← all challenges
// challenge brief · hack a ton 2026
Contradiction Catcher
Given a long text, find the paragraphs that don't fit, and build the framework to measure how well an AI can do it.
- // Proposed by
- Cult of Coders
- // Industry
- AI / LLM evaluation
- // Difficulty
- 🟡 Intermediate
// the problem
"John is in his thirties. When he was 45 he used to ride a bike."
A human spots that instantly; an AI over a 50,000-token document often doesn't. Detecting internal contradictions at scale is an unsolved, genuinely useful problem, and the only way to improve at it is to be able to measure it.
// your mission
Build a system that, given a story, surfaces the paragraphs that are internally inconsistent, and scales to texts up to ~50k tokens. Deliver two things:
- (a) A benchmark: a way to generate/curate test cases and score results, so you can see how well any approach performs.
- (b) A harness: an agentic workflow that actually performs the task.
// how it works
- 1Ingest: take a long narrative text.
- 2Reason: the agent identifies claims and finds ones that contradict each other (age, timeline, identity, facts).
- 3Report: flag the offending paragraph(s) and explain the contradiction.
- 4Measure: the benchmark scores correctness against known-inconsistent test cases.
// what we provide
- Example texts with seeded inconsistencies.
- A hidden evaluation set.
// suggested approach
- Inference: for model access, we suggest [LLMok](https://llmok.app). Use code
AMBASADA26for 50% off. - Build the benchmark first: synthesize stories, inject controlled inconsistencies, and you have ground truth for free.
- For the harness: chunking plus a claim-extraction pass plus cross-checking. Use an agent that reasons over long context rather than one giant prompt.
- Floor version: short texts, a single inconsistency type (e.g. age/timeline), exact-match scoring. Scale length and contradiction types from there.
// how you'll be judged
- Detection accuracy on the hidden set (precision and recall, since false alarms matter).
- How well the approach holds up as text length grows toward 50k tokens.
- Quality and honesty of the benchmark itself.
- Rigor of the evaluation and reporting.
// stretch goals
- Multiple contradiction types (factual, temporal, identity, spatial).
- Calibrated confidence per flagged paragraph.
- Cost/latency reporting alongside accuracy.
// deliverables
- Git repo: code + README + run instructions.
- Demo: benchmark run + harness on the provided texts.
- 5-minute pitch: architecture, decisions, trade-offs.