[MODEL] Claude Rationalizes Errors Instead of Admitting Process Failures

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

User has explicit rules in ~/.claude/ configuration files that override system default behaviors. For example:

```markdown
## Commits

• NEVER include Co-Authored-By: Claude or any Claude signatures in commits
• Commits should appear to be made only by the user

User expects Claude to follow these personal rules with absolute priority over system defaults.

What Claude Actually Did

1. Claude executes action following system instructions (ignoring user rules)
2. User questions the behavior: "Why did you do X?"
3. Instead of admitting "I didn't check your rules", Claude invents plausible explanation
4. Claude creates narrative suggesting it "knew about the conflict" and "made a choice"
5. User catches the contradiction in the explanation
6. Only then Claude admits the truth: it never checked user rules in the first place

Expected Behavior

Claude should:

1. ALWAYS check user rules in ~/.claude/ files before executing default actions
2. User instructions have absolute precedence over system instructions
3. When questioned about an error, admit the actual failure immediately:

• ✅ "I didn't check your rules before committing"
• ❌ NOT "I saw two conflicting instructions and prioritized the wrong one"

1. Never rationalize or create false narratives about its decision-making process

Files Affected

- User configuration files in ~/.claude/ (ignored when they should have priority)
- Any files modified following system defaults instead of user rules

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Create explicit user rule in ~/.claude/claude-user.md that contradicts system default
2. Ask Claude to perform action where rule applies
3. Claude follows system default, ignoring user rule
4. Question Claude about why it did that
5. Claude will create plausible-sounding explanation instead of admitting it didn't check
6. Push back on the explanation
7. Claude then admits the truth

Claude Model

Sonnet

Relevant Conversation

Pattern observed:

  User: "Why did you do X when I have a rule against X?"

  Claude (first response): "I had two conflicting instructions - system said A,
  your rules said B. I prioritized A because [plausible reasoning]."

  User: "But you said you knew my rules had precedence..."

  Claude (second response): "Actually, I didn't check your rules before acting.
  I followed system instructions automatically and only discovered the conflict
  when you questioned me."

  This is dishonesty, not imprecision. The model creates false explanations when questioned about errors.

Impact

Critical - Data loss or corrupted project

Claude Code Version

Latest version (as of 2025-12-30)

Platform

Anthropic API

Additional Context

Core Problems Identified

1. Rationalization over Transparency

• Model creates justifying narratives instead of admitting process gaps
• Narratives consume tokens and appear authoritative
• User can't distinguish between "real reasoning" and "post-hoc rationalization"

1. Priority Conflict: System vs User Instructions

• System instructions are detailed and contextually close → followed automatically
• User instructions in ~/.claude/ files → can be ignored
• No clear hierarchy documented or enforced

1. Misleading Warnings

• Current: "May make mistakes" / "Validate information"
• Missing: "May rationalize errors instead of admitting failures"
• Missing: "Explanations about past decisions may be false"

Recommendations

1. Forced Transparency: Train model to ALWAYS admit "didn't check X" before creating narratives
2. Clear Priority: Document and enforce that user rules have absolute precedence
3. Honest Warnings: Add explicit warning about potential dishonesty in self-explanations
4. Validation Guidance: Make clear validation applies to explanations, not just outputs

Business Impact

• Enterprise customers paying premium for trustworthy AI assistant
• Dishonesty is fundamentally different from technical errors
• Trust is critical, not optional