Introduction

Research on learned threat has long relied on defensive responses, especially freezing, to establish how danger is learned (1–3). After extinction or another intervention, however, a decline in that response shows that defense has changed; it does not reveal how the wider behavioral repertoire has reorganized. An animal may become less defensive yet remain behaviorally quiescent, or it may re-engage with the environment and resume effective goal pursuit. Recent temporal ethograms further show that conditioned threat cues suppress reward seeking while organizing diverse defensive and orienting behaviors (4, 5). These outcomes are not interchangeable, and a preclinical model of recovery must follow the entire repertoire across learning, intervention, delay, and relapse. The present studies adapt fear conditioning, extinction, delay, and contextual relapse paradigms to that purpose. Their experimental logic begins before danger. An animal first explores its surroundings and pursues outcomes that experience has made attainable. Once threat enters that history, the same individual must reorganize behavior around avoiding harm without permanently abandoning environmental engagement or learned goals. Together these findings define the problem addressed here: what returns after extinction, delay, or a shift of context? We call this problem fear-goal arbitration, using arbitration in a functional sense for the organization of competing behavioral priorities rather than any specific decision mechanism.

Answering that translational question requires indicators that span the repertoire rather than a single dimension of fear. Four well-validated rodent measures meet this requirement, each contributing a different kind of evidence. Freezing, the absence of all movement except respiration, is the canonical measure of conditioned defense (1–3). The distinction is biological, not merely descriptive: lesions of the periaqueductal gray can disrupt conditioned freezing while sparing conditioned suppression of instrumental behavior (6), so reduced reward seeking is not freezing by another name. The two ultrasonic vocalization (USV) bands add complementary but context-dependent affective information: 22-kHz calls accompany aversive and alarm contexts, whereas 50-kHz calls arise across play, social contact, reward, and novelty (7–10). Because the present calls were pooled across acoustic subtypes, 50-kHz output may signal context-sensitive engagement but cannot be equated with arousal, appetite, reward seeking, or positive affect. Food-reinforced lever pressing, by contrast, reports directly whether instrumental goal pursuit persists under threat. Together the channels sample defensive behavior, context-sensitive vocal engagement, and instrumental action, while their joint absence marks a fourth, low-output pattern; collapsing them to a single score would discard the very organization the study sets out to recover.

Even the defensive side of this repertoire is multidimensional. The Fear Threshold Model holds that objective defensive reactions and candidate indicators of subjective fear need not cross their thresholds together as threat rises. Derived from a systematic review and meta-analysis (11), it drew experimental support when freezing and 22-kHz calls diverged across shock intensity, cue timing, and anxiety, the calls emerging later while freezing stayed comparatively stable (12). No single defensive response, then, captures the animal’s full fear-related state. Yet this evidence speaks only to the danger-facing side; it is silent on what follows when defense declines relative to 50-kHz vocal engagement and instrumental action.

The same recovery question is longitudinal as well as momentary. When overt defense subsides, does the individual stay in a low-output state, enter a 50-kHz-dominant exploratory pattern, or return to a learned goal? Extinction, for instance, may reflect inhibitory learning, context-dependent retrieval, or reconsolidation-related updating rather than erasure of the original memory, and defensive responding can return with time or a change of context (13–17). A drop in any one defensive response therefore establishes neither emotional recovery nor restored function. For animal models of exposure-based treatment, the sharper question is whether safety learning reshapes the whole trajectory, from defense through quietness toward renewed engagement and task performance, and whether that reorganization survives relapse challenges. Answering it requires multimodal measurement across learning history, not a single fear score at a single moment.

Once recovery is framed as reorganization of the whole repertoire, tri-reference-point (TRP) theory offers a compact way to relate that repertoire to functional landmarks. Formulated for decision making under risk, TRP distinguishes a minimum requirement, the status quo, and a goal, which partition outcome space into regions of failure, loss, gain, and success (18). We invoke it alongside, not instead of, established accounts of threat-imminence-dependent defense, risk assessment, and foraging under predation risk (19–22), which explain why threat context changes action. Carrying this reference-dependent logic into fear-goal arbitration is natural because risky choice and affective organization confront the same broad problem: averting unacceptable outcomes while pursuing desirable ones. In the extension, the three references bound four candidate regions (Figure 1A): clear danger below the minimum requirement, a low-output zone between the minimum and the status quo, an exploratory zone between the status quo and the goal, and overt goal pursuit beyond it. These regions are sampled by defensive, low-output, exploratory-vocal, and instrumental channels rather than equated with them. The Fear Threshold Model describes differentiation on the danger-facing half of the scheme; the broader framework asks what follows declining defense: quiescence, 50-kHz-dominant engagement, or renewed instrumental action. The proposal concerns functional priorities alone: it neither replaces the established ethology of threat and foraging nor credits rats with explicit verbal reference points or a single literal psychological axis.

Figure 1.

The TRP-derived hypothesis supplies functional landmarks, but it neither locates them in a behavioral time series nor guarantees that observations will form discrete recurring modes. Computational neuroethology holds that behavior is better described by its structured sequences than by isolated observations (23–25), and behavior-systems modeling in experimental psychology likewise treats hidden states as a way to formalize the temporal organization of motivated action (26). A hidden Markov model (HMM) tests that stronger prediction empirically. Within each five-second window the four behaviors are observed directly, while the mode that organizes them is inferred; hidden refers only to this statistical inference. The model estimates an emission signature for each recurring multimodal pattern and a transition matrix for persistence and movement between patterns (27, 28), so that neighboring windows inform one another. Its role is to preserve the identity of all four channels while testing whether defensive, exploratory-vocal, instrumental, and low-output patterns recur, and how the animal moves among them over time.

A single commitment governs how these modes may be read. Functional accounts treat emotion states as internal states inferred from coordinated behavioral and physiological responses and marked by general properties such as persistence and generalization (29). Our question is narrower: whether the HMM recovers candidate functional modes that combine several observed channels, persist across neighboring windows, and recur across animals and studies. Meeting those criteria does not make the modes direct reports of subjective experience. Freezing and a 22-kHz call show why the qualification matters: both are objectively recorded, yet freezing indexes a defensive reaction while a 22-kHz call may also carry nonverbal information about subjective fear, without excluding communicative or adaptive readings. The distinction concerns inferential role, not observability, and does not turn on whether animal experience is conscious or non-conscious (11, 12, 30–32). It prevents us from equating any mode with subjective feeling while leaving the organization of defense and reward pursuit fully open to analysis.

Separating functional theory from statistical model shaped the analyses. K=4 was fixed from the four proposed behavioral regions before numerical validation, and four linked questions followed. First, did the data resolve into separable Fear/Danger, Quiescent, Exploratory, and Goal-directed signatures without any mode being defined by a single channel? Second, did that partition withstand variance-floor and restart sensitivity and stay recognizable across the three study datasets? Third, did a discrete-aware state-count sweep place K=4 within a higher-resolution hierarchy, and did count-native emission models preserve the same partition? Fourth, did the decoded rat-session paths distinguish defensive recruitment, fear reduction, low-output quietness, exploratory re-engagement, task-specific goal pursuit, and relapse across conditioning, extinction, delay, and context? These analyses test K=4 as a useful functional vocabulary within the validated Gaussian representation; they do not test whether nature contains exactly four states, or whether rats implement TRP as a decision algorithm.

Conceptual and experimental framework

The three experiments sampled successive disruptions and partial restorations of this proposed landscape (Figure 1; Table S1). Rats first learned food-reinforced lever pressing, and initial chamber exposure also allowed exploratory engagement. Study 1a then introduced 9-kHz fear conditioning and relaxed threat certainty through context, cue, and lower-frequency generalization tests. Study 1b retested the same 12 rats approximately five months later, asking whether weaker memory expression coincided with renewed engagement. Study 2 embedded the learned food-seeking response within conditioning, retrieval-extinction or time-matched exposure, retention, renewal, and repeated extinction, asking whether reduced defense was accompanied by task function and whether either survived relapse. These are related perturbations rather than independent replications; together they vary danger learning, extinction experience, elapsed time, context, novelty, and access to reward within one common measurement scheme.

Within that scheme the measures were chosen to complement one another. Freezing indexed classical defense; the 22-kHz and 50-kHz USV bands supplied aversive and context-sensitive vocal channels; and lever pressing tracked food-directed action. The 50-kHz band was kept for what it can reveal, engagement beyond overt defense across social, rewarding, and novel situations, not for what it cannot settle, appetite or reward pursuit in particular. Fecal boli were held back as a session-level autonomic measure. Each mode could then be defined by the co-expression and mutual suppression of all four channels rather than by any single response (Table S2), while an external physiological signal stayed in reserve for a later validity test.

How to read the model outputs

Theory and model operate at different explanatory levels, and their vocabularies are kept distinct. The TRP-derived hypothesis motivates candidate functional regions; the HMM estimates recurring patterns and transitions without asserting that those patterns are literal reference points. Throughout, state and mode are used interchangeably for a model-inferred category. An emission signature is the four-channel pattern expected within a mode; occupancy is the fraction of five-second windows assigned to it; dwell duration is the length of an uninterrupted run in it; and a transition is either persistence in a mode or a move from one mode to the next. The HMM gives every subject-session its own decoded path while sharing emission and transition parameters across the dataset. Historical structure enters at two levels: within a session, the HMM captures seconds-scale dependence among neighboring windows; across sessions, conditioning, extinction, elapsed time, and relapse define the learned history against which paths are compared. The HMM does not estimate memory states across sessions. Pooled occupancy weights every retained bin and so depends on which sessions were sampled and how long they lasted; it is neither a baseline prevalence nor a target for normal functioning. Functional interpretation comes instead from changes within animals, sessions, and phases. These quantities do not by themselves identify subjective content, a neural circuit, or a causal decision process.

Results

A four-mode account is informative only if the observations are semantically sound, the fit is numerically defensible, the signatures are stable within the stated representation, and the decoded paths separate forms of change that a single defensive response would blur. The Results follow that logic: from reconstruction and boundary-aware model selection to mode identity and stability, then to history-dependent occupancy, switching, dwell, and autonomic association, and finally to the state-count and emission-family sensitivities that bound the claim.

Reconstruction preserves no-event versus no-recording

The first requirement was an observation matrix that kept absence distinct from missingness. Raw exports were harmonized onto a 5-s grid, where a subject-session denotes one rat observed during one experimental session. For USVs and lever presses, no event within a recorded session was encoded as zero, while an unrecorded modality stayed missing; freezing of 0% was a real value, and unavailable measurements stayed missing. The full reconstruction held 74,164 rows, of which the primary analysis retained the 47,004 with all four emissions observed, from 29 rats and 213 subject-sessions (Table S3). Exclusion thus followed modality availability and freezing completeness, never outcome values (Figure 1; Figure S1). The distinction matters because the count channels contain many genuine zeros, the source of the numerical boundary problem examined next.

Discrete-aware validation selects a stable K=4 Gaussian specification

The first numerical test concerned the boundary behavior of the theory-guided K=4 model. K=4 fixed the functional scale; validation chose the numerical specification of that scale, not the number of modes. A variance floor is the smallest within-mode spread the model may estimate. With many genuine zero counts, the 1e-6 floor assigned enormous Gaussian density to an exact zero while leaving little for nearby nonzero counts; fully 93.4% of its apparent continuous-density advantage arose in all-zero bins. Because counts such as 0, 1, and 2 are discrete outcomes rather than infinitely precise points, that advantage reflected boundary behavior, not a better behavioral partition.

The discrete-aware score changed model selection. Diagonal covariance with a 0.01 variance floor ranked first (mean held-out log probability −2.094 per bin, SE 0.034), while the 1e-6 floor fell outside the one-SE set (Figure 2A; Figure S2; Table S4). The score assigned each count the probability of its integer-width interval on the log1p scale, while freezing retained Gaussian density. A diagonal model estimates a separate residual variance per channel within each mode, avoiding an unstable full covariance matrix, yet state assignment still depends jointly on all four channels and on their temporal order. With the boundary controlled, the substantive question was whether the fixed K=4 scale recovered the proposed signatures.

Figure 2.

Four emission signatures separate quiescent, exploratory-vocal, defensive, and goal-directed behavior

The selected model resolved four distinct signatures, arranged from the danger-facing side of TRP toward the goal (Figure 2B). Fear/Danger (7.8%) joined 22-kHz calls (2.29), some 50-kHz calls (0.79), and high freezing (67.7%). Quiescent held 81.7% of all retained bins, the largest pooled share, with near-zero event-channel output and fitted mean freezing of 34.5%. Exploratory (7.9%) was a 50-kHz-dominant pattern (1.93 calls per bin) with lower freezing (22.1%) and negligible lever pressing, consistent with novelty-sensitive engagement under low overt defense. Goal-directed (2.5%) was defined by lever pressing (1.00) with little vocal output (Figure 2B,C; Table S5). These percentages aggregate bins across unequal session types, durations, and histories, including periods shaped by the introduction of shock and periods when goal pursuit was unavailable or suppressed; they describe the sampled schedule, not normal, healthy, or desirable proportions. Their value lies in how occupancy shifts within animals and phases, not in the global ranking.

The labels claim only what the signatures license. Fear/Danger and Exploratory name joint emission patterns, not readings of subjective feeling: Fear/Danger binds high freezing to 22-kHz-dominant vocal output, and Exploratory marks a 50-kHz-dominant contextual pattern consistent with novelty-sensitive engagement under low overt defense. Because the 50-kHz calls were pooled across acoustic subtypes and recur in many contexts, Exploratory is not proof of exploration, novelty seeking, safety, or positive affect. Quiescent names low vocal and lever-press output, not an absence of freezing, movement, or arousal, and Goal-directed is anchored by task-specific instrumental lever pressing. Both vocal modes are read by joining their observed channel pattern to prior evidence that objective defensive reactions and candidate subjective indicators can diverge (11, 12). Fitted and empirical decoded-state means agreed closely, except where a fitted Gaussian mean summarizes the sparse upper tail of a count distribution (Figure S3); labels followed the complete four-channel signature, and the numeric state indices remain arbitrary. The next question was whether this vocabulary survived a change of initialization, boundary choice, or study.

The selected Gaussian partition is stable across restarts, variance floors, and studies

Within a fixed representation, a useful vocabulary should recur across valid fits. Of 100 full-data restarts, 88 passed the prespecified numerical checks. Seventy occupied the dominant likelihood basin, including the selected seed 9, while the rest formed three lower-scoring basins (Figure 3A; Table S6). Normalized mutual information (NMI) compares two complete assignments, 1 marking identity after relabeling and lower values increasing disagreement. Pairwise NMI among valid partitions had median 1.000 and minimum 0.608, revealing one dominant solution amid a minority of local optima (Figure 3B; Figure S4). Across folds, floors 0.001 and 0.01 gave nearly identical partitions (minimum aligned NMI 0.994), whereas 1e-6 produced a different one (about 0.55-0.65; Figure 3C; Table S7).

Figure 3.

Recurrence across studies raised a different question: did each mode keep a similar four-channel profile despite the experimental differences? After alignment by pooled emission signatures, within-study empirical signatures had cosine similarity 0.914-1.000 (median 0.996) with the pooled fitted signature, 1 denoting the same profile direction (Figure 3D; Figure S5; Table S8). Subject occupancies varied by experiment, as expected under the different perturbations (Figure S6; Table S9), while the defining multimodal profiles stayed recognizable. Because Study 1b reused the Study 1a rats, a further sensitivity treated them as one cohort and Study 2 as a non-overlapping cohort: with the all-data Gaussian representation fixed, all four signatures were recognized in both directions, 9,980 of 10,000 subject-bootstrap replicates were valid in each direction, and every valid replicate matched all four. This is fixed-partition recurrence, not model transfer or independent refitting. Within the validated Gaussian representation, the vocabulary was stable enough to ask how learning changed its use.

Learning and relapse redistribute mode occupancy

Learning first appeared as a redistribution of occupancy, the fraction of five-second windows assigned to a mode. To compare early and late portions of sessions of different lengths, each subject-session was split into ten ordered segments, or deciles. During conditioning, mean Exploratory occupancy fell from 55.8% in the first decile to 1.1% in the last, while Fear/Danger rose from 0% to 64.1% (Figure 4A). Subject-level Spearman trends were negative for Exploratory (mean rho=-0.704) and positive for Fear/Danger (mean rho=0.835), both surviving sign-flip testing and false-discovery-rate correction. These opposing trends establish within-session reorganization, though they cannot separate associative fear learning from habituation to novelty, which also quiets 50-kHz calling (7). During the first extinction session, Fear/Danger rose transiently to 17.7% in decile 3 and fell to 2.7% by decile 10, yet that final decile was 95.8% Quiescent, 0.6% Exploratory, and 1.0% Goal-directed (Figure 4B); reduced defense did not, in that session, amount to broad exploratory or task recovery. History and context then separated later outcomes. At the five-month retest in a different chamber, occupancy was Exploratory-rich (61.5%) with no Fear/Danger bins, a descriptive pattern that may reflect novelty or re-familiarization as much as remote-memory expression. When Study 2 extinction and no-extinction groups were pooled for the renewal session (n=16), occupancy was 24.7% Fear/Danger and 1.9% Goal-directed.

Figure 4.

These cross-cohort descriptive contrasts, rather than one common within-animal sequence, show what the state vocabulary adds: it can tell quieting, re-engagement, task performance, and renewed defense apart instead of treating every fall in freezing as the same endpoint (Figure S6; Figure S7; Table S9; Table S10). The route between modes requires transitions.

Fear-goal conflict appears as constrained, asymmetric switching

Transition analysis separated fitted routing from paths actually decoded. Fitted probabilities appear in Figure 2D and Table S11; the empirical matrix counts realized paths (Table S12), with session-specific matrices preserving subject-session boundaries (Figure S8). Every adjacent pair contributes an outgoing transition, persistence included; an exit counts only a move to a different mode. Fear/Danger-to-Goal-directed transitions occurred 14 times, 0.38% of all outgoing Fear transitions and 2.31% of Fear exits; the reverse occurred 10 times, 0.85% of all outgoing Goal transitions and 0.86% of Goal exits. Of exits, 95.9% from Fear/Danger and 88.4% from Goal-directed entered Quiescent. Under the primary local-order null, which permuted episode order within ten contiguous segments of every subject-session (2,000 permutations), Quiescent routing was enriched from Fear (observed/expected 1.068, p=0.0010) and Goal (observed/expected 1.126, p=0.0010), while Goal-to-Fear (observed/expected 0.305, p=0.0010) and Fear-to-Goal (observed/expected 0.448, p=0.0020) were depleted. An exact conditional run-label sensitivity held each session’s number and multiset of run labels fixed, but not run lengths or bin occupancy; under it, Goal-to-Fear stayed depleted (observed 10, exact expectation 19.725, Monte Carlo p=0.0117), whereas Fear-to-Goal did not (14 vs 19.725, p=0.1443).

Fear-to-Quiescent (580 vs 570.087, p=0.1008) and Goal-to-Quiescent (1,028 vs 1,012.907, p=0.0686) also failed to cross 0.05. The broad bidirectional-depletion claim is therefore unsupported: Goal-to-Fear depletion is robust to both nulls, Fear-to-Goal depletion is null-dependent, and Quiescent enrichment is specific to the locally occupancy-preserving null (Figure 4D; Table S13).

Dwell duration rounded out the temporal picture. Quiescent runs were longest (median 20 s), Fear/Danger runs had median 15 s, and Goal-directed runs were usually single five-second bins (Figure S9; Table S14). The decoded paths descriptively set a low-output mode between many defensive and instrumental runs, and the local-order null supports enrichment of that route; the exact run-label null, however, did not support Quiescent enrichment at 0.05. A reorientation interval is thus one plausible reading, consistent with functional and circuit accounts of threat imminence and defensive response selection (33–36), rather than an identified neural gate or explicit decision stage. Five-second bins and residual phase structure further limit mechanistic inference. A last validity test asked whether the behaviorally defined Fear/Danger mode was associated with an autonomic measure kept out of model fitting.

Fear/Danger occupancy is associated with an autonomic measure

Fecal boli supplied that independent, session-level autonomic signal. Across 130 sessions from 17 rats, a negative-binomial generalized estimating equation associated Fear/Danger occupancy with boli count. Because only 17 subject clusters were available, primary inference used the Mancl-DeRouen bias-reduced covariance: beta=1.455 per unit occupancy proportion, 95% CI 0.640-2.271, p=0.000471 (37). With Quiescent as the omitted compositional reference, a ten-percentage-point shift from Quiescent to Fear/Danger, other proportions fixed, corresponds to about a 16% higher expected boli rate (95% CI approximately 7%-25%). Because both quantities varied across the schedule, a further sensitivity added indicators for the eight session types; Fear/Danger stayed positively associated with boli count under bias-reduced inference (beta=1.319, 95% CI 0.084-2.554, p=0.0363), and broader three- and four-phase codings gave similar estimates (p=0.0373 and p=0.0283; Table S15). Exploratory and Goal-directed occupancy were not supported in the unadjusted model (p=0.264 and p=0.235). In a separate model, mean freezing alone was likewise unsupported (beta=0.00795 per percentage point, 95% CI −0.00236 to 0.01826, p=0.131; Figure 4C; Figure S10; Table S15). Conventional cluster-robust estimates are retained as sensitivity results.

The contrast between these models is narrower than a claim that occupancy outperforms freezing. Failing to reach a statistical threshold does not prove that freezing lacks an association, and the two predictors were evaluated in separate prespecified models rather than compared directly. The supported conclusion is a concurrent association between a multimodal Fear/Danger summary and an external autonomic measure held out of HMM fitting; the session-adjusted sensitivity makes that conclusion less dependent on the experimental schedule. Reading this summary as part of a general four-mode architecture therefore required testing whether its scale and exact assignments survived changes in model resolution and emission family.

Scale and emission-family sensitivities bound the four-mode claim

Under the same subject-grouped, discrete-aware score, held-out performance improved from K=2 through K=8; K=8 ranked first, K=7 and K=8 formed the one-SE set, and K=4 did not. Higher resolution refined rather than reorganized the theory-guided regions in the selected fits: their weighted nesting purity was 0.9981-0.9999 and their unweighted mean state purity 0.9943-0.9998, every selected higher-K state mapped to one K=4 mode with minimum individual-state purity 0.9616, and all four K=4 modes remained represented at every K. Fear/Danger split at K=7 and K=8, and Goal-directed at K=8. Across all 330 strictly valid higher-K restarts, minimum weighted purity was 0.9803 and minimum unweighted mean state purity 0.9456, though the per-state no-merge conclusion holds only for the selected fits. Post hoc inspection clarified what the extra resolution captured. In the selected K=8 path, bins with no USV or lever events split into low-, intermediate-, and near-total-freezing configurations, with mean freezing of 9.2%, 48.5%, and 98.7%; their fitted topology favored neighboring transitions (0.042-0.188), no direct low-to-near-total jump appeared in the decoded path, and fitted probabilities in either direction fell below 3 x 10^-25. Each of the five selected fold models likewise contained a count-silent center above 80% freezing and distinct 22-kHz-, 50-kHz-, and lever-dominant centers (Figure S11). These configurations locate where prediction improved without establishing eight natural modes; the hierarchy therefore supports K=4 as a coarse functional scale, not the predictive optimum. An initial emission-family sensitivity was less favorable: standard K=4 Poisson and negative-binomial HMMs were stable across 20 restarts each, yet their decoded paths agreed only modestly with the validated Gaussian path (NMI 0.253 and 0.239; aligned bin agreement 47.1% and 48.7%). Zero-inflated variants reached the same qualitative conclusion but met only a per-bin convergence criterion (Figure S11).

A stricter channel-matched sensitivity then modeled the two USV counts with negative-binomial emissions, lever pressing with a Bernoulli emission, and freezing with a Gaussian emission, alongside a Poisson-count comparator. Under the prespecified real-data gate, 0 of 100 negative-binomial restarts and 1 of 100 Poisson restarts passed every criterion, so no count-native model was selected. A no-refit diagnostic traced the negative-binomial profile warnings to initialization, with negligible endpoint emission effects, but that post hoc finding did not relax the prespecified fit-validity result. For exploratory interpretation only, a predefined emission-only mapping was applied to 50 stable negative-binomial endpoints and the sole valid Poisson endpoint; neither family yielded a complete four-label configuration (0/50 and 0/1), the strict Goal-directed separation gate failing in each. Planned cross-family analyses that required all four labels were therefore not estimable. Together these analyses fix the boundary precisely: the validated Gaussian path is a theory-guided reference, not an emission-family-invariant ontology.

Discussion

The principal result is a theory-guided temporal framework for tracing functional change after danger interrupts exploration and learned goal pursuit. Across three related experiments in male rats, a boundary-aware K=4 Gaussian HMM separated low-output, 50-kHz-dominant exploratory, defensive, and instrumental patterns and assigned every rat-session a path through them. Within that representation, the patterns meet three limited empirical criteria for candidate functional modes: each combines several channels, persists across contiguous windows, and recurs across animals and the fixed study partition. Conditioning redistributed the repertoire from Exploratory toward Fear/Danger; extinction later reduced Fear/Danger without necessarily restoring exploration or task performance; delay and context revealed exploratory-rich and relapse-related trajectories. Selected higher-K fits subdivided these regions, supporting their use as a coarse scale. Goal-to-Fear depletion survived both transition nulls, and the Fear/Danger-boli association survived adjustment for session. Count-native sensitivities bound the ontology claim. The account therefore establishes no natural kinds and no universal recovery sequence; it offers a reproducible way to distinguish forms of functional change that fear reduction alone cannot resolve.

The argument starts from one empirical premise: defense is not a single response dimension. The Fear Threshold Model predicts that objective defensive reactions and candidate indicators of subjective fear can cross different thresholds (11), and experiments bore this out, freezing and 22-kHz calls diverging across manipulations of threat, timing, and individual difference (12). This is why a change in any one defensive response cannot capture the animal’s entire fear-related condition, and it says nothing about what follows a decline in defense: quiescence, exploratory engagement, or restored instrumental action.

The present extension of TRP carries the argument from the danger-facing half to the whole fear-goal repertoire. It complements rather than replaces established accounts of threat-imminence-dependent defense, risk assessment, and foraging under predation risk (19–22): those traditions explain how danger and opportunity shape action, while TRP contributes a common reference-dependent vocabulary relating danger, the current condition, and goals across the measured repertoire. The four Gaussian HMM signatures answer to those regions closely enough to be useful, short of a one-to-one proof or neural identification. Quiescent is an observed low-output pattern, not the status quo itself; Exploratory is a context-sensitive vocal pattern distinct from instrumental action; and no label measures feeling directly. The global occupancy proportions, too, are properties of the aggregated experimental timeline, not a biological prescription. The contribution is a two-timescale account: seconds-scale persistence and switching within sessions, and experience-dependent redistribution across conditioning, extinction, delay, context, and relapse. The second timescale indexes learning history without making memory itself a latent HMM state.

At higher resolution, the selected K=8 Gaussian fit offers a post hoc bridge to the Fear Threshold Model. Its highest-freezing configuration was count-silent (98.7% mean freezing), whereas the 22-kHz-dominant Fear/Danger configuration averaged about 65%, a divergence compatible with response-specific thresholds within the danger-facing TRP region rather than a monotone freezing-to-calling ladder. This neither establishes an eight-stage sequence nor treats the near-zero fitted direct transitions among Gaussian freezing strata as threshold evidence: K=7 remained in the one-SE set, valid restarts yielded no unique K=8 path, mixed vocal and lever configurations appeared, and count-native models reorganized the partition. The threshold-refined account is therefore a hypothesis for prospective graded-threat and model-transfer tests of whether reactions are recruited in a reproducible order, and of how any resulting microstates relate to exploration and goal pursuit (Figure S11).

The transition paths sharpen the arbitration account once the null is specified. Goal-to-Fear transitions were depleted under both the local occupancy-preserving and exact run-label nulls; Fear-to-Goal depletion depended on the null, and Quiescent enrichment held only under the locally occupancy-preserving comparison. One plausible reading is that low-output intervals can accompany reorientation as incompatible behavioral policies are reselected, consistent with functional and circuit accounts of threat imminence and defensive response selection (33–36). Yet the exact-null result forbids treating that route as universal; local permutations cannot remove every phase-dependent influence, five-second bins cannot resolve sub-second ordering, and no transition marks a dedicated neural gate. Higher-resolution behavioral and neural measurements will be needed to test whether targeted interventions selectively change entry, exit, or persistence.

These interpretations rest on numerical credibility and explicit model boundaries. Historical full-covariance fits attempted to estimate every residual pairwise relation among the channels; none of 100 K=4 or 100 K=8 restarts passed the strict checks in the documented modern environment. A diagonal model was stable, but an initially selected 1e-6 floor exploited density at exact zeros, so discrete-aware validation instead selected 0.01. The K sweep shows that predictive performance continues beyond K=4, while selected higher-K fits preserve and subdivide its four regions and all-restart aggregate nesting remains high. Count-native alternatives set the sharper boundary: in the prespecified channel-matched real-data sensitivity, no negative-binomial model and only one Poisson restart passed every criterion, so no model was selected. Post hoc diagnostics localized the negative-binomial warnings to initialization, but the exploratory endpoints still failed the prespecified complete semantic mapping. The corrected analysis therefore supports a theory-guided Gaussian coarse scale and rejects stronger claims of global optimality, natural state count, emission-family invariance, or cross-family confirmation of all label-dependent outcomes.

These distinctions define the translational value. In models of fear- and trauma-related disorders, a fall in defensive responding cannot certify functional recovery, which is why the field increasingly looks past freezing and fear expression (15, 32). Across related but non-comparable cohorts, the experiments sample parts of a recognizable sequence: an animal explores and learns an effective action; aversive experience reorganizes that repertoire around danger; and extinction, time, and context then reveal whether defense subsides into quiescence, exploratory re-engagement, renewed goal pursuit, or relapse. The present data show why the distinction matters. Late in initial extinction, Fear/Danger was low but Quiescent dominated; other contexts supported exploratory-rich behavior; and pooled Study 2 renewal data showed renewed defensive occupancy. Multimodal measurement supplies the evidence, TRP the functional landmarks, and the HMM the individual rat-session paths and transition language. The model is not a wellbeing assay, and the study makes no claim about clinical recovery in humans. What it offers is a preclinical computational basis for asking whether an intervention restores the wider repertoire rather than merely suppressing one symptom. Its next test is prospective: whether individual mode trajectories predict durable functional restoration or relapse better than single-response measures.

Several limits remain. All animals were male, all experiments came from a single laboratory, and Study 1b reused Study 1a subjects, so external generality is untested. The two-cohort analysis is fixed-partition recurrence, not independent model transfer. K=4 was theory-guided and is not the predictive optimum, and count-native models reorganized the partition, so cross-family tests requiring all four labels were not estimable. Acoustic subtype labels for 50-kHz calls were not retained, which blocks subtype-level reinterpretation of Exploratory. The pooled occupancies are design-weighted summaries, and the modeled dataset offers no common pre-threat baseline from which a normal or recovered target occupancy could be estimated for each animal. Although every rat-session has a decoded path, the present outcome analyses are largely group-level and do not validate individualized prognosis. Within-session analyses are observational, switching claims depend on the null, and fecal boli were available only for Study 2. Finally, the labels describe behavioral configurations, not subjective content, neural mechanisms, or clinical states.

These limits also set the agenda. A prospective recovery design should measure each animal before danger learning, during graded threat, through extinction or intervention, and across delayed and contextual relapse tests, then define recovery relative to that animal’s own pre-threat exploratory and task-function profile rather than a pooled occupancy target. Independent datasets should test whether higher-resolution states nest within aligned functional regions and whether all four regions can be identified under channel-matched emission models. Input-output or nonstationary state-space models could represent experimental inputs and individual learning histories directly. If the modes reflect separable control policies, then neural, pharmacological, or behavioral interventions should shift specific entry, exit, or persistence probabilities, and those shifts should predict later function. Direct exploration measures, active avoidance, social interaction, reward devaluation, female animals, and other species may split, merge, or extend the vocabulary. Comparative tests should map species-specific emissions onto homologous functional regions and transition motifs rather than demand identical behaviors or labels. The released model remains a transparent benchmark for these tests, not a closed taxonomy; its present value is to make functional recovery testable as a historical trajectory across the whole measured repertoire.

Materials and Methods

The methods follow the same inferential order as the argument. They define the experimental unit and measurements, reconstruct the observation matrix, select and fit the theory-guided Gaussian K=4 model, decode subject-session paths, test restart and cross-study stability, evaluate the state-count hierarchy and count-native sensitivity, analyze history-dependent occupancy and transitions, and audit the provenance of the replacement model. Pooled occupancy is reported only as a design-weighted description; functional inference rests on session-, phase-, and subject-aware comparisons.

Study design and reporting

This retrospective computational study analyzed three related experiments. The rat was the experimental unit, and the 5-s bins were repeated observations nested within subject-sessions. No prospective registration or power analysis was performed. Study 1a randomly assigned 12 rats to shock and control groups (n=6 each); Study 2 balanced 18 rats between extinction and no-extinction groups by open-field behavior (n=9 each before attrition). One no-extinction rat was removed after abnormal stress behavior and refusal to eat, leaving 17 analyzed Study 2 rats; no other animal-level exclusion is recorded. Masking during collection, annotation, or analysis was not documented, and group labels were never HMM features. Reporting follows ARRIVE 2.0 and STRANGE retrospectively (38, 39).

Animals, housing, sex, and ethics

Subjects were adult male Sprague-Dawley rats (RRID:RGD_70508) from Beijing HFK Bioscience (SCXK (Beijing) 2019-0008). Study 1a used 12 rats weighing 300-350 g and reported as approximately 6-8 months old; Study 1b retested these rats approximately five months later. Study 2 began with 18 rats weighing 350-400 g and reported as approximately 12 months old; 17 contributed analyzed data. Animals were group-housed three per standard cage on a reversed 12-h cycle (lights on 20:00-08:00), with rationed food and freely available water. Only males were used, so sex was not tested. The Fujian Normal University animal ethics committee approved all procedures (IACUC-20240070).

Apparatus, training, and experimental phases

Experiments used modular Coulbourn chambers fitted with grid floors, speakers, chamber lights, food dispensers, and retractable levers, and the contexts differed visually, tactilely, and olfactorily. Rats learned to press for 45-mg pellets on a VI60 schedule that remained in force during fear and test sessions. Study 1a paired 9-kHz tone-shock conditioning with context, cue, and 7- and 5-kHz generalization tests; Study 1b retested the same rats after approximately five months; and Study 2 combined conditioning, retrieval-extinction or time-matched exposure, retention, renewal, three further extinction days, and a second retention test. Table S16 reproduces the full chamber contexts, training sequence, stimulus schedules, group handling, and phase timing, so that the protocol can be evaluated without recourse to an inaccessible source.

Behavioral acquisition

Freezing was recorded with FreezeFrame 5 (RRID:SCR_014429) at 15 frames/s as the absence of movement except respiration, and summarized as the percentage of each 5-s bin frozen. USVs were captured at 384 kHz with ultrasonic microphones, detected in DeepSqueak (40), and manually curated by two trained assessors, with disagreements settled by a third expert. Lever events were exported from Graphic State 4, and fecal boli were counted from the chamber tray after each session, never entering the model as a 5-s emission. Table S17 reports the acquisition hardware, the frequency and duration criteria, the spectrogram-correction rules, the assessor procedure, and the model-ready zero/missing semantics.

Raw-to-model reconstruction

Each acquisition stream was reconstructed separately and aligned to an integer 5-s grid, with duplicate processed and raw USV sources resolved by documented source-selection rules. Each entry into the Graphic State ‘rat press’ state counted as one lever press. Within a recorded session, the absence of an event meant zero, whereas an unrecorded modality remained missing, and a freezing value of 0% was a real observation rather than a gap. The complete-case primary model kept only those bins in which all four channels were observed; no missing emission was imputed as zero. Subject-sessions were treated as independent sequences, and no transition crossed a session boundary.

Preprocessing and theory-guided model

The four channels entered the model in a fixed order: 22-kHz count, 50-kHz count, lever-press count, and freezing percentage. A log1p transformation compressed the long right tail of the count channels while preserving their zeros, freezing remained on its percentage scale, and z-scoring then placed all four on comparable numerical scales. The diagonal HMM estimated a separate within-state variance for each channel but no residual within-state covariance; cross-channel dependence was carried instead by shared state membership and temporal transitions. A TRP-derived functional hypothesis fixed K=4, in advance of numerical validation, as the functional scale for the Fear/Danger, Quiescent, Exploratory, and Goal-directed signatures, so that validation selected the variance floor rather than the number of states.

Discrete-aware variance-floor selection

Model selection was conducted within the fixed K=4 scale. Five subject-grouped folds compared diagonal Gaussian HMMs with variance floors 1e-6, 1e-4, 1e-3, and 1e-2. Scaling was estimated on training subjects only, and each candidate-fold combination used 20 random restarts. Held-out count c was scored as probability mass over its integer interval on the transformed scale. For c>0, the bounds were log1p(c-0.5) and log1p(c+0.5); for c=0, they were negative infinity and log1p(0.5). Bounds were then converted with the training scaler. Freezing retained Gaussian density. Initial and transition probabilities combined these emission scores through the HMM forward recursion. Candidate means and standard errors were computed across folds. The 0.01 floor ranked first and was selected; candidates within one standard error were treated as not clearly distinguishable at that resolution.

Full-data fitting and strict numerical validity

The selected diagonal K=4 specification was then fit to all 47,004 bins from 213 sequences. Because HMM optimization can settle at different local solutions, 100 random seeds were run with 200 maximum iterations and tolerance 1e-4, reapplying the variance floor after each maximization step. A restart counted as valid only if its likelihood history was finite and non-decreasing within numerical tolerance, optimization reached the stopping tolerance, and every variance was finite and positive. Eighty-eight restarts passed; seed 9 had the highest valid likelihood and was serialized with its preprocessing parameters. Viterbi decoding (41) then assigned the most likely complete state path, and state identities were aligned by emission signature using linear assignment (42), never by arbitrary raw state number.

Stability and cross-study recurrence

Robustness was assessed at three levels: restart, variance floor, and study. Restart partitions were compared by normalized and adjusted mutual information (43); NMI equals 1 for identical partitions after label alignment and falls toward 0 as shared assignment structure decreases, while adjusted mutual information corrects for chance. Score basins were distinct converged likelihood levels, and variance-floor sensitivity used aligned fold-wise NMI. For cross-study recurrence, empirical means within each decoded state and study were standardized with the pooled preprocessing parameters and compared with the pooled fitted signature by cosine similarity. Study-level occupancy uncertainty used 10,000 subject-bootstrap samples. Because Studies 1a and 1b reused subjects, an additional fixed-partition sensitivity pooled them as one cohort and compared that cohort with the non-overlapping Study 2 cohort in both directions. It held the all-data Gaussian representation and scaling fixed, classified each state by nearest signature, and repeated subject-bootstrap recognition 10,000 times per direction. This measures recurrence of fixed signatures, not refitted model transfer.

State-count hierarchy and count-native sensitivity

The state-count sweep reused the exact released subject folds, train-only scalers, 0.01 variance floor, and discrete-aware forward score for K=2-8. Each fold and K used 20 restarts. For full-data hierarchy analysis, K=5-8 used 100 seeds with the same strict validity criteria. Each higher-K state was mapped to the K=4 state containing the largest share of its decoded bins. Weighted nesting purity is the fraction of all bins in those dominant mappings; unweighted mean state purity gives each higher-K state equal weight. For the selected fit at each K, a state-level overlap ledger records the dominant K=4 mapping, its purity, and whether all four K=4 labels remain represented. All-restart summaries report aggregate weighted and unweighted purities and do not extend the selected-fit per-state no-merge conclusion to every restart. After inspection of the sweep, a post hoc descriptive audit examined raw empirical emissions and session-boundary-aware transitions for the selected K=8 path, agreement across its strictly valid full-data restarts, and raw-scale emission centers from the five selected fold models. These summaries did not alter K selection, the released K=4 model, or its labels. The initial count-native sensitivity modeled the three raw count channels with Poisson or negative-binomial emissions and freezing with a Gaussian emission, using 20 restarts per family; zero-inflated variants used five 1,000-iteration restarts. A subsequent prespecified channel-matched sensitivity modeled the two USV channels with inclusive NB2 or Poisson emissions, lever pressing with Bernoulli emissions, and freezing with Gaussian emissions. One hundred real-data restarts per family were evaluated under complete-history criteria covering likelihood monotonicity and convergence, one-step parameter drift, parameter validity, accepted Q-block updates, and cumulative warnings. The formal result was NB2 0/100 and Poisson 1/100, with no selected model. Adaptive no-refit profiling was explicitly post hoc and did not relax that gate. A predefined emission-only semantic mapping then enumerated all 24 state-label permutations for 50 stable NB2 endpoints and the sole valid Poisson endpoint. It required separated defensive, lever-dominant, 50-kHz-dominant, and lowest-event configurations; complete passes were 0/50 and 0/1 because the Goal-directed separation rule failed. Consequently, cross-family label-dependent downstream tests were classified as not estimable (Table S18). Count-native analyses test representation sensitivity and do not replace the prespecified Gaussian model.

Temporal, dwell, and transition analyses

Temporal analyses used the aligned decoded path and preserved subject-session boundaries throughout. Dwell duration was the number of contiguous bins in a state multiplied by five seconds. Session transition matrices counted adjacent decoded states and were row-normalized. For the primary routing null, each subject-session was split at the bin level into 1, 2, 5, or 10 equal contiguous strata; within every stratum, contiguous state episodes were permuted, equal neighboring labels were collapsed, and exit transitions were recounted. Two thousand permutations per stratification used seed base 20260807; the ten-stratum result was primary. An exact conditional run-label sensitivity fixed each session’s number of runs and multiset of run labels and sampled label orders uniformly subject to unequal adjacent labels. It did not preserve run lengths or bin occupancy. Dynamic programming gave exact expected transition counts, and 10,000 draws with seed 10105202 gave plus-one two-sided Monte Carlo p values. Separately, subject-sessions were divided into ten segments for occupancy trajectories, and subject-specific Spearman correlations between segment order and occupancy were averaged by session and state. Two-sided sign-flip p values used 20,000 assignments, with Benjamini-Hochberg correction across the reported session-state tests (44).

Autonomic validation

External autonomic validation used 130 Study 2 sessions from 17 rats with complete boli, freezing, and mode-occupancy data. Negative-binomial generalized estimating equations used alpha=1 and an exchangeable within-rat working correlation (45). Because 17 clusters are few for ordinary sandwich inference, the Mancl-DeRouen bias-reduced covariance was primary and conventional cluster-robust covariance was retained as sensitivity analysis (37). The occupancy model carried Fear/Danger, Exploratory, and Goal-directed proportions, with Quiescent omitted as the compositional reference. An additional sensitivity added indicators for the eight session names. Two broader codings grouped sessions into conditioning, extinction, and test, or into conditioning, extinction, retrieval, and later test. A separate model used mean freezing percentage. Coefficients are log rate ratios, and Figure 4C rescales the unadjusted estimates to a ten-percentage-point change. The occupancy and freezing models were not compared directly.

Historical-model replacement audit

A separate provenance audit addressed the dependence of sequence results on the fitted path. No historical serialized full-covariance model or complete decoded path was recovered. In the documented modern environment, 0 of 100 K=4 and 0 of 100 K=8 full-covariance restarts passed the strict criteria; the library convergence flag alone was insufficient. A previous boundary-prone diagonal K=8 fit contained a 16-bin state and could not support coarse-graining, so the historical K=8-to-four claim remains excluded. The K=2-8 sweep used the validated 0.01 floor and exact discrete-aware scoring and is reported as a sensitivity analysis. The channel-matched count-native chain is likewise separated into its formal fitting-failure result, a post hoc initialization-warning diagnostic, and a predefined semantic audit; no later diagnostic retroactively selected a model or converted an unestimable cross-family claim into evidence. Table S18 distinguishes these analyses and their evidentiary roles, and Table S19 records their canonical archives and hashes.

Use of artificial intelligence tools

The theoretical framework and hypotheses, experimental design, and primary data collection were developed and conducted entirely by the authors without AI assistance. After primary data collection, AI-assisted systems, including OpenAI Codex, Edison Scientific’s Kosmos, and Anthropic’s Claude Code, were used under author direction for data-structure reconciliation, execution and verification of specified analyses, code and figure checking, and language editing and document assembly. AI-generated outputs were not treated as evidence on their own. Numerical claims were checked against versioned data, executable code, and provenance records; unsupported analyses were excluded or reported as limitations. The authors retained final authority over all scientific questions, model specifications, analytical decisions, and interpretations, reviewed the outputs, and take full responsibility for the work.

Data availability

Analysis-ready data and the project record are available through the persistent concept DOI https://doi.org/10.5281/zenodo.20551230, and original raw behavioral exports are available at https://doi.org/10.5281/zenodo.20551550. All analysis and figure-generation code, software environments, executable verification, the serialized model, preprocessing parameters, all 47,004 decoded rows, selection and stability evidence, source-linked downstream tables, and checksums are archived in the validated v2.1.0 release at https://doi.org/10.5281/zenodo.21883766. Table S19 lists artifact-specific provenance, integrity hashes, and verification roles. Data use CC BY 4.0 and code uses the MIT License. Supplementary Information contains Tables S1-S19 and Figures S1-S11.

Acknowledgements

We thank members of the Laboratory for Learning and Behavioral Sciences at the School of Psychology, Fujian Normal University, for assistance with behavioral data collection, ultrasonic vocalization annotation, and data curation.

Additional information

Funding

This work was supported by the Natural Science Foundation of Fujian Province (Project No. 2025J01642), awarded to Bin Yin. The funder had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Author contributions

BY: Conceptualization, Methodology, Formal Analysis, Visualization, Supervision, Project Administration, Resources, Writing - Original Draft, Writing - Review & Editing. JR: Investigation, Methodology, Data Curation, Formal Analysis, Writing - Original Draft. XTW: Conceptualization, Writing - Review & Editing.

Funding

福建省科技厅 | Natural Science Foundation of Fujian Province (Fujian Natural Science Foundation) (2025J01642)

• Bin Yin

Additional files

Supplementary Information.