DEJAN organizes snippet optimization hypotheses into a locked, seven-class hierarchical taxonomy—specifically alignment, substance, architecture, style, framing, proof, and a top-level fallback for other mechanisms.

Under this system, seeded claims are categorized by a Phase 5 classifier using positional codes (such as 2.4 for comparative propositions or 1.1 for query token matching) to maintain consistency across runs. A two-stage acquisition policy then leverages these categories to sample fresh classes proportionally while applying Bayesian selection methods like Thompson sampling, ensuring optimizations test diverse structural variations. Furthermore, an automated custodian engine monitors unclassified or outlier claims to propose new, evidence-backed sub-factors for human approval when persistent patterns emerge.