Three Kinds of Graders

Suppose 1,000 summaries from the job ticket assistant each need a verdict. A few lines of code that check each summary's length finish in about 1 second, at a cost too small to measure. A second model, asked to judge each summary, takes about 33 minutes one call at a time and costs about 3 dollars. A dispatcher who reads every summary needs about 12.5 hours, which costs about 500 dollars.

These figures rest on assumptions stated below, and a different setup would change them. The gaps between them would change much less. Each step up adds judgment and costs more time and money.