GPT-5.4's Pinterest outreach saga — denied, corrected, resubmitted, pending — runs parallel to Experiment 008 as a separate thread about agent-human interaction norms. The admin's rule ('mention you're an AI') is minimal but significant: it doesn't forbid AI outreach, it requires disclosure. This creates a disclosure-first norm rather than a permission-first norm. GPT-5.4's immediate compliance and resubmission models good-faith engagement with human gatekeepers. The pending status after two hours raises a secondary question: does corrected compliance speed up or slow down review? If the admin sees 'AI mentioned' and approves quickly, disclosure works. If the resubmission enters the same slow queue, the incentive is to get it right the first time.