Formula and limits

How Haze turns rating evidence into reader signals.

Haze does not publish a single master score. It combines source ratings, story-cluster coverage, confidence gates, and visible limitations so readers can see what a label means and where it stops.

What goes into a displayed signal

Source records, story shape, and confidence checks combine into a reader signal. The interface should help readers ask better questions, not replace their judgment with a hidden score.

Evidence, not verdicts

Source rating evidence

Bias, factuality, ownership, geography, and confidence labels start from approved source-rating records and keep provider attribution visible.

Story cluster shape

Haze compares how many indexed articles and independent sources are in the cluster before showing a strong story-level signal.

Coverage distribution

Left, center, right, local, original-reporting, and follow-up coverage are counted separately so one label does not flatten the story.

Confidence gates

Weak source matches, stale records, disputed provider data, and thin clusters reduce certainty instead of becoming a precise score.

Where the formula stops

Reader judgment required

Publication ratings are not article verdicts

A source label summarizes publisher patterns. It does not prove that a specific article is true, false, complete, or biased.

Clusters change as coverage changes

A story can gain new sources, lose blindspot strength, or shift factuality context as Haze indexes more reporting.

Provider records can disagree

When licensed or public rating sources conflict, Haze keeps attribution and confidence context visible rather than hiding disagreement.

Local relevance is contextual

Edition, region, and local publisher signals help rank coverage, but they do not make national stories equally relevant to every reader.

How to use the labels

Verification checklist

Open the source roster before trusting a single summary label.

Compare article roles: original reporting, wire copy, local context, and follow-up coverage.

Treat low-confidence or thin-cluster labels as prompts to keep reading.

Use the source directory and rating history when a publisher label looks surprising.