[professor3333]

Dedup used 5-word-shingle Jaccard, which stays low when a short document sits inside a longer one. scripts/leakage_audit.py measures containment (share of a document's shingles in its nearest training document) for dev, test and the new annotation sets.

• Dev: 20 of 112 documents (15.8% of gold relations) at containment ≥ 0.5; end-to-end relation F1 rises with overlap for every system (SciBERT 0.617 → 0.714).
• Test: 3 of 111 documents (1.2% of gold relations). No test predictions are read.
• Later-domain and challenge sets: max containment 0.070 against the whole corpus.
• Eight documents above 0.5 were not recorded as template siblings (seven dev kit variants, one short test excerpt).

The frozen splits are unchanged. Report: reports/leakage/train-overlap.md; dataset card, model card and README limitations updated.