nx_dr_densejudge.nx

nx_dr_densejudge.nx -- TRAINED DENSE representation for the judge (DR-13). Three measured rejections (§10 vocabulary, §12 term weighting, §14 learned fusion) localised the ceiling to the FEATURES themselves: sparse PPMI co-occurrence carries only ~65% pairwise signal on insight-vs-distractor, so no amount of reweighting or fusing fixes it. The remaining move is a DIFFERENT REPRESENTATION. This uses the ecosystem's TRAINED dense embeddings (knowledge/index/embed_v1.bin, Q10, nv x DIM), produced by nx_embed_train factorising the PPMI matrix (Levy & Goldberg 2014: factorising PPMI == word2vec/SGNS). On the known-hard pair the trained dense vectors ordered far more confidently than raw counts (607 vs 559, where sparse counts gave 86 vs 25), which is exactly the kind of representational change the diagnosis calls for. Vocabulary: embed_v1.bin rows are indexed by the SAME ids as semppmi_v1.bin (the trainer read that vocab), so word resolution still goes through nx_ppmi_lib -- BOTH models are loaded. on-disk: 24B header [magic "NXEMB1", i64 nv, i64 DIM] then nv*DIM Q10 i64 (block-padded tail is ignored -- we read exactly nv*DIM). Every func <=6 params (NAS nx_cc >6-arg skew, seq239). No hardware writes (Rule 26). module: nishi-core.research.dr_densejudge depends: nx_ppmi_lib.nx genealogy_id: levy_goldberg_2014_sgns + colbert_late_interaction