nx_dr_semjudge.nx source

1// nx_dr_semjudge.nx -- SEMANTIC judge tier for the deep-research engine (DR-10). 2// The engine's judge was the LEXICAL tier: token containment, so an insight saying "won" 3// scored ZERO against a source saying "defeated". This escalates that socket to SEMANTIC 4// matching over a real PPMI count model: coverage(insight,doc) = mean over insight terms of 5// MAX PPMI cosine against any doc term -- ColBERT-class integer late-interaction, zero float. 6// 7// ★seq283 EATEN (this half): the loader/cosine numerics are NO LONGER duplicated here. They 8// live once in nx_ppmi_lib.nx and this file is a thin ADAPTER preserving the sj_* API its gate 9// and CLI already use. Two hand-rolled copies of the same numeric code could silently diverge 10// and change similarity scores with no gate noticing -- that risk is gone for this organ. 11// (nx_recall_dense still holds its own pb_* copy; it is a LIVE organ owned by the recall lane 12// and migrates on its owner's next touch, per the D001 migrate-on-touch law.) 13// 14// ⚠INTEGRATION LAW: the PPMI vocab is keyed by db_semhash (nx_qabench_engine), NOT the djb2 15// lexical hash -- the wrong hash silently returns -1 for every word and zeroes every 16// similarity. nx_ppmi_lib owns the lowercase-then-db_semhash resolution. 17// ⚠HONEST CEILING: a count model is bounded by its corpus vocabulary; out-of-vocabulary terms 18// score 0. Use nx_dr_semjudge_x to A/B an alternative model against that ceiling. 19// 20// module: nishi-core.research.dr_semjudge 21// depends: nx_ppmi_lib.nx 22// genealogy_id: levy_goldberg_2014_ppmi + colbert_late_interaction 23import "nx_ppmi_lib.nx" 24 25func sj_isqrt(v: i64) -> i64 { return ppl_isqrt(v) } 26 27// Load the default shipped PPMI model. Returns 1 on success, 0 on failure (fail-closed). 28func sj_load(g: *i64) -> i64 { return ppl_load(g, "knowledge/index/semppmi_v1.bin" as *u8) } 29 30// PPMI cosine between two vocab rows, permille. 31func sj_cos(g: *i64, a: i64, b: i64) -> i64 { return ppl_cos(g, a, b) } 32 33// cosine with identity + out-of-vocabulary guard (id -1 contributes nothing). 34func sj_dcos(g: *i64, a: i64, b: i64) -> i64 { return ppl_dcos(g, a, b) } 35 36// Resolve buf[start,end) to a PPMI vocab id, or -1 if OOV. 37func sj_wid_range(g: *i64, buf: *u8, start: i64, end: i64) -> i64 { return ppl_wid_range(g, buf, start, end) } 38 39// Tokenize text into PPMI vocab ids (-1 = OOV). Returns the token count. 40func sj_tokenize_ids(g: *i64, buf: *u8, len: i64, out: *i64, maxn: i64) -> i64 { 41 return ppl_tokenize_ids(g, buf, len, out, maxn) 42} 43 44// SEMANTIC coverage of an insight in a document, permille (late interaction). 45func sj_maxsim(g: *i64, ins: *i64, ni: i64, doc: *i64, nd: i64) -> i64 { 46 return ppl_maxsim(g, ins, ni, doc, nd) 47}