Evidence L (v2): matched swap. A' trained by curvature_surveyor on documents 1-10 of each RR category with exactly adapter B's recipe, LoRA init seed 2000 and data-order seed 1001; B trained on documents 11-20.
Files: swapAp_blk_*.npz (scores under A'), e31_docloss_Ap.npz (A' self-losses on the 1,500 documents), rr_adapterAprime.json (official validator output for A'), expected.json.
Run: python check_swap.py PATH_TO_UNZIPPED_ARCHIVE_G (numpy only); B's score matrices (swapB_blk_*.npz) come from archive G. Groups are reconstructed from the IDs (member flag + file order within category); scale factors are restored before subtraction.
Scores: float16 after division by a positive constant (field 'scale'); rounding creates ties, handled with half credit.
