構建 AI 作文評分工具，讓我更深入理解 AP Lang 評分標準

I spent the last few months building an AP Lang essay grader — a tool that takes a student's essay and scores it on the official College Board six-point rubric. The hard part was not the model. It was teaching the system to grade the way a trained human reader actually grades, and that forced me to understand the rubric far more precisely than I ever did as a student.

The rubric has three rows. Row A is the thesis, worth one point. Row B is evidence and commentary, worth up to four. Row C is sophistication, worth one. Naively, you would weight all three equally in a grader. That is wrong, and testing against real scored essays showed why: Row B is where almost all the variance lives. Two essays can both nail the thesis and both reach for sophistication, and still land two points apart — entirely because one explains its evidence and the other just presents it.

So the grader spends most of its attention on one question per piece of evidence: does the surrounding text explain how this supports the argument, or does it just sit there? That single distinction separates most 3s from most 5s. It is also the thing a model can evaluate reliably, because it is structural, not stylistic.

The other lesson was about sophistication, Row C. Early versions over-rewarded fancy writing, which is exactly the mistake students make. The real rubric gives that point for complex thought — engaging a counterargument, qualifying a claim, situating the argument in a broader context — not for vocabulary. Calibrating against released College Board samples, where the official scores are known, was the only way to get the tool to stop being impressed by big words and start rewarding real reasoning.

If you are curious how the three rows actually break down, I wrote it up on the AP Lang rubric page, and there are annotated sample essays showing the bands at the examples library. Building the grader made me a better reader of the rubric; I hope it does the same for the students who use it.