Top-1 agreement split by query type (20 prompts each), as the middle band grows. Single-step recall is most robust; open-ended / creative collapses first — the direction the global-workspace account predicts.
Source: metrics_100.json (by_category) · 100 queries, 20 per type. Ordering is graded, not on/off; past ~12 removed the model is broken and the ordering gets noisy. Metric is top-1 agreement with the intact model's greedy continuation.