Recognize danger, but disregard it: a mixed-methods analysis of the knowledge-action dissociation in volunteer firefighter hazard recognition

Volunteer firefighters must assess scene hazards before initiating rescues; however, line-of-duty fatality data consistently identify failures of self-protection as a proximate cause of preventable casualties. Whether mnemonic-based training, specifically the GAMS schema (Gefahr erkennen–Absperren–Menschenrettung–Spezialkräfte; identify hazard–cordon off the scene-rescue-request specialist units), translates into reliable behavioral activation under operational pressure has not been empirically tested. Using a convergent mixed-methods design, 21 active volunteer firefighters from a single department in northern Germany were randomly assigned to three groups and exposed to four sequential hazard scenarios: carbon monoxide poisoning, an unsecured aggressive dog, electrical shock in a wet environment, and a flooded basement with a missing person; the fourth was delivered as a theoretical briefing and excluded from quantitative scoring. Quantitative data were collected via standardized, structured observations across the three enacted scenarios, scored on a four-point severity scale (0–3). The qualitative data consisted of 130 transcribed speech sequences, which were analyzed using structured qualitative content analysis following Mayring's methodology, with all sequences independently double-coded. Six deductively derived main categories were differentiated into 30 inductively derived subcategories, each of which was operationalized with explicit definitions, coding rules, and anchor examples. Critical-action errors formed the single largest error category and were elevated across all enacted scenarios, with no significant between-group differences [Kruskal–Wallis H(2) = 0.633, p =.729]. Qualitative analysis identified a systematic rescue impulse that overrode hazard assessment routines at the perception-to-action transition, despite participants demonstrating intact declarative knowledge of the correct procedures during post-scenario debriefings. This knowledge–action dissociation was invariant across hazard types. The GAMS mnemonic remains stored as declarative knowledge rather than functioning as a procedural-action trigger. Mnemonic-based didactic instruction alone is insufficient and must be supplemented by repeated scenario-based rehearsal to proceduralise self-protective behaviors.

Firefighter safety rests on a deceptively simple principle: secure yourself before rescuing others. This principle is codified in the operational doctrine worldwide. However, line-of-duty fatality reports consistently identify failures of self-protection as the proximate cause of preventable casualties (1, 2). The tension between the imperative to rescue and the obligation to self-protect constitutes a goal conflict that shapes every emergency response, from structural fires to technical rescue operations (3). Understanding how this conflict manifests in real-time decision-making and why trained personnel sometimes act contrary to their safety knowledge remains a central challenge for fire service research and training design.

Research on fireground decision-making has been dominated by Klein's Recognition-Primed Decision (RPD) model, which posits that experienced commanders identify situational patterns and mentally simulate a single course of action rather than comparing alternatives (4, 5). This model explains the speed and efficacy of expert decisions under time pressure but also implies critical vulnerability. When pattern libraries are underdeveloped, as in personnel who rarely encounter non-fire hazards, the first recognized pattern may default to a deeply ingrained rescue script rather than a hazard assessment protocol (6). The situation awareness theory further specifies that effective tactical action requires perceiving environmental cues, comprehending their operational significance, and projecting future states (7, 8). Failures at any of these levels, particularly at the perception stage, truncate the decision cycle before self-protective reasoning can be engaged (9).

A growing body of evidence documents the dissociation between declarative knowledge and procedural performance in safety-critical domains. Emergency responders frequently demonstrate theoretical mastery of protocols but fail to enact them under operational stress (10). Neurocognitive research attributes this gap to stress-induced shifts from flexible, hippocampus-dependent processing toward habitual, striatum-mediated responding (11), which preserves well-rehearsed motor routines but degrades the activation of declaratively learned rules that have never been proceduralized through practice (12). Mnemonics and checklists can partially bridge this gap by cueing critical assessment steps. These have been shown to increase the use of guideline-compliant measures by 25%–29% depending on the source of the literature (13–15). However, whether mnemonic-based hazard schemas function as reliable behavioral triggers rather than merely as recall aids for declarative content has not been empirically tested in fire services.

In German-speaking fire services, the GAMS rule is the standard mnemonic for initial hazard management. It is systematically taught during basic training in accordance with the official fire service regulations (16). Originating from chemical, biological, radiological, and nuclear (CBRN) hazard response, the acronym GAMS denotes hazard identification (Gefahr erkennen), scene control (Absperren), life-saving measures (Menschenrettung), and the request for specialized emergency response units (Spezialkräfte anfordern) (16, 17). Despite its ubiquity, no study has examined whether GAMS is activated as a procedural action trigger under simulated or real operational conditions. This gap is particularly consequential for Germany's volunteer fire service, which comprises over 90% of the nation's approximately 1.38 million firefighters (18). In the largest empirical survey of volunteer firefighters in a German federal state, volunteers train substantially less than career professionals, face fragmented scheduling, and encounter non-fire technical assistance scenarios infrequently (19). On physical, motoric and cardiovascular measures, however, volunteers are broadly comparable to career firefighters (20), locating the disparity in training exposure and oversight rather than in physical capacity. This creates conditions under which mnemonic schemas may remain declarative knowledge and never become proceduralized. Existing evaluation instruments, such as the FIRE questionnaire family (21), capture summative measures of training quality but do not assess scenario-specific behavioral errors or the cognitive mechanisms underlying performance failure.

Addressing these gaps requires an approach that integrates quantitative performance measurements with qualitative exploration of the underlying cognitive and motivational processes. Mixed-methods designs have been shown to yield more complete and actionable assessments of emergency training outcomes by linking measurable behavioral changes to contextual explanations of why certain patterns emerge (22). Qualitative content analysis, in particular, can reveal recurrent behavioral patterns, decision rationales, and group dynamics that standardized observation alone cannot capture (23).

Throughout this article, we use the term rescue impulse to denote an automatic, pre-decisional action tendency in which the perception of a person in distress directly activates an intervention response (i.e., person in distress → intervene immediately) before hazard cues are consciously evaluated or self-protection rules are retrieved. Crucially, this differs from a deliberate goal conflict in which competing objectives (mission vs. safety) are consciously weighed and a trade-off is selected (3). The rescue impulse operates at the perception-to-action transition and preempts the deliberative stage entirely. We conceptualize it as a strongly consolidated “default option” within Klein's extended RPD framework (24), reinforced by firefither role identity and social expectations, that becomes behaviorally dominant under time pressure and acute stress. In the present study, this construct was operationalized as the qualitative category HK3.1 (rescue impulse overrides hazard assessment) (Box 1).

Definition: Throughout this article, the term rescue impulse denotes an automatic, pre-decisional action tendency in which the perception of a person in distress directly activates an intervention response (person in distress → intervene immediately) before hazard cues are consciously evaluated or self-protection rules are retrieved from memory.

Key distinction: The rescue impulse is not a deliberate goal conflict. Whereas goal conflict implies that competing objectives (mission vs. self-protection) are consciously weighed and a trade-off is selected (3), the rescue impulse operates at the perception-to-action transition and pre-empts the deliberative stage entirely.

Theoretical grounding: It is conceptualized as a strongly consolidated “default option” within Klein’s extended Recognition-Primed Decision framework (24), reinforced by firefighter role identity and social expectations, and it becomes behaviorally dominant under time pressure and acute stress.

Operationalization: In the present study, the construct was operationalized as the qualitative category HK3.1 (rescue impulse overrides hazard assessment) and triangulated with the invariant self-protection errors recorded by the quantitative H-CRIT instrument.

Therefore, the present study employed a convergent mixed-methods evaluation to investigate hazard recognition and self-protection behavior in volunteer firefighters confronted with nonfire emergency scenarios. Specifically, we asked the following questions:

To what extent do volunteer firefighters activate trained hazard-assessment protocols when confronted with realistic non-fire emergency scenarios, and how does the interplay between rescue motivation and self-protection knowledge shape their behavioral responses?

By combining structured behavioral observation with qualitative content analysis of supervised debriefings, this study provides the first empirical data on GAMS activation under simulated operational conditions and thus contributes to clarifying whether mnemonic-based safety schemes actually act as behavioral triggers under operational pressure or remain merely declarative knowledge to be recalled. This study contributes to the theory of the rescue-impulse–self-protection conflict in volunteer fire service contexts.

Design

We employed a convergent mixed-methods design (22) that integrated quantitative structured observations with qualitative content analysis. The quantitative strand provided standardized error scoring across predefined behavioral items, while the qualitative approach enabled insights into the participants' cognitive and reflective processes underlying their observed performance. Triangulation followed a convergent model in which both data strands were analyzed independently and integrated during interpretation (see Results: Triangulation).

Participants and setting

The participants were 21 active volunteer firefighters (all operational status) from a single volunteer fire department in northern Germany, randomly assigned to three groups of seven (N = 21; groups G1–G3). Group size was fixed by the number of available participants and the crew strength required to enact each scenario as a realistic first-arriving team, and the three-group structure enabled within-study replication across three independent teams. All 21 participants were observed quantitatively (TN1–TN21); of these, 16 contributed verbal statements suitable for sequence-level coding (TN1–TN16), the remaining five producing no separately codeable utterances. Five members of the training staff served as scenario facilitators, observers, and role players and were excluded from the data collection. Participants held qualifications ranging from basic training as crew members (Truppmann I/II) to advanced certifications such as crew leader or breathing apparatus wearer (BA wearer) (Truppführer, Atemschutzgeräteträger). Detailed demographic data (age and years of service) were not collected as the evaluation was embedded in a regular special training exercise. All data were anonymized (TN1–TN21; G1–G3).

The study was conducted during a scheduled special training evening (January 23, 2026) at the department's fire station, following a station rotation design. The present analysis focuses exclusively on Station 3 (hazards at the emergency scene), comprising four sequential scenarios presented as situational assessments for first-arriving units.

Scenarios

The four scenarios were selected on two grounds. First, they represent recurrent non-fire hazard types with an elevated risk profile for the department's operational area, derived from the local incident statistics and mapped to the hazard groups of the German fire-service regulation FwDV 500 (16), so that the exercise probed situations first-arriving volunteer crews are realistically likely to encounter rather than abstract textbook cases. Second, each scenario deliberately targets the interface between a fire-service operation and an acute, life-threatening emergency to a person, the constellation in which the impulse to rescue is expected to override self-protection, so that the scenarios function as ecologically valid probes of the mechanism under investigation. The four hazards additionally vary the sensory modality of the warning cues (visual, auditory, combined, and static), allowing a test of whether protocol activation is cue dependent. Each group completed four scenarios in a fixed order, with approximately five minutes per scenario, including immediate debriefing. Table 1 summarizes the characteristics of the scenarios. Scenarios 3a–3c were enacted with physical props and role players, and Scenario 3d was presented as a photographic situation assessment without physical enactment.

Quantitative strand—structured observation

Behavioral performance was assessed using a standardized observation instrument comprising 47 items across the four scenarios (3a: 10; 3b: 12; 3c: 12; 3d: 13); because Scenario 3d was delivered as a theoretical briefing, only the 34 items of the three enacted scenarios (3a–3c) entered the quantitative analysis. Each item was scored on a four-point error scale: 0 (correctly performed), 1 (minor deviation), 2 (significant error), and 3 (critical safety violation). Items were assigned to 13 error types organized along a perception, cognition, or action taxonomy: Perception (W), Cognition (K), Action (H), Critical Action (H-CRIT), Resource (R), Securing (S), Communication (COM), Medical (M), Specialist Knowledge (FW), Documentation (D), Technical (T), Hygiene (HY), and Reconnaissance (E). H-CRIT items were flagged a priori as the primary outcome indicators. Overall group performance (summed severity-weighted error points across the three enacted scenarios; maximum 76) was classified as passed (0–33), re-training required (34–65), or critical (≥66); these bands rescale the original per-scenario thresholds (passed 0–11, re-training 12–21, critical ≥22) by the factor applied when the dataset was recomputed.

Qualitative strand—content analysis

Semi-structured group discussions were conducted immediately after each scenario and during cross-scenario debriefings. To enable systematic data collection, the research team developed a standardized discussion guide iteratively based on the relevant literature and specific training objectives. The guide covered four thematic areas: (a) hazard perception and warning cue interpretation, (b) action rationale and decision-making processes, (c) knowledge gaps and protocol awareness, and (d) self-assessment and reflective learning. All verbal data were recorded and transcribed verbatim into 130 discrete speech sequences that served as the units of analysis. Transcripts were checked for completeness and anonymized by assigning pseudonymous participant codes (TN1–TN16). The dataset was stored on encrypted university servers and managed in full compliance with data protection regulations. The analysis followed the structuring variant of qualitative content analysis (23). The category system was developed in two phases. In the first deductive phase, six main categories were derived from the theoretical framework and study objectives: HK1 (Hazard Recognition), HK2 (Specialist Knowledge), HK3 (Action Patterns), HK4 (Reflection/Insight), HK5 (Need for Improvement), and HK6 (Structural Factors). In the second inductive phase, these main categories were differentiated into 30 subcategories through iterative coding of the material (e.g., HK3.1 Rescue Impulse Override, HK4.2 Retrospective Correct Identification, HK6.3 Absence of Autonomous Assessment Routine). Each subcategory was operationalized with an explicit definition, coding rule, and anchor example to ensure unambiguous assignment. Decision rules were formulated for systematically ambiguous cases (e.g., retrospective hazard identification was assigned to HK4 rather than HK1 when the statement reflected post hoc insight rather than in-situ perception). The coding process began with the open coding of an initial subset of sequences (n = 30), which was used to refine the category system and establish decision rules. The finalized coding scheme was then applied to the complete data set. The finalized coding scheme was applied to the complete dataset by a primary coder. To make the knowledge–action relationship explicit, we added a second structuring dimension to the analysis. The material was cross-classified along two axes derived from the existing category system: (i) declarative protocol knowledge (present vs. absent), evidenced by retrospective GAMS reproduction and correct hazard identification during debriefing (HK2, HK4.2), and (ii) enacted self-protection (protected vs. unprotected entry), operationalized by the critical-action (H-CRIT) errors. Scenario 3d was excluded, as it was assessed theoretically and lacked action pressure. To establish intercoder reliability, a second trained coder independently coded all 130 speech sequences at the main-category level (HK1–HK6), blind to the primary coder's assignments. Agreement was almost perfect: raw agreement was 96.2% (125/130) and Cohen's κ = 0.95 [95% CI [0.92, 0.99]; bootstrap 95% CI [0.91, 0.99]], exceeding the pre-specified threshold of κ ≥.70. Category-specific reliability ranged from κ = 0.83 (HK1, hazard recognition) to κ = 1.00 (HK5, HK6). The five disagreements all concerned the boundary between situational hazard recognition (HK1) and the adjacent knowledge, action, and reflection categories; each was resolved by consensus, with the primary coder's assignment retained as final, prompting refinement of the HK1/HK4 and HK2/HK3 delimitation rules.

Data integration

Triangulation followed a three-level model: (1) error-type triangulation, linking quantitative error categories to qualitative process categories to explain deficits mechanistically; (2) assessment triangulation, contextualizing the quantitative classification with qualitative reflection capacity and change readiness; and (3) resource triangulation, identifying competencies visible only in the qualitative strand as inputs for training redesign. The convergent, complementary, and divergent findings were systematically documented.

Research rigor and quality assurance

Several strategies were employed to ensure methodological rigor. The convergent mixed-methods design allowed triangulation that neither data strand could achieve alone: quantitative invariance of H-CRIT errors gained explanatory depth through the qualitatively identified rescue-impulse mechanism, while the qualitative resource dimension (functioning teamwork, specialist knowledge, and reflection capacity) would have been entirely invisible to quantitative scoring. Triangulation of perspectives was further supported by an interdisciplinary research team comprising experts in emergency medicine, fire service operations, public health and qualitative research methods. The use of standardized observation instruments with a priori operationalized items reduced observer bias, while the structured coding scheme with explicit decision rules ensured replicable categorization of the qualitative data. Additionally, ongoing reflexive documentation by the research team and transparent audit trails for all analytical decisions strengthened the credibility and resonance of the interpretations.

The quality of the qualitative strand was systematically anchored in the seven core criteria for qualitative research formulated by Steinke (25, 26). Intersubjective comprehensibility was ensured through complete documentation of the research process, including the discussion guide, transcription protocol, codebook with operational definitions, decision rules for ambiguous cases, and an audit trail of all analytical decisions. Indication of the research process, the appropriateness of methodological choices for the research question, is reflected in the use of structuring qualitative content analysis (23) under a deductive–inductive logic appropriate to a partially pre-specified analytical framework. Empirical anchoring was ensured by formulating each subcategory with at least one anchor example drawn directly from the transcripts. The limitations of the scope of the findings are explicitly addressed in the Limitations section. Coherence between the data, categories, and interpretation was repeatedly checked through team discussions during the iterative refinement of the category system. The relevance of this study derives from its practical implications for fire-service training reform. Reflected subjectivity was supported by the interdisciplinary composition of the research team (emergency medicine, fire service operations, public health, and qualitative methodology) and by ongoing reflexive memos documenting preconceptions and emerging interpretive shifts.

Ethical considerations

The study was approved by the Ethics Committee of the University of Lübeck (Ethik-Kommission der Universität zu Lübeck, Ref. No. 2023-123, approved 29.04.2025). The study was conducted in accordance with the Declaration of Helsinki. All methods were performed in accordance with relevant guidelines and regulations. The evaluation was embedded in a scheduled special training exercise (January 23, 2026) and served as a quality assurance assessment of operational training. All participants provided written informed consent to participate and be mentioned anonymously in the manuscript after receiving detailed study information. Participation was entirely voluntary, and participants could withdraw at any time without providing a reason. All data were fully anonymized by assigning pseudonymous codes (G1–G3, TN1–TN16); raw data were retained for remedial training planning and were subsequently destroyed. Audio recordings were deleted after transcription and verification. All other data were pseudonymized and securely stored in a password-protected file, with access restricted to the core research team. The results are reported in the aggregate.

A convergent mixed-methods analysis of quantitative observational data (47 standardized items across four emergency scenarios) and qualitative content analysis (130 speech sequences coded according to Mayring) revealed a systematic pattern of self-protection failure across all the participant groups.

Quantitative findings

None of the three groups (n = 7 each) completed any of the three enacted scenarios without making critical errors. Across the three practical scenarios, the groups accumulated 179 severity-weighted error points (G1 = 63, G2 = 64, G3 = 52; M = 59.67, SD = 6.66), all within the re-training-required band (34–65 of a possible 76). Scenario 3d (flooded basement) was presented as a theoretical briefing and is excluded from all quantitative analyses. Tables 2A,B summarizes the error distributions across scenarios and error types.

The critical finding was the invariance of H-CRIT errors across scenarios: all three groups entered the hazard zone without prior risk assessment in each of the three practical scenarios, irrespective of the hazard type (chemical, biological or electrical) or warning modality (visual, auditory, or combined). H-CRIT was the single largest error category (34 points; 19.0%) and occurred in every practical scenario (3a = 18, 3b = 9, 3c = 7). Together, the three leading categories, critical action (H-CRIT, 19.0%), cognition (K, 17.9%) and specialist knowledge (FW, 12.8%), accounted for 49.7% of all error points. A Kruskal–Wallis test confirmed no significant difference between the three groups [H(2) = 0.633, p =.729]; given the group-level unit of analysis, this test is exploratory and underpowered, and the descriptive distribution carries the primary evidential weight.

Between-group variation was small (group totals 52–64) and non-significant [Kruskal–Wallis H(2) = 0.633, p =.729], consistent with a structural rather than an individual deficit. The only between-group differences concerned isolated competencies: one group performed delayed de-energization in the electrical scenario (3c), one group applied stepwise hemorrhage control in the animal scenario (3b), and three participants in one group demonstrated advanced CO diagnostic knowledge (pulse oximetry limitations).

Because Scenario 3d was assessed theoretically only, no quantitative transfer effect can be inferred. Participants' reflections on this scenario are reported in the qualitative strand.

Qualitative findings

Content analysis of 130 speech sequences yielded six categories. Independent double-coding of all sequences indicated almost perfect intercoder agreement [Cohen's κ = 0.95, 95% CI (0.92, 0.99)]. Reflection/insight (HK4) dominated with 27.7% of coded sequences, followed by Specialist Knowledge (HK2, 20.0%) and Need for Improvement (HK5, 18.5%). Table 3 summarizes this distribution.

The central qualitative finding was the systematic overriding of hazard assessment by the rescue impulse (HK3.1), which was consistently observed across all practical scenarios. Participants perceived warning cues (posted signs, aggressive barking, and electrical appliances in water) but failed to evaluate them as stop signals (HK1.6). The pattern is captured in cross-scenario meta-reflection: “ It was the same at all three stations: we see someone who needs help, and we run in. Regardless of whether it is CO, a dog, or electricity ”, we always act first and think afterward (Participant in debriefing, TN1).

Qualitative analysis differentiated three distinct knowledge types with divergent training implications: (a) general factual knowledge deficit (properties of carbon monoxide, e.g., that it is colourless and odourless, and organisational responsibilities), requiring didactic input; (b) available but non-activated knowledge (GAMS schema known to all, applied by none), requiring procedural automatization; and (c) heterogeneous specialist knowledge (pulse oximetry, ventricular fibrillation), concentrated in individual participants and unexploited as a group resource. A participant exemplified type (b): “ GAMS—sure. Recognize hazards, cordon off the area, rescue, and call the specialists. I can do this in my sleep. But in the scenario I didn't think about it for a single second (Participant in debriefing, TN3).”

Participants attributed errors partly to the absence of an internalized decision-making schema for first-on-scene situations (HK6.3). In routine operations, hazard assessment is performed by the incident commander; when arriving first without this leadership structure, participants lacked an autonomous assessment routine. The GAMS schema could bridge this gap, but it was not activated despite being declaratively available.

Despite pervasive failures in self-protection, qualitative analyses have identified robust competencies. Patient care and team coordination functioned reliably after initial entry (HK3.9), sensory perception of hazard cues was intact (HK1.6), and the reflection capacity was high. No participant defended the observed errors; instead, they formulated concrete behavioral intentions (HK4.6) and constructive improvement proposals (HK5), including checklists, regular scenario training, and inter-agency cooperation.

Knowledge–action cross-classification

Cross-classifying the enacted scenarios along declarative knowledge and enacted self-protection produced an unambiguous pattern (Tables 4A,B). In all three practical scenarios and across all three groups, declarative GAMS knowledge was present, every group reproduced the schema in the debriefing, yet self-protection was not enacted. Critical-action errors (H-CRIT) were the single largest category (34 points; 19.0%) and occurred in every scenario, with no significant between-group difference [Kruskal–Wallis H(2) = 0.633, p =.729]. No case showed the trained combination (knowledge present and enacted), and no unprotected entry was attributable to a gap in GAMS knowledge itself. Differentiating the knowledge dimension further showed that only circumscribed factual content (e.g., properties of carbon monoxide) was genuinely absent, whereas the load-bearing procedural rule was present but non-activated. The deficit is thus located in activation, not acquisition, the empirical core of the knowledge–action dissociation this study set out to test.

Triangulation

Systematic triangulation revealed convergent, complementary, and divergent findings across both data strands. Table 5 summarizes this integration.

The strongest convergence concerns the core finding: quantitative maximum-error scores across all critical self-protection items are mechanistically explained by the qualitatively identified rescue impulse–self-protection conflict. The qualitative strand revealed that the deficit was not located at the knowledge level (hazards were recognized retrospectively) but at the level of action regulation under pressure, with a distinction invisible in error-point scoring alone.

The most informative divergence concerned Scenario 3d: low quantitative error scores suggested adequate performance, yet qualitative analysis revealed that this reflected a transfer effect from prior failure experience rather than autonomous hazard awareness. This highlights the limitations of a purely quantitative evaluation in competency assessment and underscores the value of mixed-methods designs for training evaluation in safety-critical domains. Because the organisational structure, training system, and the GAMS scheme itself are specific to German(-speaking) fire services, the error frequencies reported here are not directly generalisable; what is more likely to transfer is the underlying mechanism, a rescue impulse that pre-empts hazard assessment at the perception-to-action transition, which is grounded in domain-general models of stress-induced habitual responding rather than any national protocol.

Notably, the qualitative strand identified a resource dimension absent from quantitative scoring: functioning teamwork, specialist knowledge among individual participants, a high capacity for reflection, and concrete behavioral change intentions. These resources constitute favorable preconditions for targeted remedial training and suggest that interventions should focus on the procedural automatization of hazard assessment rather than knowledge acquisition.

This study examined the extent to which volunteer firefighters activate trained hazard-assessment protocols in realistic non-fire emergency scenarios and how rescue motivation and self-protection knowledge interact to shape behavioral responses. The answer is unequivocal: trained protocols were not activated. Despite universal declarative mastery of the GAMS schema, no group applied systematic hazard assessment before entering any practical scenario, resulting in invariant critical action errors across chemically, biologically, and electrically hazardous environments. The quantitative and qualitative data converge on a single explanatory mechanism: an automatic rescue impulse that overrides self-protection knowledge at the perception-to-action transition, producing a consistent pattern of seeing without evaluation (Figure 1). Figure 1 integrates the two data strands into a single causal pathway: the trained sequence proceeds from perception through hazard comprehension, deliberation and scene control to a protected rescue, whereas the rescue impulse branches at the perception stage and bypasses the deliberative core (stages 3–5), yielding unprotected entry and a critical self-protection error. This is a form of altruism that remains unreflected in the heat of the moment.

The invariance of this rescue impulse across heterogeneous hazard types is both theoretically significant and practically applicable. Klein's Recognition-Primed Decision model predicts that experienced operators match perceived cues to stored action scripts, bypassing deliberate analysis (5). However, our participants were not like the experienced fireground commanders studied by Klein; they were volunteer firefighters with limited exposure to non-fire scenarios. Rather than matching a “ correct” pattern, they defaulted to the only script available: person in distress → intervene immediately. This finding extends RPD theory by demonstrating that pattern-matching failures in novice or inexperienced responders do not manifest as hesitation or confusion, which is the predicted outcome when no matching pattern exists in the response. However, it activates a dominant but contextually inappropriate rescue script. The rescue script appears to function as a “default option” in Klein's extended RPD framework (24). When no specific pattern is recognized, operators fall back on the most readily available action schema. For volunteer firefighters whose operational identity centres on helping others, the rescue script constitutes the most deeply consolidated default. Converging field evidence comes from Cohen-Hatton, Butler, and Honey, who analyzed helmet-camera footage of 33 real incidents attended by professional incident commanders. They found that situation assessment was typically followed directly by plan execution, with little prospective evaluation of the consequences of action, a reflexive shortcut moderated by operational experience (27). Thus, deliberate assessment does not appear to be the default, even at the command level, but rather depends on experience. In our sample, this experience-based, command-level moderation was structurally absent, as participants operated as first-on-scene responders without an incident command hierarchy (28, 29).

Endsley's three-level model of Situation Awareness provides a complementary lens (8). Our qualitative data show that Level 1 SA (perception) was largely intact, as participants detected warning signs. Level 2 SA (comprehension), which involves integrating perceived cues into a coherent threat assessment, was not achieved. The GAMS schema was designed to bridge this gap, yet it functioned exclusively as a retrieval structure in retrospective reflection, never as a real-time processing tool. This interpretation aligns with Berthiaume et al.'s finding that even structured VR-based hazard training improved recognition and identification but did not automatically translate into changed behavioral priorities when competing goals were present (30).

The knowledge–action dissociation observed in this study aligns with neurocognitive models of stress-induced behavioral shifts. Schwabe and Wolf demonstrated that acute stress promotes a transition from hippocampus-dependent declarative processing to striatum-dependent habitual responding (11), effectively bypassing the cognitive architecture in which procedural rules, such as the GAMS, are stored. Andersen et al. documented similar effects in police officers and emergency responders, showing that stress impairs declarative memory retrieval while leaving over-learned motor responses intact (12). In our participants, the rescue impulse may constitute such a habitual response, which could be consolidated through years of firefighter identity and social reinforcement. Simultaneously, GAMS, taught through didactic instruction without scenario-based rehearsal, remains trapped in declarative storage. This interpretation is supported by McLaughlin's automatization framework, which requires consistent mapping across many repetitions for procedural transfer, a condition that the typical volunteer firefighter training schedule cannot satisfy (31).

Beyond this neurocognitive shift, the operational reality of emergency response is increasingly characterized as a volatile, uncertain, complex, and ambiguous (VUCA) environment (32), in which the preconditions for retrieving and applying declaratively stored rules are unstable. Within crisis resource management, certain action tendencies under such conditions have been described as ballistic: once initiated, they can no longer be modulated by deliberation (33). In the helicopter emergency medical services literature, this phenomenon is observed as a helping impulse so dominant that even brief reflective pauses, such as the 10-for-10 procedure, cease to function. Our finding that participants moved directly from perception to rescue action, with no observable hesitation across all three practical scenarios, is consistent with this description. The operational implication is fundamental: training cannot rely on the assumption that personnel will be cognitively available to retrieve memorized rules under pressure. The rule itself must be installed below the threshold of deliberate retrieval through repetition in a realistic stress environment.

A notable contribution of this study is the first empirical test of GAMS's effectiveness as a behavioral tool. Widetschek introduced the GAS rule in 1978. The GAS rule was later extended to the four-element GAMS, and it has since been embedded in FwDV 500 as the standard initial hazard management protocol for German fire services (16). Our data provide a clear answer: in this sample, it did not. All 21 participants could reproduce the acronym, but none activated it under pressure in the scenario simulation. Mnemonics function as retrieval aids for declarative content (34). However, they do not automatically generate procedural routines. Hardison and Gray demonstrated that energy-based hazard recognition training, which categorizes hazards by physical energy source rather than mnemonic, improved hazard identification scores in 43 firefighters, suggesting that classification-based approaches may offer a more robust bridge from recognition to action than acronym-based schemas (35).

It is important to acknowledge the conceptual limitations of GAMS. As a four-element acronym, it offers a heuristic entry point into hazard management but does not encompass the broader construct of safety-conscious operational behavior. More elaborate frameworks have been developed in adjacent high-reliability domains, most notably the 10-for-10 principle from crisis resource management, which is a brief structured pause in which the team takes ten seconds to share a mental model before committing to ten minutes of action (33, 36), and the behavioral safety protocols of the Stop–Think–Plan–Act family used in process industries. Unlike GAMS, these frameworks explicitly install a deliberation phase between perception and action and are coupled with team-level communication rules. Therefore, they may offer a more powerful behavioral architecture than acronym recall alone and constitute a plausible target for future intervention design in volunteer fire services.

The rescue impulse–self-protection conflict identified in this study resonates with the broader goal-conflict literature on firefighting. Hagemann, Heinemann et al. documented systematic goal conflicts between safety compliance and mission achievement in German firefighters, identifying role identity and social pressure as the primary drivers of risk acceptance (3). Our data extend this finding from the decision-making level to the pre-decisional level: the participants did not decide to override safety protocols. They never reached a decision point. The rescue impulse operated automatically, preempting the deliberative process entirely. Regardless of the operational context, the quickest decisions humans can make in time-pressured, high-risk situations are often recognition-primed, intuitive, value-driven, or based on heuristics. These approaches are commonly referred to as naturalistic strategies (37, 38). Research frequently highlights that personnel such as firefighters, police officers, paramedics, and military members rely on these rapid decision-making methods in critical, time-pressured, and high-risk scenarios. Estimates suggest that up to 80% of the decisions made by first-arriving fire officers follow a naturalistic, recognition-primed approach (38). Maglio et al. reported a parallel pattern in PPE decision-making, in which firefighter identity systematically undermined protective equipment use despite complete knowledge of the risks (39). Together, these findings suggest that the rescue–safety conflict is not a deliberative trade-off but a motivational capture phenomenon that bypasses the rational weighing of alternatives.

Because Scenario 3d was assessed theoretically only, we make no quantitative transfer claim. The debriefing did, however, appear to activate metacognitive processing, reflected in the high proportion of reflection/insight coding (HK4) and concrete behavioural-change intentions, consistent with Waring et al.'s finding that debriefing tied to exercise objectives is the strongest predictor of learning in full-scale emergency exercises (40). In a comprehensive systematic literature review on decision-making in high-risk, time-pressured environments, Penney et al. highlighted the importance of opportunities for feedback and structured reviews in developing expertise. This expertise includes the ability to employ various decision-making strategies and use intuition safely and effectively. Reflective practice is also considered essential for acquiring and maintaining expertise, as it enables individuals to reorganize their knowledge by assessing experiences and evaluating the outcomes of their actions. These review and reflection processes help refine mental models or schemas that link experiences to outcomes, thereby supporting the development of intuitive decision-making. For instance, paramedics reported analyzing cardiac registry data (41) and seeking constructive feedback from hospital staff regarding patient outcomes and treatment approaches. However, receiving negative feedback from hospitals was associated with delayed and more risk-averse decision-making in subsequent situations. The high proportion of reflection and insight coding (27.7%) and concrete behavioral change intentions (18.5%) in our qualitative data suggest that the integrated debriefing format successfully activated metacognitive processing. However, the absence of action pressure in Scenario 3d precludes the definitive attribution of the improvement to cognitive learning rather than situational facilitation.

The educational mechanism by which scenario training generates competence lies primarily in structured post-event reflection, which Eppich and Cheng termed the learning conversation of debriefing. Frameworks such as PEARLS (42) and the advocacy–inquiry approach (43) operationalize this conversation as a guided process in which trainees surface their cognitive frames, examine the gap between expected and observed outcomes, and reconstruct their mental models for future situations. From a learning-theoretical perspective, this reconstruction is the proposed mechanism by which scenario exposure produces durable behavioral change: the repeated activation and revision of internal schemata through reflective practice is a precondition for the experience-dependent restructuring described by Kolb (44) and required by Endsley's (8) Level 2 situation awareness. For volunteer fire services, this implies that scenario training without structured debriefing will, at best, produce isolated experiences; only when paired with disciplined reflective dialogue does scenario training generate the procedural automatization required to override the rescue impulse.

A further implication of these findings concerns the conceptual distinction between exposure, experience, and competence, three constructs that are often conflated in fire-service workforce planning (45) but that map to distinct developmental processes. Exposure denotes mere presence at events; experience denotes the accumulation of structured engagements with operational situations; and competence (46), context-sensitive expertise evolves and emerges only through the reflective integration of experience over time (47, 48). For German volunteer fire services, where the average operational frequency is far below that of career personnel and non-fire technical incidents may occur only sporadically over a member's career, this hierarchy is consequential, as the structural conditions for experience-based competence development are not satisfied by operational exposure. Compensation requires didactic architectures that substitute or supplement scarce real-world events through deliberate practice, scenario-based rehearsal under realistic stress, structured reflection, and sufficient resources for high-quality practice and training. This places the burden of competence development squarely on training design rather than on operational frequency, and provides a theoretically grounded rationale for the procedural-automatization recommendations advanced in this study.

Strengths and limitations

Several design features strengthen confidence in these findings. The convergent mixed-methods design allowed triangulation that neither data strand could achieve alone: quantitative invariance of H-CRIT errors gained explanatory depth through the qualitatively identified rescue impulse mechanism, while the qualitative resource dimension (functioning teamwork, specialist knowledge, reflection capacity) would have been entirely invisible to quantitative scoring alone. The use of standardized observational instruments with systematic error coding (13 categories, three-level scoring), combined with Mayring-based content analysis of 130 speech sequences, provides methodological rigor exceeding the typical evaluation practice in German volunteer fire services, where structured assessment instruments are largely absent (21, 49).

This study has some important limitations. The sample of 21 volunteer firefighters from a single department limits generalizability, although the invariance of the core finding across all groups provides internal replication. With over 90% of German firefighters serving as volunteers (18), the sample was representative of the population most likely to encounter non-fire emergencies as first responders. However, replication across departments with different training cultures and operational frequencies is required. Scenario-based design, while ecologically valid for training evaluation, cannot fully capture the physiological and psychological stress of real emergencies; the effects of actual rescue impulses may be stronger than those observed here. Scenario 3d was assessed theoretically rather than practically, confounding the transfer effects with reduced action pressure. The cross-sectional single-session design precludes the assessment of training durability. Finally, intercoder reliability was established through independent parallel double-coding of all sequences at the main-category level; the 30 subcategories were not subjected to independent verification, and the reported κ therefore pertains to main-category assignment. Because κ is sensitive to category prevalence, we report raw agreement alongside it. As a single primary coder produced the final codes (with consensus adjudication of discrepancies), systematic idiosyncrasies of one coder cannot be fully excluded.

In addition, the present design captures performance at a single point in time and provides no insight into the development of longitudinal competence. Given that competence in the Dreyfus and Dreyfus (46) sense emerges only from the reflective integration of experience over time, the cross-sectional findings reported here cannot indicate whether the observed competence gap could be closed through a systematic training intervention. Longitudinal designs that combine baseline assessment, structured intervention with embedded debriefing, and follow-up measurements are required to test the proposed proceduralization mechanism.

Declarative safety knowledge does not automatically translate into procedural safety behavior under operational pressure: volunteer firefighters in this study universally failed to activate their trained hazard-assessment protocols despite intact declarative knowledge, with the rescue impulse, the automatic, pre-decisional activation of intervention before hazard evaluation, as defined above, systematically overriding self-protection at the perception-to-action transition for all hazard types. These findings demonstrate that mnemonic-based safety training, specifically the GAMS schema, fails to bridge the gap between knowledge and behavior under pressure in scenario-based situations. Rescue motivation may preempt hazard evaluation before deliberate reasoning is activated. Moving from declarative instruction to procedural competence will require scenario-based rehearsal that specifically targets the rescue impulse–self-protection transition point, potentially integrating energy-based hazard classification with structured decision protocols practiced under realistic stress (35). Future research should examine whether repeated scenario exposure with structured debriefing can shift hazard assessment from deliberate processing toward the automatized routines that the RPD model identifies as the hallmarks of expert performance. The robust reflection capacity and change readiness identified in our qualitative data suggest that volunteer firefighters are aware of this critical gap and motivated to close it. This finding should encourage training developers to invest in the deliberate practice architectures necessary to transform knowledge into life-saving behaviors.