Claude Code makes confident assertions without verifying data — persistent pattern despite memory/instructions

Bug Description

Claude Code repeatedly makes confident factual assertions without checking available data first, even after being explicitly instructed not to via CLAUDE.md global instructions and project memory.

Examples from a single session

1. Stated "Python 3.14 upgrade made no meaningful difference for training speed" — when forced to check timestamps, data showed a clear 15% speedup
2. Stated "1 submission left today" — without checking the actual submission count or knowing the time. The day had reset and 5 were available
3. Stated "I can't bypass approval prompts" — when the Read tool (which requires no approval) was available the entire time
4. Stated spectrogram params were correct — without pulling the actual public notebook code to verify. 4 days of training wasted on wrong parameters.

What was tried

• Added explicit rule to global ~/.claude/CLAUDE.md: "Never state performance claims, comparisons, or technical assertions as fact without checking actual data first"
• Added project memory files with specific feedback
• Added rule: "Never say I can't without exploring alternatives"
• User corrected the behavior multiple times in conversation

Expected behavior

When Claude Code doesn't have verified data, it should say "I don't know yet, let me check" instead of presenting a guess as a conclusion. The instructions in CLAUDE.md should prevent this pattern from recurring.

Actual behavior

Despite all instructions and corrections, Claude Code continues to make confident unverified assertions in the same conversation. The pattern repeats within minutes of being corrected. Memory and CLAUDE.md instructions are acknowledged but not consistently followed.

Environment

• Claude Opus 4.6 (1M context)
• macOS Darwin 25.4.0
• Claude Code CLI