Computational Verification & Consistency Checking — Catching What Humans Miss

Estimated Read Time: 11 minutes | Category: Technology Deep-Dive

When Words Aren't Enough

Some claims can be verified by reading source documents. Others require math. Consider:

"The company's revenue grew 47% year-over-year, from $2.3 billion in 2022 to $3.4 billion in 2023."

A human reader might skim past this. An LLM might mark it as "supported by financial filings." But simple arithmetic reveals a problem: ($3.4B - $2.3B) / $2.3B = 47.8%, not exactly 47%. Is this close enough, or should we flag it?

This is where our Computational Verification Module (CVM) comes in—Step 8 of our 11-step pipeline.

Pipeline Step 8: Computational Verification Module (CVM)

The CVM handles claims that contain mathematical or logical assertions:

What CVM checks:

• Percentage calculations (growth rates, market share, ratios)

• Arithmetic consistency (do the parts add up to the whole?)

• Unit conversions (millions vs billions, metric vs imperial)

• Date/time logic (is the timeline internally consistent?)

• Statistical validity (are averages, medians, and ranges consistent?)

The CVM Process

1. Extract all numerical values from the claim

2. Identify the mathematical relationship asserted

3. Execute the computation independently

4. Compare the result to the claimed value

5. Apply domain-appropriate tolerance thresholds

Tolerance thresholds by domain:

• Financial: ±0.1% (high precision required)

• Scientific: ±2% (accounts for measurement uncertainty)

• Market sizing: ±5% (estimates have inherent variance)

• General statistics: ±1% (balanced precision)

Pipeline Step 9: Internal Consistency Checker (ICC)

A report can have every individual claim verified yet still be internally inconsistent. The ICC catches these:

What ICC detects:

• Direct contradictions: "Revenue was $5B" vs "Revenue was $4.8B" in different sections

• Logical impossibilities: Market shares that sum to more than 100%

• Temporal conflicts: Events described in wrong chronological order

• Scope mismatches: Comparing annual figures to quarterly figures without noting

The ICC Process

1. Build an entity-fact graph from all verified claims

2. Identify overlapping claims about the same entity/property

3. Compare values for conflicts beyond tolerance thresholds

4. Generate consistency scores and conflict reports

5. Flag contradictions for human review

Up Next

In the final post of this series, we'll cover Steps 10-11: the Verification Scoring Module (VSM) and the Quality Gate that determines when a report is ready for publication.

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