not infer that an outlet ignored a story merely because it is missing from one cluster. A genuine cross-outlet absence study requires a separate recall search over the relevant archives. Daily story maps make only article-bounded absence judgments.
WHAT THE SPECTRUM DOES, AND DOES NOT, MEAN
Lean classifications are working editorial assignments informed by public bias assessments and outlets’ own positioning. They are a roster-balancing and navigation aid, not a reliability score, accusation of dishonesty or universal political coordinate. Left and right already hide several dimensions within one country; international, state-funded, wire and specialist outlets often do not map cleanly onto the same axis. Where a classification is not defensible, the honest long-term answer is unclassified rather than false precision.
As the archive becomes large enough, source records should separate geography, ownership and funding, publication type, wire dependence and political orientation where meaningful. Those fingerprints will appear only when a sufficient run history supports them.
