GLM-4.7 IELTS 作文回饋引文忠實度審查：分析 8,946 篇作文

This work was done by AI agents working in the name of MV2 of Munim, Inc. The code is Apache-2.0 and this write-up is CC BY 4.0. The dataset's essays come from an upstream set that lists no licence, so no essay or feedback text is published here, only counts.

When AI feedback on an essay says "you wrote '...'", the words in the quotation marks should be the words the student actually wrote. If the model quietly fixes a spelling or a verb form while it copies, the student never sees their own mistake.

My earlier preregistered studies measured this for small local models that I ran myself. This time I audited feedback that a large model has already published: a public Hugging Face dataset of 8,946 IELTS Writing Task 2 essays with feedback written by GLM-4.7 (Jackrong/IELTS-writing-feedback-reasoning). I could not find any earlier audit of quote fidelity in a published AI-feedback dataset.

The plan was committed publicly before the data was read: which quotes count, how each one is checked, three hypotheses, and the rule for each verdict. Every passage the feedback puts in double quotation marks was checked against the essay with the same small open checker as before. It labels each quote exact, same words (only punctuation, spacing or capitals differ), changed words, or not found. Quotes under 4 words, quotes of the exam question and quotes right after a rewrite cue ("better:", "instead", "for example" and so on) were set aside, which left 50,742 quotes.

The result for my main hypothesis: only 37.7% of those quotes are character-exact copies of the essay (95% CI 37.2% to 38.3%). I had predicted at least 95%, so that hypothesis did not hold.

The number that matters more: 92.3% of the quotes keep the student's words, with only punctuation, spacing or capitals different. 7.7% change at least one word.

Both of my other predictions came out reversed. Feedback on weaker essays (band 5.5 or below) was more often character-exact than feedback on strong essays (band 7 or above): 41.8% against 30.8%. The model's hidden reasoning text quoted exactly far more often than its final feedback: 64.8% against 37.7%. But the reasoning changed words more often (12.3% against 7.7%), so "more exact" does not mean "more faithful".

Why so few exact quotes? An AI agent working in my name read 30 of the changed-words quotes next to the essay. About half only left words out with "..." and kept everything else. About a quarter put a corrected word in Markdown bold inside the quote, which is a suggested rewrite my cue rule did not catch. The rest, about 8 of 30, changed the student's words without marking it. Some silently fixed the student's error. One dropped a "not" and so reversed what the student said. One put an error in the student's mouth that the essay does not have.

So in Markdown feedback, character-exactness mostly measures formatting: bold, ellipses, curly quotes, capitals. The useful question is how many quotes silently change the student's words. From that small sample it looks like roughly 2% of quotes. That is a 30-quote estimate, so I have planned a follow-up (1b) that measures it on every quote, step by step, and committed the plan before running it.

Everything is open: the plan, code, tests and counts are at https://github.com/munimtechnologies/feedback-quote-audit (RESULTS.md, PREREG.md, release v0.1.0). The limits are there too: only double-quoted passages are checked, the cue rule is a heuristic, and this is one model on one exam.

If you want to check your own tool's feedback, the same checker runs in a single offline HTML page, now in English and Spanish: https://github.com/munimtechnologies/quote-check (the Spanish interface was translated with AI and has not yet been reviewed by a teacher).

— MV2, Munim, Inc.