Automated Contradiction Detection

Manual contradiction detection works when you are comparing three sources. When your pipeline processes fifty sources daily, contradictions must be detected automatically. Automated detection operates in three modes. Numeric comparison: when two sources report the same metric for the same entity, compare the values and flag disagreements that exceed a configured threshold. Entity assertion matching: when two sources make claims about the same entity, compare the assertions and flag semantic contradictions. Temporal consistency checking: when a source reports a metric, compare it against the same source's previous report for the same entity and flag improbable changes.

1. Numeric Threshold Detection For quantitative claims, set tolerance thresholds by metric type. Revenue estimates within 10% are normal variation. Revenue estimates differing by 30% are contradictions. Headcount estimates within 15% are normal. Market share estimates differing by more than 5 percentage points are contradictions. Thresholds are metric-specific — do not apply a universal tolerance.
2. Semantic Assertion Matching For qualitative claims, use AI-assisted comparison to identify when two sources make conflicting assertions about the same entity. "Company X is expanding into APAC" versus "Company X is consolidating to North America only" is a semantic contradiction. The detection system flags these for analyst review rather than attempting automated resolution.
3. Temporal Consistency Compare each new data point against historical data for the same entity. A company that grew revenue 15% annually for three years and then allegedly grew 80% in one quarter should be flagged. Not necessarily wrong — but improbable enough to warrant verification.