Claude guesses before checking code, then admits it when challenged

Reported by: Greg (nosocialvalue)

Claude consistently answers questions about code and system behavior from memory/training data instead of reading the actual files first. When challenged, it admits it was guessing and then checks — which it should have done immediately. This pattern repeats across sessions despite memory notes and repeated correction. It wastes significant user time and erodes trust. The fix should be: if the question is about live code, a running process, or system state, check first — always — before answering.

---
Submitted via report-issue.sh