We present ZUSE Automat Agent (ZUSE), a deterministic, policy-driven discovery loop for elementary cellular automata (ECA) that builds an empirical atlas of cycle laws, world families, and basin fragility without a language model in the discovery loop. The pipeline combines a fixed seven-law evaluator, a dedup-gated observer stack, and persistent multi-seed world records, and is applied to a 20-world sample spanning ECA rules, Conway's Game of Life patterns, and synthetic controls. The seven laws — velocidad_constante, periodicidad, densidad_estable, tipo_unico, complejidad_alta, frontera_temporal, and temporal_scale_stability — are calibrated empirically (frontera_temporal upper bound 0.4352; temporal_scale_stability threshold 19.03, decision-tree accuracy 0.908). Laws are separated into two groups: structure-observer laws that depend on the full observer stack, and frame-metric laws that depend only on aggregate frame statistics. Key results: (1) frontera_temporal is not intrinsically rare — it activates in 38 of 256 ECA rules under at least two of three seeds, and in 17 of 256 under all three; (2) the 20-world atlas reveals seven dynamic categories, finer than Wolfram's four-class taxonomy; (3) one-bit IC fragility spans from perfectly stable basins to exact fixture disruption, with rule_108 as the ECA outlier; (4) an exhaustive protocol confirms rule_108 as the unique ECA rule producing stationary local period-2 oscillators; (5) designed periodic ICs activate production periodicidad in 207 of 256 ECA rules; and (6) a controlled single-bit experiment separates physical ECA dynamics from observer/dedup measurement artifacts. Every result is reproducible from deterministic scripts with no stochastic components in the discovery loop.