What the bias bar measures, where the data comes from, and what it doesn't tell you.

1. The bias bar is downstream of three vendors with documented conflicts.

The 5-axis fingerprint (interventionist, zionist, atlanticist, statist, financialized) and the bloc classification (Western, Non-Aligned, Adversarial) are built on top of bias ratings from AllSides, Ad Fontes Media, and Media Bias / Fact Check. Each is a single-vendor heuristic. Each has a structural conflict of interest we inherit by using it:

• AllSides sells Balance Certification to newsrooms — the same newsrooms it then rates on its bias chart. Client list includes Newsweek, AP, and Axel Springer. Direct conflict of interest between the rating product and the certification product.
• Ad Fontes Media uses a "proprietary, patent pending" weighting algorithm. Its methodology page describes a 3-rater pod system (one right-leaning, one center, one left-leaning per article) but publishes no inter-rater reliability statistic — no Cohen's kappa, no Krippendorff's alpha, no ICC. The 93,200+ article ratings cannot be independently replicated.
• NewsGuard (used secondarily for factuality only) upgraded the Daily Mail from "untrustworthy" to "trustworthy" in January 2019 after a private meeting with a Daily Mail executive. NewsGuard publicly said "we were wrong" on multiple criteria — but a private meeting with a rated party should not flip a content-based rating. As of December 2024, Publicis Groupe (a top-4 global ad-agency holding company) holds a minority equity stake in NewsGuard. 76% of NewsGuard's rated sources are US-based; zero Global South coverage.

2. "Blindspot" is a single-vendor formula, not a peer-reviewed metric.

The BLINDSPOT badge on a story means the upstream clusterer flagged it using a 4-condition heuristic: fewer than 10 left-leaning sources reporting, right-leaning coverage ≥ 33%, left-leaning coverage ≤ (R/L% − 33) × (30/37), and not more than 35% low-factuality coverage. The thresholds (10, 33, 30/37, 35) are stated in prose on the vendor's methodology page. They are not justified, not peer-reviewed, and not independently auditable.

The FAccT '24 literature on recommender systems is the academic framing: "diversity of what" is an essentially contested concept, and any single-vendor score is a normative choice wearing the mask of a measurement. We use the heuristic because it surfaces stories that are structurally under-covered by the working set. We do not claim it is a fact.

3. The geographic spread (geo_gap) is honest about its own limits.

Each story has a geo_gap field with two normalized metrics:

• entropy — Shannon entropy of the geographic source distribution, normalized 0.0–1.0 by the count of non-zero geo cells.
• non_west_pct — fraction of sources that did not originate in the "west" bloc.

What this measures: the spread of the working source set across geographic regions, not the spread of the actual global news ecosystem. A story with non_west_pct: 1.0 means every source we ingested for that story was non-Western — it does not mean the Western press ignored the story (it may not have, or it may have; we don't have the comparative denominator).

What it doesn't measure: global coverage, editorial reach, agenda-setting power, or the relationship between what we ingested and what got missed. It is a self-honest number about the corpus, not a claim about the world.

4. The source registry is where the structural bias actually lives.

The corpus is a working set, not the global news ecosystem. The 1,090-Western / 0-Non-Aligned framing that this site carried in early 2026 was a structural fact — a registry populated almost entirely from US, UK, and European outlets. The number on this page is the current registry, which has shifted through a 2026-07-05 expansion that added 288 sources across 27 regions.

Live registry snapshot · 2026-07-06

loading...

The earlier framing (84% Western, 0% Non-Aligned) was correct as of 2026-07-04 and obsolete as of 2026-07-06. The page updates when the registry updates. Numbers in commentary age faster than numbers in the table.

5. Per-story coverage is closer to 50/50 than 84/0.

Across the 160 stories in the current corpus, 94 have non-zero geo_gap entropy. The world section averages 59% entropy and 43% non-Western coverage. Sports is the actual disaster (entropy 0.17, 8% non-Western). The bias bar is not the whole story — the section matters.

The per-card "why this gap" line is the receipt. Each card shows the working source count, the bloc spread, and the geographic cell count. If a card shows "1 source: 1 Western" and the story is in climate, that's a fact the user can act on. The page does not pretend the corpus is balanced where it isn't.

6. What this site is not claiming.

• It is not claiming the bias bar is a neutral ground truth.
• It is not claiming the BLINDSPOT badge is a peer-reviewed metric.
• It is not claiming the registry is the global news ecosystem.
• It is not claiming the geo_gap is global coverage — it is working-set spread.
• It is not claiming that more sources means more truth. Source quality is a separate axis from source count.

7. What this site is claiming.

• Every story link is a primary-source URL, filed, no inference.
• The bias classification is reproducible from the working set (you can run the same source through the same vendor ratings and get the same answer).
• The substrate is named. The vendors are named. The conflicts are named.
• The page is wrong in places. Where it is, the byline and the date of the assertion are right here, and the corrections are filed in /_index/.