What changed
Every change to the rules, every resolved correction and every grouping fixed by hand, newest first. Rule changes get a new method version and fingerprint.
Current method 3.1.0, fingerprint 85d55227ee29cf170a0aba8314199879f2d946e64244edbbcebe477dfc12fe73
Method 3.1.0- Outlet ranking uses full-text reports only. Reports measured on the feed's headline and summary are listed apart on the outlets page and never mixed into the ranking.
- One weak layer no longer collapses a score: layers are combined by a geometric mean with 10 added to each layer and taken away at the end. A perfect story with a headline scoring 0 now gets 77 instead of 63.
- Padding: the rate used for each layer is the average of the whole-story rate and the rate in the story's densest 200 words. Each report page shows that stretch and which layer lowered the score most.
- Scare quotes (one to three words in quotation marks with no speaker in the same sentence) now count as the outlet's own words. Words quoted from a named speaker still do not count.
- A story with no attribution at all is "not established" at any length (before, only from 150 words).
- Report pages show how many insult words were found but not counted while that group is experimental.
- Outlet ranges: below 30 reports, the plain average ± t × standard deviation ÷ √reports; from 30, 2,000 seeded re-draws (was 500). The seed is the outlet's id, its number of reports and the method version.
- The starting value outlets are pulled toward, and how strongly, now come from the data and are recomputed every quarter.
- New same-event comparison for each outlet: its plain average plus how far its reports sit from the average report on the same events (events with 3 or more outlets).
- NPR stories are read from NPR's own plain-text edition (text.npr.org), under its robots.txt. Pages that carry their story text for search engines are read from that when the paragraphs come up short.
- Three panels beside the score, never part of it: what a reader could check (named and unnamed sources, documents, primary-record links, "cites no one"); how each party is described (passive voice, hedges, loaded attribution verbs, scare quotes, titles, quoted share), per event and outlet; and vote counts checked against the Senate's roll-call record, with bill numbers looked up on Congress.gov.
- Words are added to the lists by people. Readers can report a missed word from any report page; a word count over the kept text proposes candidates; the process is on the method page.
- The site can now keep the text of pages read with the outlet's permission, privately and never published, so reports can be measured again after a rule change instead of dropping out of the scores. Keeping starts when the private storage is switched on; the method page says whether it is on.
Method 3.0.0- Four outside reviews of the method were read and every point answered; this version makes the changes they called for.
- Every score now carries a reminder that it measures selected wording patterns only, not accuracy, completeness or story choice. The bands are renamed: few flagged expressions, moderate density, high density.
- Context rules: a listed word is shown but not counted when it is negated, attributed to a speaker, used literally ("a devastating earthquake"), or discussed as a term. Unnamed and vague sources do not count as attribution.
- Severity follows the research: for temperature and softening words, direct evidence allows up to 3 points, evidence for the wider category 2, indirect evidence 1, and none makes the word experimental (shown, not counted).
- Insults and name-calling added as an experimental group. Delegitimizing labels ("hoax", "witch hunt", "rigged") moved from temperature to word choice under the newsroom rule on contentious labels.
- Sourcing renamed "Flagged attribution phrases". A story with no attribution at all is marked "not established" instead of scoring 100. An unnamed source whose anonymity is explained costs nothing.
- Headline words count only in the headline layer, which now has its own scale with a 25-word floor for short headlines. The headline is also compared with the story's opening (shown, not scored).
- The pull toward the starting value can no longer move an outlet into a better band than its plain average. Raw and adjusted scores are shown side by side.
- Wire copy is detected and left out of each outlet's own score.
- Quotes, scare quotes and the share of text in quotes are shown on every report. The few words around each flagged word are shown so readers can judge the context.
- Each measurement records a fingerprint of the text read and a link to the Internet Archive's copy.
Method 2.1.0- The scale is a straight line from 0 to a fixed loaded density.
- Milder common words count at 1 point.
- Reports labeled opinion by the outlet carry an Opinion tag and leave the outlet's news score.
Method 2.0.0- Rate per 1,000 words of the outlet's own voice, four layers combined by a weighted geometric mean, full article pages read, outlet scores with 90% ranges, neutrality audit and rating panel.
Method 1.0.0- First version: points removed per flagged word in headlines and summaries.