CARE-SD: Explainable Reasoning Architecture for Local SLM Self-Disclosure Analysis in Textual Data

I. INTRODUCTION

• Interview analysis can capture personal security incidents, behaviors, privacy concerns, and trust issues that technical assessments may overlook.These insights are relevant to security awareness, organizational policy, and privacy-conscious system design.Source: §I · p. 1
• Context-dependent responses may combine overlapping disclosure categories or discuss hypothetical risks, making keyword detection insufficient and potentially mischaracterizing participants.The categories are C1 security experience, C2 security/privacy behavior, and C3 security/privacy concern; earlier taxonomies do not establish accurate multi-label reasoning for these interviews.Source: §I · p. 1
• Fluent model explanations may attribute hypothetical incidents or unsupported motives to participants, so claims and explanations must be checked against transcript evidence.Prior local-LLM work examined qualitative privacy analysis and hallucination mitigation, but these risks remain relevant.Source: §I · p. 1
• Local SLMs keep sensitive transcripts in researcher-controlled environments, but limited capacity and entangled prompting can impair contextual reasoning, multi-label decisions, and explanation fidelity.Chain-of-thought and iterative refinement do not guarantee faithful evidence selection, and explanation quality requires independent assessment.Source: §I · p. 1
• The study asks whether CARE-SD improves detection and classification over conventional classifiers and standard prompting, and whether it reduces unsupported interpretations while narrowing the gap to larger models.The work also evaluates explanation accuracy.Source: §I · pp. 1–2
• CARE-SD integrates context-adaptive batch processing, hierarchical disclosure reasoning, explanation-consistency verification, and cross-batch reconciliation for local SLM analysis.It separates model interpretation from deterministic context selection, validation, and aggregation, enabling independently assessable predictions and explanations.Source: §I · p. 2

II. RELATED WORK

• Prior detection research covers other domains and evaluation units, while extracted spans alone do not establish the correctness of generated interpretations,,,.The cited work includes a 19-category taxonomy and contextual extraction of rationales.Source: §II · p. 2
• Privacy-sensitive text detection, disclosure support, and rewriting address distinct tasks from checking whether locally deployed SLM explanations faithfully represent interview evidence.Related work includes human–LLM disclosures [29], context-aware privacy assistance,, and locally deployable reformulation.Source: §II · p. 2
• CARE-SD's distinct contribution is integrating context-selected inputs, hierarchical multi-label decisions, quotation and explanation checks, and cross-window reconciliation, assessed for accuracy and claim-level reliability.The paper does not claim the interview setting or underlying techniques as new.Source: §II · p. 2

III. PROPOSED CARE-SD ARCHITECTURE

• CARE-SD separates SLM semantic interpretation from deterministic context selection, validation, and aggregation across four modules.The architecture combines context-adaptive batch processing, hierarchical disclosure reasoning, explainability and consistency verification, and cross-batch reconciliation.Source: §III · p. 2, Figure 1 · p. 3

A. Architectural Modules

• M1 processes sentence-aligned transcript windows, initially using the participant response and interviewer question, and expands context when dependency cues or validation failures require it.Windows use 384 tokens with 64-token overlap; expansion can include the preceding two question–answer pairs, within a 2048-token input budget.Source: A. Architectural Modules · p. 2
• M2 first detects disclosure, then independently classifies C1–C3, allowing multiple positive categories and assigning C0 only when all categories are negative.This decomposes a single multi-label decision into binary detection followed by category classification.Source: A. Architectural Modules · p. 2
• M3 requires a verbatim participant-response quotation for each positive category and checks output structure, labels, category consistency, and explanation length.A failed validation triggers at most one corrective re-analysis with expanded context; outputs still invalid are flagged rather than labeled C0.Source: A. Architectural Modules · pp. 2–3
• M4 reconciles overlapping windows by participant identifiers and text offsets, removes duplicate quotations, and consolidates valid predictions per response.Responses with no valid predictions are flagged rather than automatically classified as nondisclosures.Source: A. Architectural Modules · p. 3

B. Algorithmic Implementation

• The SLM performs semantic detection and classification, while the controller explicitly controls workflow decisions and all models share the same processing parameters and task definitions.The controller sets segmentation, contextual triggers, inference order, validation conditions, retry limits, and aggregation.Source: B. Algorithmic Implementation · p. 3
• The controller segments transcripts, attaches questions and identifiers, selects context, invokes inference, retries when triggered, flags invalid outputs, and reconciles windows.Segmentation uses W = 384, O = 64, and B = 2048; context expansion is triggered by a cue or invalid output.Source: B. Algorithmic Implementation · p. 3
• Context expansion and retry decisions are controller-executed rather than generated by the SLM.The common user prompt supplies the segment identifier, interviewer question, selected preceding context, and participant response.Source: B. Algorithmic Implementation · p. 3

C. System and User Prompt Specification

• The detection prompt asks whether the participant personally disclosed an experience, behavior, or concern, excluding general observations and hypothetical statements.It returns a Boolean disclosure decision in JSON.Source: C. System and User Prompt Specification · pp. 3–4
• The dataset table is described as reporting dataset statistics and participant-disjoint partitions.The caption does not provide the statistics or partition values.Source: Table I · p. 4
• The classification prompt defines C1–C3, permits multiple categories, requires a verbatim quotation and concise explanation for each positive category, and prohibits invented experiences or motives.Explanations are limited to at most 20 words.Source: C. System and User Prompt Specification · p. 4
• The common user template includes a segment identifier, interviewer question, controller-selected preceding context, and target participant response.Prompt instructions, token limits, and correction rules remained fixed across evaluated SLMs.Source: C. System and User Prompt Specification · p. 4

• The evaluation tested whether CARE-SD improved disclosure-category accuracy and reduced unsupported interpretation.Source: §IV · p. 4
• The study combined independently annotated interviews with conventional and prompting baselines, larger-model references, and architectural ablations.Source: §IV · p. 4
• Blinded output review was included in the evaluation.Source: §IV · p. 4

A. Dataset and Human Gold Standard

• Two reviewers annotated 82 interviews and established a consensus gold standard for three potentially overlapping disclosure categories; C0 denotes no disclosure.The held-out test set contained 12 interviews and 229 responses, with all responses from each participant assigned to one partition.Source: A. Dataset and Human Gold Standard · p. 4
• Blinded reviewers assessed generated factual claims against transcripts and judged whether explanations justified the assigned categories.Invalid predictions and missing or invalid required explanations remained in evaluation and were scored according to the stated rules.Source: A. Dataset and Human Gold Standard · p. 4

B. Baselines and Reference Models

• The study compared conventional SVM and DeBERTa baselines, three local SLMs under four prompting or architecture approaches, and three larger references.Source: B. Baselines and Reference Models · p. 4

C. Inference and Quantization

• SLMs ran on a local laptop while larger models ran on a self-hosted AWS instance; hardware-dependent efficiency metrics were excluded.Dense generative models used 4-bit GGUF Q4_K_M weights, while gpt-oss-20b used released native MXFP4 weights.Source: C. Inference and Quantization · p. 4

D. Evaluation Matrix and Protocol

• Macro-F1 averaged binary F1 across C1–C3, while detection F1 treated C1–C3 as disclosure and C0 as no disclosure.Source: D. Evaluation Matrix and Protocol · p. 4, §V · p. 5
• Point estimates used participant-disjoint splits, but confidence intervals were omitted and differences were presented descriptively rather than as significance or equivalence findings.The response-level paired predictions, participant identifiers, and claim adjudications needed to rerun the bootstrap were unavailable in the result records.Source: D. Evaluation Matrix and Protocol · p. 5

E. Experimental Comparisons

• The experiments assessed prompting gains, unsupported claims, module contributions, model scale, and human-judged explanation correctness.E1 compared prompting strategies, E2 measured hallucination separately, E3 ablated components, E4 compared larger references, and E5 assessed explanations.Source: E. Experimental Comparisons · p. 5

A. E1: Architecture Effectiveness

• Macro-F1 rose from 0.72 to 0.84 for Llama 1B, 0.75 to 0.85 for Qwen 1.5B, and 0.80 to 0.88 for Llama 3B versus direct prompting.CARE-SD also exceeded DeBERTa-v3-base (Macro-F1 = 0.79) for all three SLMs.Source: A. E1: Architecture Effectiveness · p. 5

B. E2: Hallucination Reduction

• UCR fell from 0.21 to 0.10 for Llama 1B, 0.19 to 0.09 for Qwen 1.5B, and 0.15 to 0.06 for Llama 3B under CARE-SD.Two blinded reviewers classified factual claims against source transcripts; E2 measured unsupported and contradictory claims.Source: B. E2: Hallucination Reduction · p. 5, D. E4: Larger-Model Comparison · p. 6

C. E3: Architectural Ablation

• Removing M1–M4 reduced Llama 3B Macro-F1 by 4, 5, 3, and 2 points, respectively; removing M3 raised UCR from 0.06 to 0.15.E3 removed each module individually with the SLM and test set fixed; the ablation did not isolate verification-operation effects or module interactions.Source: C. E3: Architectural Ablation · p. 5, D. E4: Larger-Model Comparison · p. 6

D. E4: Larger-Model Comparison

• Llama 3B + CARE-SD achieved Macro-F1 = 0.88, matching gpt-oss 20B, exceeding Qwen3 8B (0.87), and trailing Qwen3 32B (0.89) by one percentage point.Its UCR of 0.06 equaled Qwen3 32B and was lower than Qwen3 8B (0.08) and gpt-oss 20B (0.07); the numerical agreement does not establish statistical equivalence.Source: D. E4: Larger-Model Comparison · pp. 5–6, Figure 2 · p. 5

E. E5: Explanation Correctness

• ECR rose from 0.75 to 0.88 for Llama 1B, 0.78 to 0.90 for Qwen 1.5B, and 0.84 to 0.93 for Llama 3B under CARE-SD.Each comparison is against direct prompting.Source: E. E5: Explanation Correctness · p. 6
• Under CARE-SD, Llama 3B’s human-assessed ECR rose from 0.84 with direct prompting to 0.93, matching the gpt-oss 20B reference (0.93) and approaching Qwen3 32B (0.94).Source: E. E5: Explanation Correctness · p. 6
• M3’s deterministic verification checked structural and quotation consistency, while human reviewers independently assessed semantic correctness.The ECR results complemented reductions in unsupported claims.Source: E. E5: Explanation Correctness · p. 6

F. Discussion and Limitations

• Findings apply to 82 cybersecurity-oriented interviews; the held-out test contained only 12 interviews, and sector sizes were unequal.No sector-specific scores were available from the supplied aggregates.Source: F. Discussion and Limitations · p. 6
• The aggregates lacked reproducible confidence intervals and claim-level or explanation-count denominators, limiting independent audit of UCR, CR, and ECR.Flagged-output counts were also unavailable.Source: F. Discussion and Limitations · p. 6
• Llama 3B’s numerical Macro-F1 equality with gpt-oss 20B does not establish statistical equivalence.Source: F. Discussion and Limitations · p. 6
• The joint accuracy–factuality comparison was not treated as a Pareto frontier because some reference configurations dominated others on both plotted metrics.Source: F. Discussion and Limitations · p. 6

VI. CONCLUSION

• Llama 3B’s category Macro-F1 increased from 0.80 to 0.88, while its unsupported claim rate fell from 0.15 to 0.06, a 60% reduction.Source: §VI · p. 6
• Llama 3B achieved detection F1 of 0.90 and explanation correctness of 0.93, numerically matching the 20B reference model’s Macro-F1.Source: §VI · pp. 6–7
• CARE-SD integrates context-adaptive batch processing, hierarchical disclosure reasoning, explainability and consistency verification, and cross-batch reconciliation without modifying model parameters.Source: §VI · pp. 6–7
• Component ablations identified M2 as the largest classification contributor and M3 as most influential for reducing unsupported claims.Each ablation removed one module while retaining the remaining components.Source: Table IV · p. 7, §VI · p. 7
• The authors identify larger, independent, and more diverse disclosure datasets as future evaluation needed to establish cross-domain generalizability and robustness.Source: §VI · p. 7