nx_analyst_insight.nx

nx_analyst_insight.nx -- AUTOMATED INSIGHT MINING (F1003): point at a dataset with NO target column and get back RANKED findings. Every other analyst entry point we have answers a question you already framed (ar_analyze needs a target); this one finds the question. THE HONESTY PROBLEM THIS IS BUILT AROUND. Scanning all column pairs means testing C(k,2) hypotheses and reporting the best -- which is p-hacking with extra steps unless two things are true: (1) the significance bar is FAMILY-WISE over every hypothesis actually tested (Bonferroni over the pair count, via nx_analyst_infer), not per-test .05; and (2) the report DISCLOSES how many hypotheses were tested and what level survived, so a reader can see the multiplicity instead of being handed only the winners. A miner that quietly reports "the top 5 correlations" out of 190 tested pairs is manufacturing findings. Both are enforced below and gated. FINDING CLASSES: RELATIONSHIP (significant |r|) · NONLINEAR-MONOTONE (Spearman clears the bar and beats Pearson by a configured margin -- the pair moves together but not in a straight line, which a correlation-only scan reports as "moderate" and a human then misreads) · CONSTANT · ID-LIKE · SKEWED · OUTLIERS. Thresholds and score weights are DATA (analyst_insight.conf), not literals (rule 11). license_tier: ORIGINAL No hardware writes (Rule 26).