nx_dr_fuse.nx

nx_dr_fuse.nx -- LEARNED signal combiner for the deep-research judge (DR-12). Two hand-reasoned refinements to the count model were measured and REJECTED (a domain- augmented vocabulary made discrimination worse; IDF weighting was a dead heat). What had never been tried is LEARNING how to combine the signals we already compute, from the real labels ServiceNow publishes. Model: score = w0*lexical + w1*semantic + w2*semantic_idf (integer, deterministic). Fitting: EXHAUSTIVE GRID SEARCH over integer weights on a TRAINING task range, scored by pairwise ranking accuracy (every insight must outrank every distractor within its own task, which is exactly the matched-precision metric the judge is graded on). Grid search rather than gradient descent: no learning rate, no convergence failure, and bit-reproducible. ★HONEST EVALUATION BY CONSTRUCTION: weights are fitted ONLY on the training task range and reported ONLY on a DISJOINT held-out range. Fitting and scoring on the same tasks would be the rigged-gate sin; the split is a parameter so the gate can prove the separation. Every func <=6 params (NAS nx_cc >6-arg skew, seq239). No hardware writes (Rule 26). module: nishi-core.research.dr_fuse depends: nx_syscalls.nx genealogy_id: linear_rank_fusion + perceptron_ranking