You may be looking for this note (http://www.stateofthemedia.org/2005/methodology.asp) on the methodology, particularly on the section labeled "Intercoder Reliability Testing for Text Media" (they seem to use similar methods for other media).

My guess is they'd use the former criteria: A story that reports bad behavior is considered "negative," good behavior "positive." This is far more reliable than determining how well a story "should" treat it's subject, and comparing it to how well they do treat it, a far more subjective standard.