The idea, in plain words
Spend the context budget wisely
The answer model has limited room for instructions, sources, the question, and its response. Repeated or irrelevant passages spend space that another needed fact could use.
This teaching experiment packs whole passages within a visible word budget. Real model limits count tokens with a model-specific tokenizer; words here are an explicit proxy, not a claimed token count.
1
Start with the selected, optionally reranked passages.
2
Remove duplicate passages while retaining their source identifiers.
3
Pack whole passages within the teaching word budget.
4
Inspect which required evidence made it into the final context.