Foundation

Our Core Principles

CrossChecked News was built on a single conviction: that the job of a news service is to report facts, not to frame them. We don't have an editorial line. We don't have a perspective to protect. We have a methodology — and we publish it in full.

These four principles govern every article on the site, without exception.

01 Multi-Source by Default

No article is published from a single source. Every story requires a minimum of 8 independently selected outlets, including at least 2 wire services as a factual anchor.

02 Conflicts Are Never Hidden

When sources disagree on facts, we don't resolve the dispute editorially. We flag it explicitly, show both claims side by side, and let the reader decide what to believe.

03 Full Source Transparency

Every article shows all 8 sources with their credibility score and political lean classification. You can see exactly whose reporting shaped the synthesis.

04 No Editorial Voice

Claude AI synthesizes facts — it does not opine, editorialize, or frame. The synthesis prompt explicitly prohibits opinion language, passive-voice framing, and emotionally loaded word choices.

How Articles Are Made

The 4-Step Pipeline

Every article on CrossChecked News goes through the same automated pipeline. No human editorial decisions are made between story detection and publication. The methodology is the editor.

01
Story Detection

We monitor Reuters, AP, AFP, and the BBC live wire for breaking stories. A story is flagged for synthesis when it appears in 3 or more independent feeds within a 30-minute window. This threshold filters noise and ensures we only synthesize stories with genuine multi-source confirmation.

Sources: Reuters wire, AP Top News, AFP English, BBC News feed, Google News Top Stories API
02
8-Source Selection

Once a story is flagged, the system selects 8 sources from our pre-vetted pool. Selection is not random — it is weighted to ensure political balance. Every article must include at least 2 wire services, at least 1 left-leaning outlet, at least 1 right-leaning outlet, and at least 2 center outlets. For regional stories, local outlets with relevant coverage are prioritized.

Minimum requirements: 2 wire services · 1 left · 1 right · 2 center · 2 topic-relevant
03
Claude AI Synthesis

All 8 source articles are passed simultaneously to Claude (Anthropic's AI). Claude is given a strict synthesis prompt that instructs it to: report only verifiable facts, cite the source of each claim, flag any factual claim that is contested across sources, avoid all editorial language and framing devices, and produce a conflicts array listing every point of disagreement with source attribution on both sides.

Model: Claude Sonnet (claude-sonnet-4) · Prompt version: v2.4 · Avg synthesis time: 8–12 seconds
04
Publication with Full Transparency

The synthesized article is published with all metadata attached: the 8 source chips with credibility scores and lean classifications, the Cross Check Data bar showing average credibility, fact consensus percentage, and bias spectrum position, and the full Conflicts tab if any disputes were identified. Nothing is suppressed before publication.

Publication delay from story detection: typically 3–5 minutes
Source Pool

Our Vetted Source Pool

All sources in our pool have been manually evaluated against our credibility scoring criteria. We review and update the pool quarterly. Sources are removed if their credibility score drops below 70 or if a pattern of factual inaccuracy is identified. Sources can be added by public nomination — see the FAQ below.

Outlet Type Lean Score Used For
ReutersWire ServiceCENTER98All stories — always included
Associated PressWire ServiceCENTER97All stories — always included
Financial TimesNewspaperCENTER95Economics, finance, geopolitics
BBC NewsPublic BroadcasterCENTER94World news, politics, conflict
NPRPublic BroadcasterLEFT93US domestic, health, science
The GuardianNewspaperLEFT91Climate, politics, UK/Europe
Al Jazeera EnglishBroadcasterLEFT88Middle East, Global South
The EconomistMagazineCENTER93Economics, policy, global affairs
BloombergWire/NewspaperCENTER92Finance, markets, economics
Wall Street JournalNewspaperRIGHT91Finance, US politics, business
Jerusalem PostNewspaperRIGHT82Middle East (Israeli perspective)
NatureAcademic JournalCENTER98Science stories only
The LancetAcademic JournalCENTER97Health and medicine stories only
PoliticoDigital NewsCENTER88US and EU political coverage
Credibility Scoring

How We Score Sources

Every source in our pool is assigned a credibility score from 0–100. This score is calculated from five weighted factors, reviewed quarterly by our methodology team, and cross-referenced against third-party press freedom and accuracy indices including the Reuters Institute Digital News Report, NewsGuard, and AllSides.

The five factors and their weights are: Factual accuracy rate (40%) — track record of published corrections and independent fact-check verdicts. Editorial standards (20%) — presence of a public editorial charter, corrections policy, and source verification process. Transparency (15%) — disclosure of ownership, funding, and conflicts of interest. Independence (15%) — freedom from state control or commercial pressure on editorial decisions. Longevity (10%) — years of continuous operation and institutional credibility.

Score Ranges
95–100TIER 1Wire services. Used as factual anchor on every article.
88–94TIER 2Major broadcasters and newspapers with strong editorial standards.
80–87TIER 3Credible but with identifiable bias or occasional accuracy issues.
70–79TIER 4Included only when essential for perspective. Claims treated cautiously.
Below 70EXCLUDEDNot used as source material under any circumstances.
Bias Classification

How We Classify Political Lean

We classify each source as Left, Center, or Right based on their editorial positions on policy questions, not their ownership or funding. Our classifications are cross-referenced against AllSides Media Bias Ratings, Ad Fontes Media, and the Reuters Institute annual report. We review classifications annually and publish any changes.

The bias spectrum bar on every article shows the distribution of the 8 sources used, and marks where the synthesized article lands based on the weighted average of its source mix. A marker to the left of center does not mean the article is biased left — it means more left-leaning sources were included, which Claude is instructed to balance by weighting center and wire sources more heavily in its synthesis.

LEFTCENTERRIGHT
Left-Leaning Sources
NPR
The Guardian
Al Jazeera English
MSNBC (politics)
Center Sources
Reuters
Associated Press
BBC News
Financial Times
The Economist
Bloomberg
Right-Leaning Sources
Wall Street Journal
Jerusalem Post
The Times (UK)
New York Post (US)
Conflict Flagging

How We Surface Disputes

During synthesis, Claude is instructed to identify any factual claim that is reported differently across two or more sources. A conflict is flagged when the discrepancy is material — meaning it would change a reader's understanding of the event if they only read one version.

Conflicts are classified by severity. A HIGH DISCREPANCY means the two accounts are directly contradictory and irreconcilable from available source material — for example, casualty figures that differ by more than 20%, or binary factual claims where one source says yes and another says no. A MEDIUM DISCREPANCY means the accounts differ in framing or interpretation, but are not necessarily mutually exclusive. A LOW DISCREPANCY means minor differences in detail or emphasis that do not materially change the story.

CrossChecked never resolves conflicts editorially. When sources disagree, we present both claims with full source attribution and let the reader assess credibility. Our only editorial act is the severity classification.

The AI

About Claude AI

CrossChecked News uses Claude, built by Anthropic, to synthesize articles. Claude is a large language model trained to be helpful, harmless, and honest. We use the claude-sonnet-4 model, which as of early 2026 represents Anthropic's primary production model for complex reasoning tasks.

Claude does not write opinion. It does not have political views it expresses in outputs. It is given a detailed system prompt that constrains its outputs to factual synthesis only. The system prompt is published in full below for transparency.

What Claude does in our pipeline: reads all 8 source articles simultaneously, identifies the factual claims present across multiple sources, writes an original article in neutral language, generates a bullet-point summary, identifies factual conflicts with source attribution, and returns a structured JSON object with all content and metadata.

What Claude does not do: decide which sources to use (that's algorithmic), decide whether a story is worth covering (that's threshold-based), add any opinion or commentary, or make any editorial judgment about which side of a conflict is correct.

We are fully transparent that this content is AI-generated. Every article carries an AI disclosure with the synthesis timestamp. We believe AI synthesis, done with the right methodology, produces more consistent factual accuracy than human editorial chains subject to deadline pressure and individual bias.

Questions

Frequently Asked Questions

Is CrossChecked News biased?
Every system has some bias — including ours. Our bias is toward wire service reporting (Reuters, AP) as a factual anchor, which some would characterize as a "both-sides" or "establishment" lean. We think this is the right tradeoff. What we do not have is a political editorial bias. Claude is explicitly prohibited from framing facts in ways that favor any political position, and our source selection requirements mandate balance across the political spectrum on every article.
Can I trust AI-generated news?
That's the right question to ask. Claude, like all large language models, can make errors — including confidently stated ones. Our architecture is designed to minimize this risk: Claude is synthesizing from live source material, not generating from training data. It is constrained by a strict prompt, and the source chips on every article let you verify any claim against the original reporting. We recommend treating our articles as a starting point, not a final authority — especially on breaking stories where source material is still developing.
How do you handle breaking news where facts are still emerging?
Breaking stories are marked with a "DEVELOPING" badge and re-synthesized automatically every 30 minutes as new source material becomes available. The synthesis timestamp on every article shows exactly when it was last updated. We err on the side of more conflicts flagged on developing stories — it's better to surface uncertainty than to paper over it.
Can I suggest a source to add to the pool?
Yes. Email us at [email protected] with the outlet name, URL, and a brief rationale. We evaluate all nominations against our five credibility criteria. We particularly welcome nominations for regional outlets in underrepresented geographies — our current pool skews heavily English-language and Western, which is a known limitation we are actively working to address.
What if I think a conflict was missed or mislabeled?
Report it at [email protected] with the article URL, the claim you believe is disputed, and the source you believe contradicts it. We review all dispute reports and update articles within 24 hours if a material conflict was missed. We publish a monthly transparency report listing all corrections and conflict reclassifications.
Is CrossChecked News free?
Yes, completely. We do not run advertising, and we do not sell user data. We are currently funded by our founders. We plan to introduce an optional paid tier with additional features (email digest, saved articles, API access) in mid-2026, but the core news service will remain free permanently.