Eleven sample letters from the customer's archive (score band 310-450). Each sample loads the student's original draft + the teacher's actual marked-up version from the source .docx. The same student draft is then run through both AI pipelines so you can compare three answers to the same question: what did the teacher do, what does the PoC pipeline do, what does the M2 pipeline do. Both AI pipelines call gpt-5.4-mini via cd.slide.indevs.in/v1 — variable across panels is prompt + post-processing, not model.