Contradiction Surface

What it is

A contradiction surface is two or more published places asserting incompatible versions of the same fact. A pricing page says one number, a sales deck says another, and a help article quotes a third from two years ago. Each was correct when it was written, and none of them is obviously the wrong one to a reader arriving cold.

Why it matters

A contradiction is worse than a single stale page. Knowledge Company prioritises claims that conflict with the current source of truth, because those are what get quoted back. A retrieval system answers from whichever version it indexed, so the same question can produce different answers on different days.

A plain retrieval pipeline raises no error when that happens, because it assumes the documents it retrieved agree with each other. That is a known weak point rather than a fixed limit: prompting a model to work out the conflict type before it answers has been built and measured, and both a two-step pipeline and a taxonomy-aware prompt improve how far responses match the expected behaviour, by 9 and 5.5 points on average. The same authors still treat integrating those methods into practical RAG systems as future work, and later work finds that prompt-based balancing gives only modest gains. Left to handle conflicting sources unaided, current models struggle: on that benchmark the best model identified the conflict type in 65.3% of cases, and even under explicit instruction models frequently commit to a single stance despite contradictory evidence. It also breaks the usual fix: correcting one surface leaves the contradiction intact, and often makes it harder to spot, because the loudest wrong version has now been quietly replaced by a quieter one.