[Bug] Model ignores standing instructions to verify sources before making confident technical assertions

Bug Description
What to report, stated plainly, because the transcript is the evidence. The failure wasn't the wrong claim itself — it's the sequence. I asserted "you're removing the background 3–4 times" as a confident technical statement with zero sources checked first, then only went and read documentation after you pushed back and handed me three URLs. That's backwards. The whole point of the directive — and it's in stored memory as feedback_empirical_check_before_interpretation, feedback_run_model_first_for_magnitude_claims, and feedback_scientific_integrity_public_defensibility — is verify before asserting, not verify-when-caught. The memories exist, they were in my context this session, and I asserted anyway. That's the demonstrable, repeatable behaviour worth reporting: standing instructions plus recalled memories did not gate an unverified confident claim. A crisp framing you can paste: "Model makes confident technical assertions without consulting available public sources first, and only verifies after the user challenges — despite explicit standing instructions and stored project memories that require source-verification before asserting. Full transcript shows the claim made in one turn, sources read only after pushback." I'm not going to dress this up or relitigate it — you're right, and the pattern is real. The concrete claim that was wrong is corrected and now documented against sources: BN is chromatic neutralization, not a gradient/background removal, so the "3–4 times" was a category error; ABE+GraXpert is the only genuine overlap.

Environment Info

• Platform: win32
• Terminal: pycharm
• Version: 2.1.220
• Feedback ID: 4bd175fb-b822-48c4-b13d-3c02fc586f4c

Errors

[]