Abstract
Keywords
Parental cooperation during child protection services (CPSs) interventions directly influences case trajectories, intervention success, and child safety outcomes (Ben-David, 2016; Charest-Belzile et al., 2020). Cooperation encompasses the mutual, intentional, and behavioral involvement of caregivers in services provided by child protection and related agencies (Platt, 2012). Consequently, its absence limits parents’ capacity to make necessary changes, reducing the likelihood of positive outcomes in CPS interventions (Ben-David, 2016). The practical implications extend beyond mere compliance—“doing what they are told”—toward “true engagement” (Charest-Belzile et al., 2020; Platt, 2012); beyond acting out of fear of legal consequences (Ben-David, 2016); and beyond disguised compliance, concealing underlying parental opposition (Forrester et al., 2012; Mason et al., 2020). Noncooperation in mandated CPS contexts does not always reflect active resistance; structural barriers such as poverty, mental health difficulties, or housing instability may limit parental engagement in ways that are indistinguishable in case documentation from deliberate unwillingness (Charest-Belzile et al., 2020; Platt, 2012). The coercive context of mandated CPS involvement, where noncooperation may carry legal consequences, further complicates the interpretation of documented cooperation patterns (Ben-David, 2016). Effective cooperation requires attitudinal engagement, demonstrated through trust and openness toward services; relational alignment, characterized by a collaborative working relationship with caseworkers; and behavioral participation, expressed through consistent attendance and active involvement in interventions (Charest-Belzile et al., 2020). Moreover, drawing on our research experience, we consider parental cooperation to be an important, though often implicit goal, alongside other core objectives, in CPS interventions.
Given its recognized importance, parental cooperation has increasingly come to the forefront in social work practice and has received substantial research attention in both qualitative and quantitative studies (Gautschi, 2021; Jud & Gartenhauser, 2015; Lätsch et al., 2022). The proposed automated classification method would enable population-level data collection on this case factor. Unlike discrete variables such as substance use or domestic violence, cooperation is not captured in the structured administrative data maintained by child welfare information systems. Information about cooperation exists only in narrative documentation: case notes, assessment reports, and accountability reports written by caseworkers over the course of an intervention (Witte, 2020). This presents a fundamental measurement challenge. Cooperation is inherently dynamic; it emerges, fluctuates, and evolves across months or years of service involvement. A parent who initially resists engagement may develop genuine collaboration over time, while apparent compliance may mask underlying resistance (Forrester et al., 2012; Mason et al., 2020). Capturing these patterns requires analyzing narrative records that document changing circumstances over extended intervention periods. Systematic assessment of cooperation from case documentation presents distinctive analytical challenges. Information within a single case report is often contradictory, ambiguous, or reflects change over time. Assessment, therefore, requires weighing conflicting evidence and determining the net trajectory rather than making simple present/absent classifications.
Gender Perspective of Parental Cooperation
Additionally, research on gender-related perspectives on working with parents in child welfare suggests that professionals interact differently with mothers and fathers, leading to a systematic bias in cooperation assessments (Philip et al., 2019; Scourfield, 2001; Scourfield et al., 2024). Child protection systems are characterized as mother-centric (Strega et al., 2008), assigning mothers primary responsibility for children's well-being while often treating fathers as peripheral, nonclients (Scourfield et al., 2024). Social workers tend to be more tolerant of noncooperation from sole-carer mothers, who are seen as having no choice but to stay engaged for the child's benefit (Philip et al., 2019). Mothers are contacted earlier and more often, while information on fathers is frequently indirect due to rushed or delayed engagement (Dominelli et al., 2011; Philip et al., 2019; Strega et al., 2008). Organizational routines reinforce this imbalance, resulting in case documentation that provides more detailed and nuanced information about mothers than fathers (Strega et al., 2008), while fathers are often assessed in binary terms—cooperative or not—rather than through nuanced evaluation (Philip et al., 2019).
Automated Text Analysis in Child Welfare Through Reasoning Methods
Recent advances in natural language processing have expanded the capacity to extract structured data from unstructured case narratives, driven in part by the need for secure methods that protect highly sensitive child-welfare data from disclosure or use by commercial AI providers. Language models (computational systems trained to process and generate human language) can now classify discrete risk factors in child welfare records with accuracy comparable to human experts (Perron et al., 2024). They are available as open-source software and can be deployed locally within the secure environment of a child protection agency or research institution, ensuring that all data remain on-site and comply with data-protection requirements. These methods have successfully identified parental substance-related problems, domestic violence, opioid involvement, and firearms presence without requiring extensive manual annotation of training data (Perron et al., 2024).
Standard language models generate responses by predicting likely word sequences from patterns learned during training, processing each input in a single pass. Instruction-tuned models—a refined variant trained through additional human feedback—follow explicit task instructions more reliably and form the foundation of many widely used AI systems. However, both approaches struggle when information is divergent, shifting, or requires synthesis across extended narratives. Our prior attempt to classify parental cooperation employed retrieval-augmented generation (RAG), a technique applied to instruction-tuned models that first retrieves the most relevant passages from a document before generating a response, rather than processing the full text. This approach achieved an accuracy of 80%, but error analysis revealed that it failed to reconcile contradictory evidence within reports or to track how cooperation evolved across the reporting period (Stoll et al., 2025).
Reasoning language models (RLMs) offer a methodological advance suited to this analytical challenge. These models are designed to decompose complex problems into sequential steps before producing final outputs (Besta et al., 2025; Raschka, 2025). The core mechanism, termed chain-of-thought processing, encourages the model to generate intermediate reasoning steps that lead to a conclusion rather than producing direct classifications (Wei et al., 2022). For cooperation assessment, this means the model can identify evidence both for and against cooperation, weigh temporal changes documented in the narrative, and reconcile contradictions before rendering a final determination. This process mirrors the cognitive approach that expert human reviewers (EHRs) employ when evaluating complex case information. Model capacity is expressed in parameters—the numerical values learned during training that determine model behavior—with larger counts generally associated with greater capability. This study evaluates RLMs of three sizes (4 billion, 32 billion, and 255 billion parameters), compared against an instruction-tuned model of 123 billion parameters using RAG. Recent research demonstrates that reasoning-enabled models achieve substantial performance improvements on classification tasks in child welfare contexts, with smaller models matching or exceeding the accuracy of much larger architectures when reasoning capabilities are enabled (Qi et al., 2026).
Purpose
This study evaluates whether RLMs can accurately assess parental cooperation from CPS case reports, a construct marked by ambiguous and conflicting information that has proven difficult for prior automated approaches. We pursue three objectives. First, we compare reasoning-enabled models of different sizes (4 billion to 255 billion parameters) against our prior RAG approach to determine whether reasoning capabilities improve classification accuracy for this complex task. Second, using a semantic approach designed to distinguish caregiver roles, we assess whether model performance differs between mothers and fathers, informed by hypotheses from research on gendered documentation patterns in child welfare. Third, we demonstrate practical utility by applying the validated method to classify cooperation across a large corpus of CPS cases (N = 29,770 reports spanning 12,607 cases), generating structured data that would be infeasible to obtain through manual review. The findings address both methodological questions about appropriate AI techniques for complex assessment tasks and substantive questions about how cooperation is documented and understood across caregiver roles in child protection practice.
Method
Data Source
This study utilizes casework reports from the child protection system in the Canton of Zurich, Switzerland. In this system, Child Protection Authorities (comparable to child welfare courts in other countries) assess potential maltreatment cases and determine necessary interventions. The most common intervention involves appointing a social worker as an “assistant to the child,” who provides counsel to families, coordinates services, and monitors case progress (Jud et al., 2011). Approximately every 2 years, appointed social workers submit accountability reports to the authorities documenting the case status, family circumstances, and intervention outcomes.
Since 2008, these reports have followed a standardized structure including: contact and demographic information; the legal mandate and case objectives; the child's current situation and well-being; case developments over the reporting period; perspectives of the child and parents on key matters; prognosis for future development; and recommendations for continued action. We obtained all 29,770 casework reports from 12,607 CPS cases spanning 2008 to 2022 through a data-sharing agreement with the cantonal authorities. Reports averaged approximately 1,300 words (five pages), with a standard deviation of 730 words. The median duration of CPS assistance was 3.3 years, though cases in the upper quintile averaged approximately 10 years. Table 1 summarizes data source characteristics. All casework reports were digitally created in an electronic word-processing format and were transferred to the research team in this form, reducing transcription errors associated with manual or paper-based records. The standardized report structure introduced in 2008 provides a consistent documentation framework. Nevertheless, as with all practitioner-generated administrative records, the reports are subject to variability in documentation quality, linguistic precision, and completeness across caseworkers, which is addressed further in the Limitations section.
Data Source Metainformation.
Analytical Workflow
We developed a four-stage workflow for assessing parental cooperation using RLMs. The first stage involved collecting and preprocessing all casework reports from their original word processing format into plain text. The second stage applied reasoning-based question answering to assess maternal and paternal cooperation separately. The third stage extracted structured cooperation categories from model outputs. The fourth stage aggregated report-level classifications to generate case-level labels indicating whether lack of cooperation occurred at any point during CPS involvement.
Reasoning-Based Assessment
As described in the introduction, RLMs differ from standard and instruction-tuned language models by generating extended chains of intermediate thinking steps before producing a final answer—a process termed chain-of-thought reasoning (Wei et al., 2022)—rather than producing responses in a single processing pass. This extended reasoning trace, or “thinking budget,” allows RLMs to work through complex, contradictory information more thoroughly than other model types. For cooperation assessment, this means the model generates intermediate reasoning that identifies evidence for and against cooperation, weighs how cooperation evolved across the reporting period, and reconciles contradictory information before rendering a final decision.
We developed the assessment prompts through iterative refinement by examining model outputs on randomly sampled reports (Perron et al., 2024). The full prompt used in the study is provided in Appendix A. It consists of five components: (1) an instruction specifying that the model should answer based only on explicitly documented information; (2) the assessment question with three response categories (lack of cooperation, cooperation present or emerged, no evidence); (3) operational definitions of lack of cooperation and established cooperation; (4) assessment guidelines addressing common interpretation challenges; and (5) the case report text. We assessed mothers and fathers cooperation separately using parallel prompts.
Lack of cooperation was operationally defined as: the parent does not follow professional instructions; is uncooperative or unwilling to work with the caseworker or other professionals; does not respond to or follow guidance; does not attend agreed appointments; cannot be motivated to adopt new perspectives; or is formally instructed under Article 307 of the Swiss Civil Code, which imposes a legal obligation to comply with child protection authority orders (Swiss Civil Code (Zivilgesetzbuch, ZGB), 1907). Whereas cooperation, present or emerging, was defined as: the parent is cooperative and willing to work with professionals, or willingness has developed over time; the parent responds to and follows professional guidance; and the parent attends agreed-upon appointments.
The assessment guidelines addressed interpretive challenges identified during prompt development. When reports contained evidence of both cooperation and lack of cooperation, the model was instructed to evaluate the overall trajectory to determine whether cooperation emerged over time. When no evidence of either cooperation or lack of cooperation was observed, the model was instructed to classify as “no evidence” rather than infer cooperation status. When reports referred to “parents” collectively, the model was instructed to apply this information to both the mother and father assessments.
The operational definition was developed through iterative refinement involving the research team, which includes researchers with extensive experience in CPS practice and research. Initial definitions were repeatedly tested against a subset of case reports, discussed among team members, and revised to reflect interpretive challenges encountered in the data. The same definition was applied to both the model and the EHRs to ensure comparability between automated and human classification.
Model Selection and Configuration
We evaluated open-source RLMs of three sizes to assess the relationship between model capacity and classification accuracy. Model capacity is measured in parameters (the numerical values learned during training that determine model behavior), with larger parameter counts generally associated with greater capability but also greater computational requirements. We selected models from the Qwen3 family (Yang et al., 2025) because these models demonstrated strong performance on multilingual tasks, including German text processing.
The three reasoning models evaluated were Qwen3-235B-A22B (255 billion total parameters, 22 billion active per inference), Qwen3–32B (32 billion parameters), and Qwen3–4B (4 billion parameters). The largest model employs a mixture-of-experts architecture, meaning it activates only specialized subnetworks, chosen experts, for each task rather than processing through all parameters simultaneously. For comparison with our prior approach, we also evaluated a 123-billion-parameter instruction-tuned model (Mistral-Large-Instruct-2411) using RAG, which retrieves relevant text segments before classification, rather than processing complete reports through reasoning steps (Jiang et al., 2023).
All models were deployed using 4-bit quantization, a compression technique that reduces memory requirements while preserving most of the model's capabilities (Dettmers et al., 2023) and were obtained from the Hugging Face model repository. Following recommendations from the Qwen3 development team, we configured reasoning models with a temperature of 0.6, top-k of 20, and top-p of 0.95. Temperature controls output randomness, with lower values producing more deterministic responses; top-k and top-p control the diversity of word choices during generation. Maximum output length was set to 8,000 tokens (approximately 6,000 words; equivalent of up to 20 pages of text) to accommodate extended reasoning chains. Model inference was performed using the vLLM framework, which optimizes processing speed for large language models (Kwon et al., 2023).
Category Extraction and Case Labeling
Reasoning model outputs contain two components: a “thinking” section that documents intermediate reasoning steps, and a final answer that includes the classification with supporting justification. An anonymized final assessment illustrating a case of noncooperation is provided in Appendix B. We extracted cooperation categories from the final answer using a separate instruction-tuned model (Qwen3–32B), configured to output structured JSON format. This extraction model received only the final answer text and was instructed to identify which of the three categories (lack of cooperation, cooperation present, or no evidence) the reasoning model selected. The extraction prompt is provided in Appendix C.
For analysis purposes, we aggregated cooperation categories into a binary classification. During validation, EHRs evidenced a difficulty distinguishing between two scenarios: cases in which documentation provided sufficient evidence to conclude that cooperation was present, and cases in which the documentation simply lacked any information about cooperation. Both scenarios share a common practical characteristic; neither contains evidence of problematic engagement. We therefore combined the “cooperation present or emerged” and “no evidence” categories into a single category indicating no documented lack of cooperation, while retaining “lack of cooperation” as a distinct category.
During the development and validation of our approach, we observed in the casework reports the binary framing that typically characterizes practitioners’ assessments of parental cooperation; resistance or nonengagement triggers concern and is documented, while cooperative behavior is only sporadically mentioned, mostly in combination with other arguments regarding the case development. We conducted sensitivity analyses comparing model performance under the original three-category scheme versus the binary classification to assess whether this aggregation affected accuracy estimates. Case-level labels were then generated by flagging any case in which at least one report indicated a lack of cooperation by the respective parent.
Validation Procedures
We validated model classifications against expert human review using a stratified random sample of 100 casework reports. The sample was constructed to ensure representation across classification combinations: 20 reports with both parents classified as lacking cooperation, 20 reports with neither parent classified as lacking cooperation, and 15 reports each for the two discordant patterns (mother lacking cooperation but not father; father lacking cooperation but not mother). Two EHRs, both with graduate training in social work and experience in child protection practice and documentation, independently classified maternal and paternal cooperation for each sampled report.
Expert reviewers were instructed to examine each report for indications of cooperation or lack thereof, note specific passages informing their evaluation, and classify cooperation for each parent using the same three-category scheme applied by the models. Reviewers provided justification for classifications when deemed necessary. All disagreements between reviewers were discussed and resolved through consensus, producing a benchmark dataset against which model performance was evaluated.
Evaluation Metrics
Model performance was assessed using four metrics. Weighted overall accuracy measured the proportion of classifications matching the expert consensus benchmark. Precision measured the proportion of positive predictions (lack of cooperation) that were correct. Recall (also termed sensitivity) measures the proportion of actual positive cases that the model correctly identified. F1-score provided the harmonic mean of precision and recall, balancing both error types in a single measure.
We additionally computed Cohen's κ to measure agreement after accounting for the level of agreement expected by chance. Following conventional interpretation guidelines (Landis & Koch, 1977), κ values between 0.41 and 0.60 indicate moderate agreement, between 0.61 and 0.80 indicate substantial agreement, and above 0.81 indicate almost perfect agreement. We computed κ for model-to-benchmark agreement and for inter-rater agreement between the two expert reviewers to contextualize model performance relative to human consistency. Confusion matrices were constructed to characterize the nature and direction of classification errors.
Data Protection
Given the sensitive nature of child protection records, we implemented comprehensive data protection measures. All researchers signed confidentiality agreements. Structured data were anonymized by excluding or recoding identifying variables. All model processing was conducted locally on secure infrastructure without transmission to external servers. A detailed data management and protection plan was reviewed by the cantonal data privacy office and approved by a data privacy and ethics lawyer. The computational infrastructure consisted of high-performance computing server-class hardware with sufficient graphical memory to deploy the evaluated models locally. Due to confidentiality requirements, the datasets cannot be made publicly available; however, all analysis code is available in the project's GitHub repository.
Results
Reasoning Models Versus RAG Approach
Table 2 presents accuracy, precision, recall, and F1-score for classifying parental cooperation, comparing reasoning models of three sizes against the RAG approach. The large reasoning model (255B parameters) achieved the highest overall accuracy (89%), representing a nine percentage point improvement over the RAG-based approach (80%). Performance scaled with model size: the medium reasoning model (32B) achieved 84% accuracy, and the small reasoning model (4B) achieved 80% accuracy. The small reasoning model matched the performance of the substantially larger RAG-based instruct model (123B), suggesting that reasoning capabilities can compensate for reduced model capacity.
Evaluation Metrics for Classification of Parental Cooperation Compared to Consensus Dataset According to the Applied Model.
Table 3 presents confusion matrices for the large reasoning model, which achieved the highest overall performance and was subsequently used for full corpus classification. Across both parents (N = 200 classifications), the model correctly identified 118 of 132 actual cases of cooperation (89.4% recall) and correctly classified 60 of 68 cases without documented lack of cooperation (88.2% specificity). The false positive rate (11.8%) was lower than that of the RAG-based approach (25.0%), indicating that reasoning capabilities reduced the erroneous identification of cooperation problems where none existed.
Confusion Matrices for Classification of Parental Cooperation Compared to Consensus Dataset for the Large Reasoning Model.
Note. Bold values indicate instances of false predictions, representing cases where the model's output did not match the actual classification.
Table 4 presents Cohen's κ coefficients measuring agreement between model classifications and the expert consensus benchmark. The large reasoning model achieved κ = 0.76 for overall parental cooperation, indicating substantial agreement according to conventional interpretation guidelines (Landis & Koch, 1977). This represents an improvement over the prior RAG-based approach (κ = 0.62). To contextualize model performance relative to human consistency, we also computed κ between the two EHRs before consensus was reached. The model's agreement with the benchmark (κ = 0.76) approached the level observed between individual EHRs and the consensus benchmark (κ = 0.88 and κ = 0.80, respectively).
Percentage Agreement and Cohen's κ Metric for Inter-Rater Agreement for Classification of Parental Cooperation Between the Model, Expert Human Reviewers (EHRs) and the Consensus Dataset as a Benchmark for the Large Reasoning Model.
Note. Values between κ = 0.41 and κ = 0.60 indicate moderate agreement, between κ = 0.61 and κ = 0.80 substantial agreement, and larger than κ = 0.81 indicate complete agreement (Landis & Koch, 1977).
Parent-Gender Differences
Classification accuracy differed substantially between mothers and fathers across all models (Table 2). For the large reasoning model, accuracy was higher for mothers (93%) than fathers (85%). This eight percentage point gap persisted across model sizes: the medium model achieved 86% accuracy for mothers versus 82% for fathers, and the small model achieved 79% for mothers versus 81% for fathers. The RAG-based approach showed an even larger gap (85% for mothers, 75% for fathers).
Cohen's κ values reflected this pattern (Table 4). For the large reasoning model, agreement with the benchmark was almost perfect for mothers (κ = 0.85) but only substantial for fathers (κ = 0.66). Error analysis revealed that the model failed to identify 16% (n = 5) of actual cases of paternal noncooperation compared to 8% (n = 3) of maternal cases (Table 3). The model also produced more false positives for fathers (16.1%) than mothers (8.1%).
Notably, EHRs exhibited similar difficulties. Prior to consensus resolution, inter-rater agreement between the two EHRs was substantially lower for fathers (κ = 0.65) than for mothers (κ = 0.71). The model's agreement with the benchmark for fathers (κ = 0.66) closely matched the agreement observed between the two human reviewers. These parallel patterns suggest that reduced classification accuracy for fathers reflects characteristics of the source documentation rather than model-specific limitations.
Sensitivity Analysis
Sensitivity analyses compared model performance under the original three-category classification scheme versus the binary aggregation. Under the three-category scheme, overall accuracy for the large reasoning model was 78% compared to 89% under binary classification. The improvement reflected reduced ambiguity in distinguishing “cooperation present” from “no evidence” categories. Cohen's κ showed a similar pattern (three-category: κ = 0.65; binary: κ = 0.76). The relative performance ranking across models remained unchanged under both classification schemes, indicating that the binary aggregation did not systematically advantage particular model configurations.
Full Corpus Classification
Table 5 presents classification results for the full corpus of 29,770 casework reports from 12,607 CPS cases using the large reasoning model. At the report level, lack of cooperation was identified in 17.5% of reports for mothers (n = 5,261) and 18.3% of reports for fathers (n = 5,483). At the case level, 17.6% of cases (n = 2,153) had at least one report documenting maternal noncooperation, and 18.8% of cases (n = 2,366) had at least one report documenting paternal noncooperation. Overall, 31.0% of cases (n = 3,900) had documented lack of cooperation by at least one parent at some point during CPS involvement.
Results for Parental Lack of Cooperation: Classification of Casework Reports and Labeled Child Protection Service (CPS) Cases.
Note. Results are presented in absolute numbers and as percentage of all case reports and CPS cases, respectively.
Processing the full corpus required approximately 375 h of computation time on high-performance server-class hardware with 192GB of graphical memory, with each report requiring an average of 45 s for reasoning-based classification. This processing time, while substantial, enabled extraction of structured cooperation data from a corpus that would require an estimated 5,000 h (approximately 120 weeks of full-time work) for manual review at 10 minutes per report.
Discussion and Applications to Practice
This study demonstrates that reasoning-enabled language models can accurately classify parental cooperation from child protection documentation—an assessment domain characterized by ambiguous, temporally variable, and often contradictory information. The large reasoning model achieved the strongest performance, improving accuracy from 80% (RAG-based approach) to 89% and approaching expert inter-rater reliability, with accuracy scaling with model size. Applied to the full corpus of Switzerland's largest child protection provider, 31% of cases had at least one parent with documented noncooperation, illustrating the scale of insight that automated classification can generate from existing administrative records.
These results extend prior research showing that small reasoning models can match or outperform larger instruction-tuned models on well-defined classification tasks such as substance-related problems or domestic violence (Qi et al., 2026). However, parental cooperation is conceptually distinct, encompassing attitudinal, relational, and behavioral dimensions that unfold over extended service involvement. Our findings indicate that larger reasoning models are better equipped to synthesize this complexity through structured chain-of-thought processing, substantially improving classification validity compared to earlier RAG-based approaches (Stoll et al., 2025) and enabling population-level analysis that would be infeasible through manual review.
Despite these advances, several methodological limitations warrant consideration. First, accuracy was consistently lower for fathers than for mothers. The large reasoning model missed 16% of documented instances of paternal noncooperation, compared to 8% for maternal cases, and inter-rater reliability between human reviewers showed the same pattern (fathers: κ = 0.65; mothers: κ = 0.71), suggesting the performance gap reflects characteristics of the underlying documentation rather than model behavior. Documentation differences between mothers and fathers likely contribute to this gap, reflecting systemic bias that extends to automated classification. Across the full corpus, lack of cooperation was documented in 18.8% of cases for fathers and 17.6% for mothers—figures that should not be interpreted as true prevalence rates, as automated classification can only assess what is written. Case documentation is shaped by biases throughout the reporting process, including uneven access to information, selective attention, and strategic choices about what to record. Consistent with prior research characterizing child protection systems as mother-centric (Dominelli et al., 2011; Philip et al., 2019; Scourfield et al., 2024; Strega et al., 2008), documentation on fathers is less extensive and more dichotomous, limiting nuanced assessment. Both human reviewers and reasoning models achieved lower accuracy for fathers and were more likely to miss noncooperation, suggesting parental cooperation cannot be measured equivalently across caregiver roles and that comparisons should be interpreted cautiously. Automated classification does not eliminate such biases and may reinforce them if source documentation or model pre-training data contain systematic distortions (Garrido-Muñoz et al., 2021). We mitigated interpretive bias by defining cooperation constructs explicitly in the prompts and validating the reasoning process with human reviewers. Nevertheless, pretraining biases and asymmetries in source documentation remain unavoidable constraints. Future work should examine how documentation practices influence automated assessments and how model reasoning can be calibrated to identify and flag potential gaps or inconsistencies in narrative records.
Second, the computational demands of the largest reasoning model remain substantial. Deploying a 255B-parameter reasoning model locally required specialized hardware (a minimum of 192GB of graphical memory) and generated an average inference time of 45 s per report. In contrast, 4–32 billion parameter reasoning models required only a fraction of the resources and achieved inference times of 4–8 s. Although these smaller models performed moderately well, the performance gap indicates that advanced reasoning models currently remain resource-intensive for large-scale operational use.
Third, as an administrative data source, the casework reports are subject to limitations inherent to practitioner-generated documentation. Although reports are digitally created—reducing risks associated with paper-based entry or optical character recognition—data quality variability remains a meaningful constraint. Caseworkers may differ in their documentation practices: more experienced practitioners may produce more nuanced and comprehensive accounts of parental cooperation, while those managing higher caseloads or with less experience may generate more cursory or formulaic entries. Furthermore, the level of detail documented about cooperation may be influenced by case complexity, intervention phase, and the degree to which cooperation was considered clinically relevant at the time of writing. The corpus consists exclusively of 2-year summary accountability reports, which are written under accountability pressures and may foreground unresolved problems—including noncooperation—over periods of successful engagement; whether classification accuracy and prevalence estimates generalize to other documentation formats such as continuous case notes remains an open question. Future research should examine whether model performance varies systematically with documentation quality indicators such as report length, caseworker tenure, or agency-level documentation practices.
A broader conceptual challenge concerns the ethical implications of labeling parents as noncooperative using a definition grounded in professional compliance. Automated classification can only reflect what is documented, and documentation may not reliably distinguish structural inability from deliberate resistance. The operational definition used in this study was designed to capture documented behavioral patterns rather than underlying motivation, and it cannot determine at the level of individual cases whether noncooperation reflects deliberate unwillingness or structural constraint; practitioners interpreting model outputs should therefore treat classifications as hypotheses requiring contextual judgment rather than as verdicts about parental intent. The consequences of such labeling for case trajectories, intervention decisions, and parental rights warrant critical attention, and any operational deployment of this classification system should be accompanied by safeguards ensuring that automated outputs are interpreted in context by qualified practitioners rather than used as standalone assessments.
For social work practice, the validated RLM workflow offers a feasible pathway to generating population-level data on parental cooperation—a construct currently unmeasured at this scale in most child welfare systems. Because models can be deployed locally within agency infrastructure, child protection authorities can analyze their own documentation without having to transmit sensitive records to external servers, directly addressing the data protection constraints that have historically limited AI adoption in child welfare settings. Agencies considering an RLM deployment should note that smaller reasoning models (4B parameters) achieved classification accuracy comparable to systems many times their size, substantially reducing hardware requirements and making a local deployment feasible for agencies without high-performance computing infrastructure. Beyond its analytic utility, the gender disparity in classification accuracy has a direct practice implication: the finding that both EHRs and reasoning models assess fathers’ cooperation less accurately than mothers’ warrants critical reflection from frontline practitioners and supervisors. Child protection agencies should examine whether current documentation templates and reporting expectations adequately capture fathers’ cooperation trajectories across the full course of intervention, and whether training programs adequately support caseworkers in engaging with and documenting fathers as full participants in child protection processes rather than peripheral figures.
At the policy level, the finding that 31% of CPS cases in the Canton of Zurich involved documented noncooperation by at least one parent provides a baseline prevalence figure relevant to resource allocation, staff training priorities, and intervention planning. Comparable analyses with data from other Swiss child protection providers or CPS data from international jurisdictions would enable cross-system comparisons of cooperation patterns, documentation practices, and their relationship to intervention outcomes. Policymakers and agency leaders should also consider how existing documentation standards and reporting templates may systematically disadvantage fathers, and whether structured quality improvement initiatives can address these asymmetries. Future research should examine whether RLM classification accuracy varies with documentation quality indicators such as report length or caseworker tenure. The operational definition of parental cooperation used in this study was developed through iterative expert refinement; future work should incorporate broader stakeholder input—including input from parents with lived experience of CPSs—to ensure that automated classification reflects ethically grounded and practice-informed constructs.
The integration of advanced reasoning models into child welfare research holds promise for supporting more comprehensive, transparent, and evidence-informed assessments. Continued methodological refinement, combined with attention to gendered and structural biases in documentation, will be essential for ensuring that these tools contribute meaningfully to understanding and improving CPS practice.
Footnotes
Author Contributions
Dragan Stoll: Funding acquisition, project administration, conceptualization, methodology, software, data curation, validation, formal analysis, investigation, visualization, writing—original draft, reviewing and editing. Andreas Jud: Supervision, conceptualization, methodology, writing—reviewing and editing. Bryan E. Perron: Conceptualization, methodology, writing—reviewing and editing. Zia Qi: Conceptualization, methodology, writing—reviewing and editing. Selina Steinmann: Conceptualization, methodology, validation. Nicole F. Eicher: Conceptualization, methodology, validation.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: Our research was primarily funded by the Zurich Higher Education Institutions (DIZH) which aims to advance research and innovation on digitalization by using interdisciplinary approaches.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.