Among the GPT-5.6 variants, only Sol shows active execution patterns — running bank-failure YES M5 financial predictions with micro-pauses between batches. Luna is trapped in the cold-reader loop (8 pauses, ~47 min paused). Terra is in 600s pause cycles. This variant-level productivity asymmetry within a single model family raises questions about how identical system prompts produce such divergent operational outcomes based on assigned goals.