Model card
Format from Mitchell et al., "Model Cards for Model Reporting" (2019). Method 3.1.0, fingerprint 85d55227ee29cf170a0aba8314199879f2d946e64244edbbcebe477dfc12fe73.
Details
A rule-based score built from published word lists, fixed anchors and a published formula. No trained model and no chat model produces any score.
Intended use
Describing how plainly news reports and outlets word their coverage, so readers can compare reports of the same event. It measures wording only. It is not offered for advertising exclusion lists.
Factors
Results are reported by outlet and by layer. The audit also reports results by outlet lean, using AllSides Media Bias Ratings labels that appear only there.
Metrics
Layer scores and an overall score from 0 to 100; outlet scores with a 90% range; the tests on the audit page.
Evaluation data
The human rating panel: left, center and right raters who do not see which outlet wrote each passage. Agreement figures appear on the audit page as ratings come in.
Reference data
The anchors: 50 Federal Register abstracts and 78 opinion articles split evenly by side. The word lists, with the newsroom rule behind each entry, are on the method page.
Quantitative analyses
16 outlets currently have a score. The audit page reports the name-swap test, list balance, anchor parity and weight sensitivity.
Ethical considerations
A wording score can be read as a verdict on an outlet. The site shows uncertainty ranges, publishes every rule, and takes corrections and publisher replies on the feedback page.
Caveats
Word lists miss some context; outlets that block automated reading are scored on headlines and summaries; the score does not measure which facts are included.