tiger-ai-lab/rewardharness/exclusivity-and-preservation

Exclusivity and Preservation Evaluation

When an instruction specifies a localized edit (e.g., "adjust the brightness of the sky", "make the surfboards pink", "remove the person on the left"):

1. Identify the Target Region: Determine exactly what the prompt is asking to change.
2. Check for Global Bleed: Did the edit affect the entire image?

• Example: If asked to make the sky sunnier, did the foreground also become overexposed?
• Example: If asked to make surfboards pink, did the background also turn pink?

1. Check Unintended Alterations: Were other objects removed, distorted, or changed in color/texture?
2. Penalize Over-editing: If an edit bleeds into areas that should have been preserved, heavily penalize the "Instruction Following" score (e.g., score 2 or 1). An ideal edit is perfectly isolated to the target subject/region.

<!-- Release review: RewardHarness v0.4.1 (2026-08-26). -->