nx_dr_semjudge.nx

nx_dr_semjudge.nx -- SEMANTIC judge tier for the deep-research engine (DR-10). The engine's judge was the LEXICAL tier: token containment, so an insight saying "won" scored ZERO against a source saying "defeated". This escalates that socket to SEMANTIC matching over a real PPMI count model: coverage(insight,doc) = mean over insight terms of MAX PPMI cosine against any doc term -- ColBERT-class integer late-interaction, zero float. ★seq283 EATEN (this half): the loader/cosine numerics are NO LONGER duplicated here. They live once in nx_ppmi_lib.nx and this file is a thin ADAPTER preserving the sj_* API its gate and CLI already use. Two hand-rolled copies of the same numeric code could silently diverge and change similarity scores with no gate noticing -- that risk is gone for this organ. (nx_recall_dense still holds its own pb_* copy; it is a LIVE organ owned by the recall lane and migrates on its owner's next touch, per the D001 migrate-on-touch law.) ⚠INTEGRATION LAW: the PPMI vocab is keyed by db_semhash (nx_qabench_engine), NOT the djb2 lexical hash -- the wrong hash silently returns -1 for every word and zeroes every similarity. nx_ppmi_lib owns the lowercase-then-db_semhash resolution. ⚠HONEST CEILING: a count model is bounded by its corpus vocabulary; out-of-vocabulary terms score 0. Use nx_dr_semjudge_x to A/B an alternative model against that ceiling. module: nishi-core.research.dr_semjudge depends: nx_ppmi_lib.nx genealogy_id: levy_goldberg_2014_ppmi + colbert_late_interaction