Measure and prevent fake work in Claude Code sessions. 3-layer defense (memory feedback + Stop hook + slash command) + 4 audit tools. Baseline: 24.91% loose / 8.28% strict fake work rate across 813 real sessions.
python bash hooks ai-safety hallucination ai-agent self-verification claude-code fake-work-detection
-
Updated
Apr 8, 2026 - Python