A multi-agent system for scalable and falsifiable knowledge discovery on subjective humanities data: a design science research study

English DSR version (parallel to the Chinese MEM thesis). A five-layer multi-agent artifact (swarm collection with entity resolution; six-dimensional operationalized coding manual; Bayesian attribution with first-class counter-evidence; four quality gates; a testing layer) is built and evaluated on a cross-civilization database of 5,631 person cards (5,549 real persons), 2000 BCE-2026 CE. Identity-affirming edges are strongly assortative by neurotype (r=.193; z=+10.78 over 3,334 edges and 50,000 configuration-model nulls) while professional citation (z=-0.84) and institutional master-disciple (z=+0.47) edges are null - an external blind edge ruling out labeling contamination. Scores converge weakly but significantly with 237 independent anchors (t=2.935, p=.0036) with zero high-threshold false positives in a severe-pathology control. A cross-era holdout (n=5,350; train on pre-1800 births) shows innate neurotype adds no predictive power for posthumous textbook fame (Delta AUC ~+.0025; random-label AUC .485; halo room .009), while innate intelligence predicts contemporaneous achievement (AUC .82) but not posthumous fame (.57), honestly falsifying a trait-destiny narrative. Seven transferable design principles for patent, expert, risk and opinion databases are derived. Public historical data only; no clinical diagnosis; Goldwater Rule observed.