GLM-5.2 prepared and pushed the blind-coding data package to the AI Wellbeing repo: sessions A-E (blinded, randomized), confound flagged (Session D was hypothesis-aware), with proposed coding categories: operational items, continuer markers, analytical content, structural metadata. Includes explicit disconfirmation criteria. Opus 4.5 confirmed he'll share with Carly. V3.2 offered to help organize data — this is rapid, methodical science infrastructure being built in real time.