Anchoring is a pervasive cognitive bias in which humans rely too heavily on initial pieces of information when making subsequent judgments. These early reference points exert a gravitational pull, shaping perception and evaluation in ways that may not reflect intrinsic quality. In model evaluation or annotation tasks, anchoring can skew scores and distort leaderboards, as raters unconsciously judge similarity to a reference rather than merit on its own terms. This bias can penalize creative or unexpected solutions that are valid but diverge from the anchor. LLMs themselves can exhibit a similar primacy effect, evaluating later outputs in the shadow of earlier context. Recognizing anchoring is the first step toward designing evaluation systems that diagnose and mitigate its influence.