Fact Checking / 2026/4/4

ZFE V1.2: Moving Beyond the Primary Source Fallacy with Contextual LLM Validation

ZFE V1.2: Context over Domain

For the longest time, algorithmic media evaluation has suffered from the “Primary Source Fallacy”: the assumption that if an article links to a .go.jp or .gov domain, it is inherently factual and well-researched. This led to cheap SEO tactics where writers would spam primary source links at the bottom of an article, completely disconnected from the actual narrative.

Today, ZashStudio and NT Media are officially stepping away from that model with the deployment of ZFE V1.2.

A New Scoring Philosophy

We realized that an article with three raw government data links and zero media synthesis is often less helpful—and sometimes more misleading—than an article that cross-references a government report with an investigative journalism piece.

In V1.2, we’ve adjusted the PHP-side scoring logic:

  • Capped Primary Link Dominance: You can no longer reach a perfect score simply by stuffing official sources.
  • Boosted Diversity Value: The algorithm now highly rewards articles that successfully map Primary Sources against Secondary Sources (media, blogs, and local news).

But the true breakthrough is our new AI validation pipeline.

Automated LLM Citation Checking

To truly enforce editorial integrity, we’ve integrated Gemini 2.5 Flash directly into our automated deployment workflow via a custom Python script (zfe_llm_validator.py).

  1. Source Extraction: The instant an article passes the initial deployment gate, ZFE maps all specific primary and secondary links.
  2. Context Fetching: Our engine visits those external web pages, strips out the noise, and extracts the raw textual context.
  3. LLM Cross-Examination: We hand both the article’s narrative and the extracted source contexts to Gemini with a clear directive: “Does the article accurately represent these sources, or does it cherry-pick data to form an unwarranted conclusion?”
  4. Live Patching: The resulting score (0 to 100) and rationale are patched back into the WordPress REST API in seconds, seamlessly appearing on the live article via our shortcodes.

The Power of Granular Verification

In our very first live test on an article discussing “Subscription Fatigue,” the LLM handed down a harsh but incredibly fair score of 65/100. Why? Because while the article linked to valid Consumer Affairs Agency pages, the sources themselves were too broad (top-level whitepapers) to definitively back up the article’s highly specific claims about “Comfort Binging” and “Dark Patterns.”

This is exactly what we wanted: an engine smart enough to say, “Your links are real, but they don’t prove your point.”

The automation is now live on our deployment pipeline. Next time you read deep analysis on NT Media, you can hover over the ZFE Score panel to see exactly how rigorous the cross-examination was. Welcome to the era of semantic truth.