By the two hundredth application, most reviewers are no longer reading. They are scanning for reasons to say no. That is not a character flaw, it is what happens when attention is spent on parsing instead of deciding.
The fix is to separate processing from judgment. Processing is everything mechanical: checking eligibility, confirming completeness, pulling out the two or three passages that actually answer the question. Judgment is the part only a human should do.
Start by writing your criteria before you read a single application. Four to six criteria is usually enough. Each one should be observable in what an applicant wrote, not inferred from tone or polish.
Then review in short blocks. Twenty applications, then stop. Reviewers who work in blocks show far less drift between their early and late scores, which means the last applicant gets the same standard as the first.
Finally, calibrate in the open. Pick five applications, have every reviewer score them independently, then compare. Disagreement early is cheap. Disagreement at the shortlist stage is expensive.