Contradiction detection: keeping both numbers instead of averaging them
2026-08-08 · 6 min read
Breaking news is the worst case for a summarizer. Counts move, sources lag, and a model that blends them produces a number that no source ever reported.
That blended number is the most dangerous output in AI: it looks sourced and it is invented.
The rule
Competing values are never averaged, rounded together, silently overwritten by the newest one, or quietly picked between. Each is kept verbatim, attributed, and shown next to the reason the figures differ — cut-off time, counting method, scope, definition, sample.
Casualty counts, vote tallies, financial figures and polling numbers are treated as hard contradictions on any disagreement, not noise to be smoothed.
Question
How many people died in the incident reported this morning?
Answer
Two counts are live and they don't reconcile yet. Both are shown as reported; neither is averaged.
Straight Edge score · Partial support
Multiple sources of similar rank disagree; capped while unresolved.
Contradiction logged · unresolved
Reported death toll
Wire agency A
6 dead
police briefing, 09:40 local
Wire agency B
9 dead
hospital count, 14:10 local
Different cut-off times and different counting bodies: a scene count versus a hospital intake count.
Would change this answer: An official consolidated figure from the responsible authority.
Why the log matters after the fact
The contradiction entry stays in the exported trace. Months later you can see not just what was believed, but what was contested and why — which is the part every other tool throws away.
Ask it something it could get wrong
Three questions, no card, no account. Every answer comes back with its score, its sources and an exportable trace.
Ask 3 questions free