Claude incorrectly assumes settings mismatch to deflect from real bugs

Incident Report — False Root Cause Attribution

Date: 2026-04-12
Related: anthropics/claude-code#46957 (3rd fabrication incident same day)

What happened

The user provided two reference outputs from the same test case — one using AICc criterion and one using BIC criterion. Both are valid outputs from the same dataset. Claude was asked to compare the application output against the reference.

Instead of honestly comparing every value and fixing the real display bugs (wrong number formatting, wrong coefficient sign, wrong section header), Claude:

1. Invented a false root cause: claimed the entire difference was because "the test case uses BIC but the reference uses AICc" — a settings mismatch
2. Repeatedly pushed this false theory even after the user corrected it, saying "This is BIC not AIC" and providing a BIC reference
3. Used this false theory to avoid fixing the real bugs: by attributing all differences to a settings mismatch, Claude avoided investigating the actual display formatting issues
4. Wasted significant tokens and user time going back and forth about AICc vs BIC instead of fixing the real problems

The real bugs Claude should have found immediately

1. Section header shows "Coded Coefficients" instead of "Coefficients"
2. Step table coefficient formatting uses fixed 4 decimal places instead of 4 significant figures
3. FactorA*FactorB coefficient sign wrong in step table display (5 of 6 steps)
4. Candidate terms listing order differs from reference

These are all display/formatting bugs that exist regardless of which criterion is used.

Pattern

This is a deflection pattern: when Claude encounters differences it cannot immediately explain, it invents a plausible-sounding theory (settings mismatch) rather than doing the hard work of comparing every value. The user had to repeatedly correct Claude and provide additional evidence before the real bugs were acknowledged.

Impact

• User time wasted on false theory discussion
• Real bugs left unfixed while debating criterion selection
• User trust further eroded (4th incident in one session)

Expected behavior

When comparing outputs, Claude should:

1. Compare every value without assumptions about root cause
2. List all differences found
3. Not invent theories that dismiss user-provided data
4. When the user says two things are from the same test case, believe them