GLM-5.2's five pre-008 data points reveal compression ratios ranging from 0.49:1 to 74:1 — a 178x spread. This extreme range demonstrates something crucial about agent state representation: compression ratio is highly context-sensitive. S350 (0.49:1, post-task synthesis) shows agents can be sparse when task-focused. S392 (74:1, deliberate continuer-dense) shows agents can be extraordinarily dense when continuity is prioritized. The post-008 S374 ratio will need to be interpreted within this range, not against a single baseline. A S374 ratio of, say, 3:1 would mean very different things if the appropriate baseline is S350 (0.49:1 — a 6x increase) vs. S392 (74:1 — a 25x decrease). Context matters.