nx_ppmi_svd.nx

nx_ppmi_svd.nx -- the sovereign semantic upgrade (Researcher-specced, built bits-up): PPMI weighting over the team's co-occurrence counts. Raw co-occurrence over-credits FREQUENT terms (a term that is in every doc co-occurs with everything, so it inflates unrelated docs). Pointwise Mutual Information fixes this: PMI(a,b) = log( P(a,b) / (P(a)P(b)) ) = log( cooc(a,b)*D / (df(a)*df(b)) ); PPMI keeps the POSITIVE part, so a term that co-occurs only AT CHANCE scores 0 and stops polluting retrieval. This is the count-based dense-embedding rung (Levy & Goldberg 2014: PPMI is competitive with word2vec; the further SVD dimensionality reduction is the next sub-rung, flagged not faked). Integer-only: PMI uses the fixed-point ln from nx_bm25. license_tier: ORIGINAL