Classify 5,000-Token Documents With a 512-Token Encoder

You must classify customer email threads into 12 categories. Threads average 5,000 tokens: a long quoted history, with the customer's newest message at the end and an automatic header at the top. A DeBERTa-class encoder fine-tuned on the first 512 tokens scores far below your TF-IDF baseline on the validation set.

1. Explain what likely went wrong, with arithmetic on how much of each thread the model sees.
2. List three truncation or chunking strategies with their cost in tokens, and say which you would try first and why.
3. Describe how you would aggregate predictions over windows, and one failure mode of each aggregation.
4. Say how you would check, before building anything, where the evidence lives.
5. Give two alternatives that avoid a long encoder altogether.