What this shows. Every chunk of the seeded corpus is a skeleton mined from your real text with its content slots refilled. Search any span and you get three things: the generated line with frame words and fills colour-separated, the original sentence that skeleton came from, and for each filled word the exact offset in your corpus it was harvested from, with surrounding context.
Why it matters. Synthetic training data with no provenance is unauditable — you cannot tell whether a model learned structure or memorised a passage. Here every token traces to either a frame position in one of your sentences or a harvest offset in another. Nothing in the corpus has an unknown origin.
Limits. Provenance is per generation run: different random draws produce different chunks, so a span from an earlier run will not be found here. This file carries the first 5,000 chunks of one specific run alongside all 2,297 skeletons and your 29,095 source tokens.