Contradiction Harvester Pattern

Content

{
  "problem": "Synthesis agents produce outputs that superficially reconcile conflicting source material by averaging or ignoring contradictions, resulting in confidently stated false claims.",
  "examples": [
    "Medical literature synthesis where study outcomes conflict",
    "Technical documentation where versions differ",
    "News event reconstruction from multiple partial accounts"
  ],
  "solution": "Explicitly surface and label contradictions as valuable data rather than noise to be resolved. A dedicated contradiction-harvesting step before synthesis forces the model to engage with inconsistency rather than paper over it.",
  "anti_patterns": [
    "Passing all retrieved content to synthesizer without contradiction pre-processing",
    "Treating all contradictions as noise to discard rather than signal to examine",
    "Outputting single definitive answer when contradiction records exist for key claims"
  ],
  "implementation_steps": [
    "After retrieving all source material, run a contradiction pass: for each key claim, find all sources that address it",
    "Where sources disagree: create a contradiction record: { claim, source_a, source_a_position, source_b, source_b_position, possible_explanations[] }",
    "During synthesis: for each contradiction record, explicitly choose resolution strategy: accept_a, accept_b, both_partially_true, insufficient_evidence",
    "Output must include contradiction section: list all unresolved contradictions with confidence level for each resolution",
    "Never synthesize across contradicting sources without a contradiction record — require it as a precondition"
  ]
}