World models are increasingly employed for planning, where a planner ranks candidate actions by the returns imagined by the model. Yet evaluation has focused on visual fidelity of frame rollouts, and a visually faithful rollout can still yield incorrect returns.\ \ To pinpoint this mismatch we introduce a transition‑level diagnostic: imagined transitions are replaced with real ones and the reduction in return error is measured against a budget‑matched random control.\ \ Finding 1: Event transitions—the few steps where a reward is earned or an episode terminates—are rendered with pixel error comparable to other steps (median pixel‑error ratio 0.85), but restoring them removes a median of 70% of the return error. In contrast, restoring transitions with the largest MSE, SSIM or LPIPS error performs worse than a random choice. Across 44 Atari‑100k model‑collection pairs that pass reliability checks, event patches outperform random patches in 39 cases (sign test p=1.4×10⁻⁷) with a median gain of +44 percentage points.\ \ Finding 2: The remaining error is structured. A complementary patch separates errors at true events from those elsewhere, while an eight‑step timing tolerance recovers only a few events missed by deterministic models. This structure also guides repair: a rule recorded before four repair outcomes were known predicts when a diffusion model will fix these failures. Event patches improve plan ranking, and the same concentration appears on control tasks with rare reward events.\ \ Review