CRED-1: An Open Multi-Signal Domain Credibility Dataset for Automated Pre-Bunking of Online Misinformation

Based on the provided manuscript, here is a detailed technical summary of the paper "CRED-1: An Open Multi-Signal Domain Credibility Dataset for Automated Pre-Bunking of Online Misinformation."

1. Problem Statement

The rapid proliferation of online misinformation and disinformation necessitates automated detection mechanisms. Current solutions often rely on proprietary databases (e.g., NewsGuard) or claim-level fact-checking, which can be opaque, expensive, or reactive (occurring after content spread). There is a critical need for:

• Transparent and Reproducible Data: Open datasets that allow independent verification.
• Privacy-Preserving Deployment: Solutions that can run client-side (on-device) without sending user data to servers.
• Pre-Bunking Capabilities: The ability to flag untrustworthy domains before a user consumes content (at the content delivery stage), rather than just correcting claims after the fact.

2. Methodology

The authors developed CRED-1, a dataset and pipeline designed to generate domain-level credibility scores using only publicly available data and standard software libraries. The methodology consists of two primary phases:

A. Data Acquisition and Merging (Phase 1)

The pipeline aggregates domain labels from two openly licensed sources:

1. OpenSources.co: A curated list of 825 domains categorized by type (fake, bias, conspiracy, satire, etc.).
2. Iffy.news Index: A dataset of 2,040 domains rated by Media Bias/Fact Check (MBFC) for low factual reporting.

• Normalization: Domains are standardized (lowercasing, removing www. and trailing slashes).
• Conflict Resolution: If a domain appears in both sources with conflicting categories, the lower credibility category is assigned (conservative approach).
• Result: A merged list of 2,672 unique domains.

B. Signal Enrichment (Phase 2)

Each domain is enriched with four computed signals to create a multi-dimensional profile:

1. Domain Age: Retrieved via RDAP (Registration Data Access Protocol).
2. Web Popularity: Mapped against the Tranco Top-1M list to determine traffic rank.
3. Fact-Check Frequency: Queried via the Google Fact Check Tools API to count ClaimReview annotations.
4. Threat Intelligence: Checked via the Google Safe Browsing API for malware or social engineering flags.

C. Scoring Model

A composite credibility score ($S$) ranging from 0.0 (fake) to 1.0 (reliable) is calculated using a weighted linear combination of signals:
$$S = w_{cat} \cdot s_{cat} + w_{iffy} \cdot s_{iffy} + w_{fc} \cdot s_{fc} + w_{tranco} \cdot s_{tranco} + w_{age} \cdot s_{age} + w_{fill} \cdot s_{cat}$$

• Weights: Category score ($0.50$), Iffy score ($0.15$), Fact-check ($0.15$), Tranco rank ($0.05$), Domain age ($0.05$).
• Missing Data Handling: If a signal is missing, its weight is reverted to the base category score ($w_{fill}$).
• Hard Override: Any domain flagged by Google Safe Browsing is capped at a score of 0.05, regardless of other signals.

3. Key Contributions

• CRED-1 Dataset: An open, reproducible dataset of 2,672 domains with credibility scores and metadata, released under CC BY 4.0.
• Dual Format Delivery:
  • Compact JSON (145 KB): Optimized for on-device embedding in browser extensions and mobile apps.
  • Full CSV: Contains raw enrichment signals and intermediate scores for research analysis.
• Fully Reproducible Pipeline: The entire dataset generation is implemented in Python using only standard library modules (no external dependencies). It requires only free API keys from Google Cloud.
• Pre-Bunking Focus: Specifically designed for client-side deployment to enable "pre-bunking" (blocking or warning users before they visit a site) while preserving user privacy.

4. Results and Data Characteristics

• Domain Distribution:
  • Mixed: 50.0% (1,335 domains) – The largest group, containing biased or misleading content alongside some facts.
  • Unreliable: 22.0% (589 domains).
  • Fake: 18.4% (493 domains).
  • Conspiracy: 5.7% (153 domains).
  • Satire: 3.5% (94 domains).
  • Reliable: 0.3% (8 domains).
• Score Distribution: The dataset is bimodal.
  • Mean score: 0.299.
  • 31.7% of domains score below 0.2 (highly untrustworthy).
  • 50.0% score between 0.4 and 0.6.
  • Very few domains score above 0.8.
• Signal Coverage Insights:
  • Domain Age: The median age is 14.0 years, indicating that misinformation sites are often long-established, contrary to the heuristic that bad domains are newly registered.
  • Fact-Checking: Only 2.5% of domains have specific fact-check claims in Google's database, highlighting a gap in current fact-checking coverage.
  • Malware Overlap: Only 0.07% (2 domains) were flagged for malware, confirming that misinformation sites generally operate within technical legitimacy bounds (i.e., they are not phishing/malware sites).

5. Significance and Impact

• Privacy-Preserving Architecture: By enabling on-device processing via a compact JSON file, CRED-1 allows browser extensions to filter misinformation without sending user browsing history to a central server.
• Democratization of Tools: Unlike proprietary alternatives, CRED-1 is free, transparent, and allows researchers and developers to audit the scoring logic and extend the dataset.
• Educational and Research Utility: Serves as a "ground truth" for training machine learning models, developing media literacy curricula, and studying the intersection of domain reputation and AI-generated content.
• Strategic Intervention: Supports the "misinformation kill chain" by targeting the content delivery stage, allowing for automated pre-bunking before a user engages with false information.

Limitations Noted by Authors

• Language Bias: The dataset is heavily skewed toward English-language sources.
• Temporal Validity: Domain credibility changes over time; the dataset reflects the state as of February 2026.
• Negative Signal Design: Absence from the dataset does not imply a domain is reliable; it only flags known issues.
• Upstream Bias: The dataset inherits any biases present in the original OpenSources.co and Iffy.news lists.