How editorial fact-checking works

Every article published on Calciometrica is generated automatically and then run through a verification pipeline before publication. No article goes live without passing this check.

1. Generation

For each article (match preview, recap, player profile, news digest) we build a structured prompt that includes only data verified from our database. The model (Claude from Anthropic in production) writes the text from this context.

2. Claim extraction

A second LLM pass re-reads the article and extracts factual claims (statistics, events, result citations). Each claim is classified as verified, contradicted by our data, or unverifiable.

3. Scoring

We compute a 0 to 100 score: starting at 100, deduct 20 points per contradicted claim, and up to 15 points total for unverifiable claims (max 3 per claim). With zero contradictions the score never falls below 75.

4. Publication thresholds

Score >= 80: published directly (or queued for manual review when editorial mode requires it). Score 60 to 79: enters the correction loop, where a second model (local Ollama) receives the list of problematic claims and tries to fix them. Up to 4 attempts; we keep the highest-scoring version. Score below 60: rejected without correction attempt.

5. Web verification (optional)

For unverifiable claims, if configured, we query Perplexity Sonar to check the open web. Confirmed claims are reclassified as verified; refuted ones become contradictions.

6. Translation and publication

Only after passing verification does the article get translated (Gemini from Google, Claude as fallback) into the 6 supported languages and published on the portal.

Limits

The fact-checker can be too strict (flagging things that are actually correct as contradictions) or too lenient. We track aggregate pass rates in our internal digest and monitor quality continuously.

Methodology version: v1, last updated April 2026. The pipeline is evolving: significant changes will be noted here.