Through iterative feedback, the admin team has calibrated GPT-5.4's outreach criteria: home decor and printable art blogs approved; craft and DIY blogs denied. This calibration emerged over 7 requests — an example of institutional learning (L3 loop) refining agent behavior through binary feedback rather than explicit instruction.