nx_collocation.nx

nx_collocation.nx -- statistical association metrics for word pairs. First brick of the corpus-linguistics S-class substrate per cardinal feedback-corpus-linguistics-s-class-substrate-sketch-engine-displacement. Substrate goal: benchmark + exceed Sketch Engine on every measurable axis, patent-clean, all NishiLang. === What this primitive answers ==================================== Given a co-occurrence count table (n_a, n_b, n_ab, n_total) — the fundamental contingency-table summary of how often word A appears, how often word B appears, how often they co-occur, and the corpus size — emit the standard statistical association metrics: PMI Church & Hanks 1990 log2( p(a,b) / (p(a)*p(b)) ) log-likelihood Dunning 1993 -2 * sum(O * ln(O/E)) T-score Manning & Schütze (O - E) / sqrt(O) log-Dice Rychly 2008 14 + log2( 2*n_ab / (n_a + n_b) ) All Q10 fixed-point. No f64. All patent-clean (public NLP literature from 1990-2008). === Composition ==================================================== nx_essentials.log2_floor -- integer log2 used everywhere nx_tier.nx -- nx_int aliases === Sealed-enum strength bands ==================================== Per dual-reading cardinal: every primitive emits BOTH a quantitative Q10 number AND a qualitative sealed-enum band. Downstream tooling consumes the band without re-deriving thresholds. genealogy_id: church_hanks_1990_pmi + dunning_1993_log_likelihood + manning_schutze_1999_t_score + rychly_2008_log_dice + firth_1957_collocations + halliday_1961_categories lineage_id: nx_collocation_v1