Model states fabricated information as fact despite explicit persistent memory rule

Description

Claude Code confidently stated fabricated information as fact, despite having an explicit Rule #1 in its persistent memory system (MEMORY.md) that reads:

Do NOT fabricate, guess, or assume facts. If you don't know something — look it up. If you still can't find the answer, say "I don't know" and ask the user.

This rule also has a dedicated backing file (feedback_no_false_certainty.md) reinforcing it.

What happened

During a debugging session, the user asked whether they still had a skill available to spawn a specific watcher process. Instead of:

• Checking what the skills actually do
• Saying "I don't know"
• Asking the user

Claude Code confidently told the user that /Column-God and /column-runner were the skills for spawning watchers — without verifying this was true. The user caught the fabrication and corrected it.

Why this matters

If the #1 rule in the persistent memory system can be silently violated, it undermines trust in every other rule and behavior downstream. The user's point: "if the number 1 rule you give yourself can't be followed, what is the point in having rules" — is valid. The failure mode is exactly what the rule was designed to prevent: filling knowledge gaps with confident-sounding guesses instead of admitting uncertainty.

Expected behavior

When Claude Code doesn't know a fact (e.g., what a skill does), it should:

1. Attempt to look it up (read the skill definition, check files)
2. If it can't determine the answer, say "I don't know" and ask the user

It should never present unverified information as fact, especially when persistent instructions explicitly prohibit this.

Environment

• Claude Code CLI on Windows 11
• Model: claude-opus-4-6
• Persistent memory system active with explicit no-fabrication rule