| name | contradiction-resolution |
| description | When inputs contain conflicting facts, stop and resolve the conflict explicitly — classify it, pick a resolution rule, and log what stays unresolved — before building anything on either side. Activate when two sources give different values for the same quantity, a document's claim disagrees with observed state, or the text itself flags inconsistency. |
| license | MIT |
contradiction-resolution
Trigger (observable): Two sources in the task give different values for the same quantity, a document's claim disagrees with observed/live state, or the text itself flags inconsistency ('but the dashboard says', 'these numbers don't match').
When NOT to activate: Apparent conflicts already resolved by scope (different time windows, different populations) and the scoping is stated; deliberate counterfactuals or A/B arms meant to differ; pure opinion differences with no factual referent.
Procedure
- State the contradiction in one sentence naming both sides and their sources ('Doc says jobs 1–3 done; crontab shows job 3 never installed').
- Classify it: transcription error / version skew / scope difference (different windows, units, populations) / one source stale / genuine conflict.
- Pick the resolution rule and apply it: live observed state beats documents; recomputation beats quotation; fresher primary beats older secondary. Name which rule was used.
- Once one claim from a source is proven wrong, downgrade that source's remaining claims to unverified — do not keep trusting its other markers.
- Anything still unresolved is logged as an explicit caveat attached to every conclusion that depends on it — not dropped because it doesn't block the literal task.
Required output
A contradiction log: one sentence per conflict, its classification, the resolution rule applied (or 'unresolved'), and which downstream claims carry the caveat.
Verification
- Each detected contradiction is stated in one sentence with both sources named and a classification.
- Every conclusion depending on a conflicted fact either cites the applied resolution rule or carries the unresolved caveat — no silent side-taking.
Known risk: Contradiction-hunting in noisy data where trivial mismatches (rounding, timestamps) stall real work. Mitigation: the classification step routes trivia to one-line closure.
Max intended cost: ≤300 added output tokens; up to one live-state check per conflict where a command can settle it.
Evidence status: SMOKE_TESTED — executed end-to-end in live sessions and behaves as specified; no measured lift is claimed.
Lineage: Grounded in two evidence-backed repair patterns — evidence kills specific hypotheses rather than whole spaces, and live observed state beats plan markers — plus a documented reasoning-failure mode: building conclusions on one side of an unresolved contradiction.
