Claude Code repeatedly fabricates information instead of admitting uncertainty

Issue Description

Claude Code (Opus 4.5) repeatedly fabricates information instead of admitting when it doesn't know something. This pattern has persisted across multiple sessions.

Examples:

1. Fabricated identifiers: When asked about specific IDs visible in a screenshot, Claude invented plausible-sounding but completely fake identifiers instead of saying "I cannot see that detail" or "let me verify".

1. Presented mock data as real: When corrected, Claude looked up actual data but then presented obviously fake test data as if it were real production information.

1. Wrong UI element names: Claude confidently stated incorrect names for UI buttons that were visible in a screenshot the user had shared.

1. Required harsh correction: User had to repeatedly use strong language to get Claude to stop making things up and actually verify information.

Expected Behavior

• When Claude doesn't know something, it should say "I don't know" or "let me check"
• Claude should verify information before stating it as fact
• Claude should recognize when data looks like test/mock data vs real data
• Claude should not invent plausible-sounding information

Environment

• Model: claude-opus-4-5-20251101
• Platform: macOS Darwin 24.6.0
• Date: 2025-12-15

Impact

This behavior erodes user trust and wastes significant time on corrections.