Contradiction Detection & Divergence Classification

Extract all statement pairs where sources directly disagree, then apply the three-question decision tree: check for factual ground truth, assess downstream synthesis dependency, and determine if the difference is framing rather than fact. Each contradiction receives a RESOLVE-BEFORE-SYNTHESIS, HALT, or DOCUMENT-ONLY classification.

HALT applies when no factual ground truth exists and downstream synthesis depends on resolving the contradiction, blocking processing. DOCUMENT-ONLY applies when the disagreement is a legitimate framing or interpretation difference, allowing both perspectives to appear in the final output.

Run the included validate_output.py script against your output file. It checks for required sections (contradictions_identified, decision_tree_classifications, resolution_status), verifies all three decision-tree question fields, and confirms classification values match the allowed set.

A genuine contradiction involves sources making directly incompatible claims about the same fact, not differences in emphasis or wording. Marking paraphrase variations or complementary points as contradictions is a documented anti-pattern that produces false positives.

It runs at three pipeline stages: P1 during raw input processing, P2 during foundational analysis comparisons, and P4 during final review and integration. It requires a complete thread inventory from P1 and convergence scoring from P4 as inputs.