Claim Divergence Detection
A method for detecting whether factual claims in a text are stable. It extracts claim-type tokens such as numbers, time units, and percentages and determines whether they appear consistently across multiple samples.
Timeline (last 90 days)
By computing the average pairwise similarity of normalized token shingles and separately detecting diverging claim-type tokens (numbers, magnitudes, durations, URLs), a verdict can be assigned to document passages.
The diverging claim check deliberately overrides the similarity score, because a passage that confidently cites wrong details is worse than a passage that cites no details.
All Facts (2)
By computing the average pairwise similarity of normalized token shingles and separately detecting diverging claim-type tokens (numbers, magnitudes, durations, URLs), a verdict can be assigned to document passages.
50%UnverifiedThe diverging claim check deliberately overrides the similarity score, because a passage that confidently cites wrong details is worse than a passage that cites no details.
50%