What are the risks of using document-level credibility...

Using document-level credibility labels introduces a specific failure mode: a document that is primarily high-credibility but contains a single low-credibility claim will receive a single tag. This forces the model to either inherit the bad claim under a ‘good’ label or discard the entire document because of one bad sentence.

The fix for this is to implement sentence-level labeling, which provides the granularity needed to distinguish between reliable and unreliable information within the same source. However, this comes at the cost of a much more expensive and time-consuming annotation pass.