The Three-Stage Content Grounding Process — How We Verify Every Claim Against Source Documents

Estimated Read Time: 14 minutes | Category: Technology Deep-Dive

Why Source Verification is Hard

It's not enough to find a source that mentions the same topic as a claim. Consider this claim: "Apple's revenue was $383 billion in 2023."

A naive verification system might find a Forbes article titled "Apple's Revenue Growth" and mark the claim as verified. But what if the article says Apple's revenue was $394 billion? Or what if it's discussing a different fiscal year?

True verification requires semantic alignment—proving that the source actually supports the specific claim being made. This is what our Content Grounding Analyzer (CGA) does, and it's the core of our differentiation.

Pipeline Steps 6-7: Citation Resolution & Content Grounding

After claims are extracted by the CEE (Step 5), they enter our verification pipeline:

Step 6: Citation Resolution Service (CRS)

The CRS is responsible for finding and validating source documents for each claim:

• URL validation: Confirms cited URLs actually exist and are accessible

• Content retrieval: Fetches the full text of each source document

• Archival lookup: For paywalled or deleted content, checks archive.org and other archives

• Source metadata: Extracts publication date, author, and domain authority signals

Step 7: Content Grounding Analyzer (CGA)

This is where the magic happens. The CGA performs three sequential checks on each claim-source pair:

Stage 1 — Relevance Check:

• Question: Is this source actually about the same topic as the claim?

• Method: Compute semantic similarity between claim and source content

• Threshold: Cosine similarity ≥ 0.70 to proceed

• Failure mode: Source discusses related but different topic

Stage 2 — Support Check:

• Question: Does the source content actually back up this claim?

• Method: LLM-based entailment analysis with chain-of-thought reasoning

• Threshold: Support score ≥ 3 out of 5

• Failure mode: Source mentions topic but doesn't support specific claim

Stage 3 — Fidelity Check:

• Question: Does the claim accurately represent what the source says?

• Method: Cross-check for exaggeration, misattribution, or false precision

• Threshold: Fidelity score ≥ 0.85

• Failure mode: Claim overstates, rounds incorrectly, or takes out of context

The Three Failure Modes

Each CGA stage catches different types of hallucinations:

1. Topic Drift (Relevance failure): Claim: "Apple's iPhone revenue was $200B." Source: An article about Apple's Mac revenue. Fails Stage 1.

2. Unsupported Claims (Support failure): Claim: "Google laid off 12,000 employees." Source: Article mentions Google layoffs but doesn't specify numbers. Fails Stage 2.

3. Misrepresentation (Fidelity failure): Claim: "Revenue grew 50%." Source: Says revenue grew "nearly 50%" (actual: 47.3%). Fails Stage 3.

Up Next

In the next post, we'll explore the Computational Verification Module (CVM) and the Internal Consistency Checker (ICC)—the systems that verify mathematical claims and detect contradictions.

PromptReports.ai is a Verified Intelligence Platform that delivers AI-powered analyst reports with claim-level source verification. Generate your first verified report →