A canonical four-state macro-architecture governs multimodal fear-goal arbitration across learning and relapse

Fear, safety, and goal pursuit are often studied with separate behavioral readouts, leaving open whether multimodal behavior is organized by continuous response intensity or by discrete latent modes. We reconstructed a four-channel, 5-s time-binned rat dataset spanning fear learning, threat generalization, long-delay retest, extinction, retrieval, and renewal. The primary complete-case analysis contained 47,004 four-channel bins from 29 rats and 213 subject-sessions. We compared continuous latent-factor baselines with hidden Markov and semi-Markov sequence models using freezing, 22-kHz ultrasonic vocalizations, 50-kHz ultrasonic vocalizations, and lever pressing. Discrete sequence models strongly outperformed PCA/factor-analysis baselines. Although empirical fit improved up to eight microstates, these states coarse-grained into four biologically interpretable macro-modes: Quiescent/Baseline, Safety/Appetitive, Danger/Fear, and Goal-directed behavior. The four-state macro-model captured 96% of the optimal cross-validated likelihood and was highly stable across random initializations and leave-one-subject-out validation. Across conditioning, extinction, and renewal, emission signatures remained comparatively stable while transition matrices reconfigured strongly. Session-level fecal boli were predicted by decoded state occupancy better than mean freezing alone. These findings support a canonical discrete macro-architecture for fear-goal arbitration, with learning and relapse expressed primarily as context-dependent transition reweighting. Author SummaryWhen an animal faces both danger and opportunity, it must do more than turn its responses up or down: it must choose which behavior takes control, whether to freeze, to seek reward, or to act. Fear-conditioning research has largely tracked a single behavior, freezing, along a simple threat-versus-safety scale. We asked whether the fuller picture of behavior is better captured by such a continuous scale or by switching among a small set of discrete behavioral modes. Recording four behaviors at once in rats, freezing, alarm 22-kHz calls, appetitive 50-kHz calls, and lever pressing for reward, across fear learning, extinction, and relapse, we found that discrete-state models fit far better than continuous ones, and that behavior organizes into just four recurring modes: quiet baseline, safety and reward-seeking, fear and distress, and goal-directed action. Learning and relapse mostly rewire the transitions between these modes rather than changing the modes themselves, and an independent physiological stress marker tracked the models states. This offers a reusable, computational way to read emotional-motivational behavior and to view relapse as a change in behavioral routing.