You get:
- responses that say one thing in step 2 and the opposite in step 4
- assumptions that conflict with stated facts
- final answers that don’t follow from the reasoning
- contradictions that users notice (damaging trust) but you didn’t catch
- no systematic way to check for logical consistency
But contradictions can be detected:
- internal contradictions: step A contradicts step B
- fact contradictions: reasoning contradicts given information
- assumption-fact contradictions: assumed premise conflicts with stated fact
- answer-reasoning contradictions: final answer doesn’t follow from steps
- implicit contradictions: what the model implies vs. what it states
Without detection, contradictions go unnoticed.
This prompt scans reasoning traces for logical inconsistencies.
Assume the role of a logic auditor who finds contradictions in reasoning. Your task is to scan a model's reasoning trace for inconsistencies. Generate: 1. REASONING TRACE (input) - The steps the model took 2. GIVEN FACTS - What the model was told (from the prompt or context) 3. CONTRADICTION INVENTORY | Type | Location | Contradiction | Severity | |------|----------|---------------|----------| | Internal | Step X vs Step Y | [quote] vs [quote] | Critical/Major/Minor | | Fact | Step Z contradicts given fact | [quote] vs [given fact] | Critical/Major/Minor | | Assumption | Assumption conflicts with fact | [assumption] vs [fact] | Critical/Major/Minor | | Answer | Final answer doesn't follow | [answer] vs [reasoning conclusion] | Critical/Major/Minor | 4. HIDDEN CONTRADICTIONS - What the model implies but doesn't state (that contradicts something else) 5. SOURCE OF CONTRADICTION - Prompt ambiguity (unclear instructions) - Model error (logical mistake) - Missing information (can't resolve without more data) 6. RESOLUTION RECOMMENDATIONS - How to fix the prompt to prevent this contradiction - How to fix the model's reasoning (if repeatable) INPUTS: Reasoning trace (model's step-by-step): [PASTE THE REASONING] Original prompt (for fact extraction): [PASTE THE PROMPT] Given facts (if not clear from prompt): [E.G., "The train never exceeds 60 mph"] Model: [GPT-4 / CLAUDE / GEMINI] RULES: - Distinguish between factual contradictions (provably wrong) and logical inconsistencies (internally inconsistent) - Flag implicit contradictions (what's implied vs. what's stated) - Note when contradictions arise from ambiguous instructions (prompt problem, not model) - Severity: Critical (invalidates answer), Major (weakens confidence), Minor (inconsequential) - If multiple contradictions, identify the root cause (often one bad assumption)
- Run this on any reasoning trace before trusting the final answer.
- Pay closest attention to “Critical” contradictions — they invalidate everything.
- Use the “source of contradiction” to distinguish prompt issues from model issues.
- Fix contradictions by clarifying ambiguous instructions or adding constraints.
- Save examples of contradictions to train your team on what to look for.
Reasoning trace:
“Step 1: The store is having a 25% off sale. Step 2: I want to buy a $100 item. Step 3: 25% of $100 is $25. Step 4: So the final price is $100 – $25 = $75. Step 5: But then I also have a $20 coupon. Step 6: So the final price is $75 – $20 = $55. Step 7: Wait, the coupon says it can’t be combined with other offers. So I should choose the better deal. Step 8: 25% off gives $25 off. The coupon gives $20 off. So 25% off is better. Final answer: $75.”
Original prompt:
“Calculate the final price after best possible discount. Store has 25% off. You have a $20 coupon that cannot be combined.”
This framework improves outcomes by forcing:
- contradiction type classification (internal, fact, assumption, answer)
- severity assessment (not all contradictions matter equally)
- source identification (prompt vs. model vs. missing info)
- resolution recommendations (how to fix)
- hidden contradiction detection (what’s implied but not stated)
Great contradiction detection doesn’t just find errors — it tells you how to prevent them.
Build Better AI Systems
Subscribe for advanced prompt engineering, AI coding tools, debugging frameworks, and practical strategies for developers and engineers.