PaperlayerView exact revision on arXiv

Source-linked AI summary

AI for Computational Design Science: A Responsible Human-AI Framework and Case Study on Short-Form Video Safety Surveillance

Wenli Zhang, Jiaheng Xie, Zhihe Pan, Yidong Chai, Xiao Fang, Sudha Ram

arXiv:2609.05270v1cs.AI

TL;DR

Prior CDS work does not specify how AI participation should be organized across the lifecycle, while generic safety signals cannot stand in for child appropriateness. The paper presents AI4CDS to organize and govern AI participation in CDS; its proposed architecture satisfies four axioms.

  • Problem

    Generic safety signals cannot serve as proxies for child appropriateness in short-form video detection.

    Source: §2.1 · p. 7

  • Method

    AI4CDS organizes and governs AI participation in CDS, including a phase in which AI explores resources.

    Source: §1 · p. 6, §3 · p. 14

  • Results

    The proposed architecture satisfies the four axioms.

    Source: §2.4 · p. 14

  • Takeaways & Limitations

    A child-oriented detector must preserve useful generic-safety information while identifying additional risk specific to children.

    Source: §2.1 · p. 7

    Prior CDS literature does not specify how AI participation should be organized across the CDS lifecycle.

    Source: §1 · p. 3

Abstract

from arXiv · show

Artificial intelligence (AI) is transforming not only what information systems researchers design, but also how design research is conducted. Yet existing literature offers limited guidance for computational design science (CDS) when AI actively participates in problem formulation, resource construction, design search, evaluation, and knowledge abstraction. We develop AI for Computational Design Science (AI4CDS), a five-phase methodological framework in which AI expands problem and design search while researchers retain responsibility for domain grounding, admissibility, verification, and scientific judgment. Collaboration is governed by graduated trust, reversibility, auditability, and differentiated reproducibility. We instantiate AI4CDS through ChildRiskGuard, an interpretable artifact for detecting short-form videos inappropriate for children, while documenting AI interactions, rejected alternatives, corrections, and audit trails. The case translates audience-dependent safety and explanation faithfulness into three technical challenges and develops an artifact that separates generic from child-specific risk, represents distinct developmental-risk mechanisms, and makes concept-level explanations part of the predictive computation. ChildRiskGuard achieves an F1 score of 0.769, substantially outperforming direct application of a general-purpose content-safety model while remaining competitive with strong benchmarks. The primary contribution is AI4CDS as a responsible framework for AI-enabled CDS; ChildRiskGuard provides process and artifact evidence of how AI-expanded, researcher-governed design can generate and evaluate novel computational design knowledge.

1. Introduction

The introduction proposes AI4CDS, a five-phase framework for governing AI’s participation across computational design science while researchers retain scientific authority and accountability. It introduces ChildRiskGuard as a consequential case for translating child-oriented video-safety needs into computational design requirements.

1. Introduction

  • Across five phases, AI supports problem formulation, resource construction, design search, evaluation, and knowledge abstraction, while researchers make and verify consequential design decisions.Graduated trust, reversibility, auditability, and differentiated reproducibility govern collaboration across the lifecycle.Source: §1 · pp. 3–4
  • ChildRiskGuard is an interpretable short-form-video safety system that instantiates AI4CDS and demonstrates how domain knowledge can guide computational design.The case translates developmental considerations into requirements for harmful-content detection and explanation.Source: §1 · pp. 5–6
  • Child-oriented video safety exposes gaps in treating safety independently of audience and producing explanations separately from predictions, motivating three technical challenges.The challenges are preserving generic safety while isolating child-specific risk, representing heterogeneous developmental-risk mechanisms, and ensuring concept-level faithfulness.Source: §1 · p. 5
  • AI4CDS is the primary contribution, organizing AI participation across the research lifecycle while researchers retain authority over domain grounding, admissibility, verification, and scientific accountability.ChildRiskGuard provides the artifact and process evidence through which the framework is instantiated and examined.Source: §1 · p. 6

2. Related Work

The review frames child-oriented moderation as a design problem: existing moderation can miss audience-specific and developmental risks, while concept-based prediction alone does not ensure faithful explanations. It motivates distinguishing general safety from child-specific risk, representing distinct developmental mechanisms, and tying explanations faithfully to decisions.

2. Related Work

  • The review uses AI to broaden literature retrieval, while researchers verify and synthesize evidence to define the problem and translate domain requirements into technical challenges.The review can be reopened when emerging requirements expose missing knowledge.Source: §2 · p. 6

2.1. Audience-Specific Video Safety and Inappropriate-Video Detection

  • General moderation cannot proxy for child appropriateness: a detector must preserve generic-safety information while identifying additional child-specific risk, not merely lower its threshold.Existing methods often define risk from content without representing the audience for whom it is harmful.Source: §2.1 · p. 7

2.2. Interpretable Video Prediction

  • Concept-based prediction is favored because named concepts can represent developmental mechanisms and participate in the decision, but their presence alone does not guarantee faithful explanations.A concept may capture information unrelated to its intended meaning or additional information about the target label.Source: §2.2 · p. 9

2.3. Concept Bottleneck Models

  • Concept bottleneck models place named concepts on the predictive pathway; here, external developmental knowledge defines child-risk constructs, while computational models measure evidence for them.This separates what the artifact represents from how evidence for each construct is measured.Source: §2.3 · pp. 9–10
  • ChildRiskGuard uses externally grounded mechanisms and concept measurements, learns from overall child-appropriateness labels without mechanism supervision, and computes their contributions within the final prediction.This approach addresses the cost of collecting sample-level labels for each developmental-risk mechanism.Source: §2.3 · p. 11

2.4. Research Gaps and Design Challenges

  • Existing approaches often treat harm as content-only and may not ensure that interpretable concepts capture intended evidence or faithfully explain predictions.The section calls for separating generic safety risk from additional child-specific risk, with developmental mechanisms directly participating in prediction.Source: §2.4 · p. 12
  • The child-specific pathway uses generic safety as an explicit reference and must capture incremental child risk without offsetting, miscalibrating, or duplicating generic risk.The generic signal may require calibration for child-oriented moderation, while the child-specific component should avoid reproducing information already represented by that signal.Source: §2.4 · pp. 12–13
  • The artifact must separately represent heterogeneous, differently evidenced and unequally frequent developmental mechanisms, learn them from overall labels, and remain modular for expansion.The primary model uses mechanisms involving imitable dangerous actions, sexualized or adult-like performance, and self-harm or depression-related content.Source: §2.4 · p. 13
  • Faithfulness requires limiting semantic leakage from concept scores while balancing interpretable attribution against the flexibility needed for thresholding or saturation.A simple linear mapping can be too restrictive, while an unconstrained nonlinear mapping can make contributions difficult to identify.Source: §2.4 · pp. 13–14
  • The proposed architecture satisfies four mapping axioms: completeness, sign consistency, unique attribution, and non-masking.These require contributions to reconstruct mechanism risk, follow theoretically specified directions, map to specific concepts and mechanisms, and prevent positive child-specific risk from being canceled elsewhere.Source: §2.4 · p. 14

3. Research Design

The research design uses AI to expand resource and design search while researchers validate measurements, assess candidate designs, and retain or reject alternatives. It operationalizes child-oriented video moderation by distinguishing generic safety from child-specific risks, encoding those risks through domain-grounded mechanisms and gates, and constraining semantic evidence to preserve interpretable component contributions.

3. Research Design

  • AI explores resources and generates design alternatives, while researchers test construct validity, inspect measurements, set training-data normalization and thresholds, and evaluate alternatives before selection.Researchers retain, revise, or reject candidates and record consequential decisions; this process yields identifiable residual learning, mechanism gating, semantic budgeting, and faithful additive hazard mapping.Source: §3 · p. 14

3.1. Computational Problem Formulation

  • The formulation retains useful generic-safety capabilities while explicitly and interpretably learning additional child-specific risks, distinguishing the two and identifying concepts contributing to the child-specific assessment.The problem requires both identifying videos inappropriate for children and explaining their underlying risks; child appropriateness overlaps with but differs from generic content safety.Source: §3.1 · pp. 14–15

3.2. ChildRiskGuard

  • ChildRiskGuard is a mechanism-aligned concept-bottleneck additive-hazard model with a frozen measurement stage and a trainable child-oriented risk-assessment stage.Its design coordinates identifiable residual learning, mechanism gating, semantic budgeting, and faithful additive hazard mapping.Source: §3.2 · pp. 16–17

3.2.1. Design Element 1: Identifiable Child-Specific Residual Learning

  • Under noisy-OR composition, the exponential hazard link is unique up to positive rescaling among continuous, strictly increasing links that support additive decomposition.The result establishes the link from the requirements that risk quantities add on the hazard scale while probabilities compose as noisy-OR.Source: §3.2.1 · p. 18
  • Under stated block-separability, zero-anchoring, and joint-support conditions, the residual decomposition uniquely assigns components and discourages child-specific mechanisms from duplicating generic-safety evidence.The separation penalty is a soft diagnostic and does not require strict independence.Source: §3.2.1 · pp. 19–20

3.2.2. Design Element 2: Knowledge-Anchored Mechanism Structuring and Differentiable Gating

  • Three domain-grounded child-risk mechanisms use fixed, non-overlapping concept blocks, with evidence families encoding substitutable cues within families and conjunctive requirements across families.A continuous gate is used during both training and inference.Source: §3.2.2 · pp. 20–21

3.2.3. Design Element 3: Semantic Contribution Budget

  • The semantic contribution budget limits VLM-derived evidence relative to deterministic evidence, and mechanism-level budgets remain meaningful after gating and aggregation.The primary model includes per-concept caps, a soft training penalty, and inference-time projection.Source: §3.2.3 · pp. 22–23

Appendix B.3.3.

ChildRiskGuard’s retained design combines four elements addressing three technical challenges, while its monotone additive mapping makes child-specific faithfulness structural and exposes prediction decompositions. The appendix also limits the semantic contribution budget to the mechanism-intensity pathway.

Appendix B.3.3.

  • The semantic contribution budget applies only to the mechanism-intensity pathway; activation-family effects are governed separately by the knowledge-anchored gate, while contextual attenuation concepts affect only generic safety.Source: Appendix B.3.3 · p. 23

3.2.4. Design Element 4: Axiomatically Faithful Monotone Additive Hazard Mapping

  • The child-specific pathway satisfies A1–A4 as structural artifact properties, and the same forward pass yields generic-safety, mechanism-level, and concept-level decompositions without a separate post-hoc step.The mapping retains nonlinear concept–risk relationships while keeping realized contributions explicit.Source: §3.2.4 · pp. 23, 24, 26
  • With semantic projection inactive, a concept contribution equals its Shapley value conditional on the realized mechanism gate; with projection active, reconstruction and monotonicity remain, but unconditional Shapley equivalence is not claimed.The exact attribution result applies to the additive intensity game.Source: §3.2.4 · p. 24

3.3. Design Novelty of ChildRiskGuard within AI4CDS

  • ChildRiskGuard combines identifiable residual learning, knowledge-anchored mechanisms, a semantic contribution budget, and monotone additive mapping to address three technical challenges.These elements separate generic from child-specific risk, represent heterogeneous developmental risks, constrain input-side faithfulness risks, and preserve identifiable concept-level attribution.Source: §3.3 · p. 26
  • The four retained elements emerged from a broader design search rather than being specified ex ante, with alternatives accepted, modified, reclassified, or rejected before final specification.Source: §3.3 · p. 26

4. Evaluations

The evaluation assesses ChildRiskGuard’s predictive utility and the contribution of its structured risk mechanisms through benchmark comparisons, ablations, robustness and generalizability checks, and verified interpretability analyses. It examines whether explanations reflect the model’s computation while distinguishing this internal faithfulness from evidence of real-world harm causation; AI-generated evaluation analyses were reviewed by researchers.

4.1. Data and Evaluation Design

  • The evaluation covers 7,070 videos, using 6,010 for development and an untouched 1,060-video test set that includes 150 positive cases.The videos combine expert-annotated TikGuard examples and additional manually reviewed TikTok videos; fitting and threshold selection use development data only.Source: §4.1 · pp. 26–27

4.2. ChildRiskGuard Prediction Utility

  • F1 = 0.769, exceeding Random Forest (0.593), LM4CV (0.758), and frame-wise ResNet-18 (0.711), indicating value beyond frozen measurements, conventional concept models, and learned representations.The comparison supports the value of structuring measurements around domain-specific requirements, rather than relying on the measurements or task-specific representation learning alone.Source: §4.2 · pp. 27–28, Table 3 · p. 28
  • ChildRiskGuard raises F1 from 0.632 for frozen LlavaGuard to 0.769, but the results do not establish universal superiority over foundation models.InternVL3 reaches 0.758 and fine-tuned LlavaGuard 0.679; unlike these baselines, ChildRiskGuard provides structured attribution without sample-level mechanism annotations.Source: Table 3 · p. 28, §4.2 · p. 28

4.3. Ablation, Sensitivity, and Generalizability Analyses

  • Removing the calibrated generic channel and monotone additive experts reduces F1 by 0.085 and 0.077, respectively; the semantic budget adds 0.012 and mechanism gating 0.006.All four design innovations improve performance when each focal element is removed while the rest of the architecture is held constant.Source: Table 4 · p. 29, §4.3 · p. 29
  • Adding M1, M2, and M3 increases F1 by 0.035, 0.011, and 0.150, respectively; M3, the self-harm and depressive-risk mechanism, contributes the largest incremental gain.The ablations show that M1–M3 make non-redundant predictive contributions in the focal dataset; the architecture can accommodate M4 with mechanism-specific measurements without redesigning the core risk-assessment architecture.Source: §4.3 · p. 29

4.4. Interpretable Insights: Faithfulness and Mechanism-Level Interpretation

  • With only video-level binary training labels, ChildRiskGuard produces mechanism- and concept-level attributions within the same forward pass that reconstructs the child-specific risk score.These attributions are computed as part of the prediction, without requiring additional labels for individual mechanisms or concepts.Source: §4.4 · p. 30
  • For a dangerous fire-performance video, ChildRiskGuard predicts 0.738 versus LlavaGuard’s 0.221, and removing M1 lowers the prediction to 0.412 and flips the decision to safe.The M1 risky-action concept contributes 0.702 hazard, and the model’s 0.738 prediction exceeds the 0.612 threshold.Source: §4.4 · pp. 30–31
  • For an adultized performance with skin exposure, ChildRiskGuard assigns 0.707 while LlavaGuard classifies it as safe, combining exposure with contextual evidence of adultized or sexualized performance.The generic and M2 channels contribute hazards of 0.938 and 0.286, respectively; the explanation uses contextual signals rather than exposed skin alone.Source: §4.4 · pp. 30–31
  • Removing M2 lowers the probability from 0.707 to 0.610, below the operating threshold, and reverses the final decision.Source: §4.4 · p. 31
  • 0.860 for ChildRiskGuard versus 0.109 for LlavaGuard; removing M3 lowers ChildRiskGuard’s prediction to 0.377 and reverses its classification.M3 contributes a hazard score of 1.493, above the generic-channel hazard of 0.454, with evidence from self-harm semantics, darkness, and depressive affect.Source: §4.4 · p. 31
  • Across all three cases, removing the focal mechanism reverses the final decision, indicating that explanations reflect the model’s decision process but not that those mechanisms cause real-world harm.Substantive interpretation of the mechanisms depends on external domain grounding.Source: §4.4 · p. 32
  • AI-assisted evaluation produced invalid analyses and premature interpretations that researchers removed, revised, corrected, or rejected after review.Examples included using test data to select an operating point, misdescribing the M2 gate, and recommending mechanism removal from one split; these errors created verification burdens when outputs became evidence or claims.Source: §4.4 · p. 32

5. Discussion and Conclusion … A.1 Phase 1: AI-Expanded Research Problem Formulation

This section presents AI4CDS as a researcher-governed framework for AI participation in computational design science, with ChildRiskGuard as an instantiation. It connects the literature-grounded problem, domain requirements, and technical challenges to governance practices, domain implications, and boundaries for the framework and artifact.

5. Discussion and Conclusion

  • Phase 5 assembles cross-phase records, traces how requirements and evaluation evidence shaped design, and uses AI to surface patterns that researchers test before distilling transferable, theory-grounded knowledge.The process also yields methodological implications, governance principles, reproducibility guidance, domain implications, and boundary conditions.Source: §5 · p. 32

5.1. AI4CDS as a Prescriptive Methodological Framework

  • AI4CDS prescribes phase-contingent AI discretion across the research lifecycle while researchers retain authority over domain grounding, admissibility, verification, and scientific commitments.It makes roles, transitions, and responsibilities explicit so AI-enabled CDS processes can be examined, documented, and refined.Source: §5.1 · pp. 32–33

5.2. Constitutive AI in CDS: Breadth, Depth, and Artifact Development

  • AI materially expanded the breadth, depth, and feasible scope of ChildRiskGuard’s development within the available research window, while its outputs required researcher verification, revision, or rejection.The collaboration was most valuable when AI expanded inquiry while researchers retained authority over scientific commitments.Source: §5.2 · pp. 33–34

5.3. Governance Principles for AI-Enabled CDS

  • AI4CDS addresses uncertain reliance, path-dependent commitment, and weakened provenance through governance principles and differentiated reproducibility, reversibility, and auditability.Governance effort should scale with scientific consequence and propagation risk.Source: §5.3 · p. 35

5.4. ChildRiskGuard as Consequential Instantiation and Domain Implications

  • ChildRiskGuard separates generic safety from child-specific risk and represents developmental-risk mechanisms so outputs can support human review, escalation, and auditing.The model estimates computationally defined risks and does not establish causal harm to individual children or independently determine policy or intervention.Source: §5.4 · p. 36
  • ChildRiskGuard substantially improves on the reused generic-safety model and performs competitively against strong foundation-model alternatives, without claiming universal predictive superiority.Its domain-grounded design combines competitive prediction with structured, faithful attribution.Source: §5.4 · p. 36

5.5. Boundary Conditions and Future CDS Research

  • AI4CDS is supported by one intensive project and does not establish that its phases are optimal or that it reduces effort or produces superior artifacts; ChildRiskGuard’s evaluation is limited to one setting and a bounded mechanism set.Further work should compare AI4CDS across research processes and test artifact generalization beyond the evaluated setting.Source: §5.5 · pp. 36–37

Trail

  • The appendix explains how generative AI was used in the research process that produced ChildRiskGuard and complements the technical specification in Appendix B.Source: Trail · p. 42
  • Researchers retained authority over problem definition, admissible sources and designs, domain validity, evaluation, research integrity, and interpretation, while AI supported search, implementation, and diagnostics.Consequential AI outputs remained subject to researcher verification and authorization.Source: Trail · p. 42

A.1 Phase 1: AI-Expanded Research Problem Formulation

  • The review covered descriptive knowledge on child development and media safety and prescriptive knowledge on detection and interpretable prediction, with researchers verifying sources and making final scope decisions.AI supported retrieval and synthesis, organized evidence, questioned draft formulations, and discussed alternatives.Source: §A.1 · p. 43
  • AI helped identify field terminology and benchmarks for bibliographic queries, while unsupported gaps or evolving design choices prompted renewed literature retrieval.Searches later expanded from detection methods to interpretable video models, concept bottleneck models, and concept leakage.Source: §A.1 · pp. 43–45
  • Researchers defined the problem as detecting short videos inappropriate for children and specified users and uses supporting moderation or guardianship decisions.Phase 1 produced this problem definition, two literature gaps, domain requirements, and challenges guiding resource construction and design search.Source: §A.1 · pp. 44–45
  • Phase 1 derived three challenges: preserve general safety while detecting child-specific risks, represent heterogeneous developmental mechanisms, and make concept-level explanations faithful to evidence and prediction computation.The literature review found that appropriateness depends on audience, risks arise through distinct developmental mechanisms, and many explanations are separate from prediction.Source: §A.1 · p. 44
  • Verification was repeated when sources were selected, information extracted, and citations inserted, reducing reliance on a single initial check.Most corrections occurred when retrieved records were converted into summaries or manuscript claims; Table A1 describes verification errors and safeguards but is not exhaustive.Source: §A.1 · pp. 45–46, Table A1 · p. 46

A.2 Phase 2: AI-Augmented Data & Resource Construction

Phase 2 used AI to explore candidate concepts, measurements, and resources, while researchers governed their validity, structural roles, and admissibility. The resulting resource base defined fixed measurements and child-risk mechanisms for later design and prediction.

A.2 Phase 2: AI-Augmented Data & Resource Construction

  • The fixed measurement stage uses a frozen LlavaGuard safety reading for trainable calibration, fixed-prompt semantic scores from a frozen visual representation, and fixed deterministic extractors.These produce a frozen general-safety reading and normalized concept measurements.Source: §A.2 · p. 46
  • The three child-risk blocks combine distinct signals: M1 dangerous-action motion and semantics, M2 skin exposure, suggestive context, and adultized-performance concepts, and M3 darkness, self-harm recall, and depressive affect.These blocks represent imitable dangerous action, adultized performance, and self-harm or depressive risk, respectively.Source: §A.2 · p. 47
  • Mechanism activation is selective: M1 uses a motion-measurement family, M2 a conjunctive skin-exposure/context gate, and M3 darkness; other assigned concepts affect intensity without independently opening a gate.Source: §A.2 · p. 47
  • AI proposed concepts, prompts, extractors, and mechanism assignments; researchers judged developmental fit, stability, specificity, and each concept’s role before fitting risk parameters.Researchers decided whether concepts should affect activation, intensity, contextual attenuation, or none of these.Source: §A.2 · p. 47
  • Resource construction applied four checks: construct validity, measurement behavior, structural role, and separation of data partitions.Normalization statistics and fixed thresholds were estimated from the relevant training partition and held fixed outside it.Source: §A.2 · pp. 47–48
  • The design addresses input faithfulness through fixed concept definitions, structural assignments, semantic caps, and mechanism budgets, and mapping faithfulness through an anchored, monotone, additive hazard construction.Source: §A.2 · p. 48

A.3 Phase 3: AI-Generative Research Design Search & Refinement

AI broadened the design search, while researchers governed which candidates were admissible and retained. The process translated child-safety and explanation requirements into constrained, testable design choices and documented their evaluation.

A.3 Phase 3: AI-Generative Research Design Search & Refinement

  • The artifact had to preserve general-safety information while identifying child-specific risks, representing developmental mechanisms, and maintaining faithfulness from evidence to prediction.Mechanism labels had to be developmentally grounded; the method had to operate without concept-level annotations and leave original labels unchanged.Source: §A.3 · p. 48
  • The researchers translated challenges into testable requirements, generated component-level candidates, and specified an evaluation before selecting each candidate.Candidates were retained when an ablation, diagnostic test, or theoretical condition could assess them.Source: §A.3 · p. 49
  • Researchers critically reviewed claims and constructions, then documented candidate origins, decisions, and rationales to track how the final architecture developed.AI-supported critiques surfaced unsupported guarantees, incompatible requirements, and arguments misaligned with the architecture; researchers could narrow claims, revise constructions, or reject components.Source: §A.3 · p. 49
  • AI contributed more to generating candidates than selecting the final design, with researchers retaining responsibility for admissibility.Candidates were revised or rejected after theoretical review, measurement analysis, or development-data diagnostics.Source: §A.3 · p. 51

A.4 Phase 4: AI-Enabled Multi-Faceted Evaluation … B.3 Theoretical Results and Proofs

This section documents the researcher-governed evaluation and analysis of ChildRiskGuard, from frozen measurements, risk-model specification, training and thresholding to mechanism-share definitions and AI-assisted principle distillation. It also explains how evaluation evidence and the research record support distinguishing case-specific findings from transferable design knowledge, with interpretation and final wording retained by researchers.

A.4 Phase 4: AI-Enabled Multi-Faceted Evaluation

  • Phase 4 separated specification, execution, and interpretation, with researchers reviewing specifications and outputs before accepting evidence-based interpretations.This separation made discrepancies among the designed method, implemented code, and reported claims easier to identify.Source: §A.4 · p. 51
  • The evaluation checks set rules in advance, isolated focal changes, reconciled code with design, compared benchmarks under comparable rules, and kept interpretation under author responsibility.Test labels were excluded from model selection, and AI-generated explanations were treated as proposals.Source: §A.4 · p. 52
  • The evaluation suggests human review is especially important at specification-to-execution and execution-to-interpretation handovers, where safeguards addressed analyses that could make results appear stronger.Examples included test-informed operating points, broad abstention, and unqualified point estimates.Source: §A.4 · p. 53

A.5 Phase 5: AI-Assisted Design Principle Distillation

  • Phase 5 used the Phase 1–4 record and evaluation evidence to distinguish case-specific findings from transferable design knowledge, with researchers verifying AI-supported organization and candidate patterns.The record preserved consequential decisions, alternatives, corrections, and verification evidence rather than being reconstructed only at project end.Source: §A.5 · p. 54
  • Researchers tested candidate abstractions against the record for case-specific dependence, contradictions, and evidential support, rejecting unsupported claims and narrowing partially supported ones.Surviving patterns could be documented as transferable design or methodological insights, governance or reproducibility principles, or boundary conditions, using established theory to explain mechanisms and scope.Source: §A.5 · p. 55
  • AI supported organizing and refining candidate principles, but researchers retained substantive interpretation, abstraction level, and final wording, linking claims to theory and cross-phase evidence.Source: §A.5 · p. 55

B.1.1 Frozen Video and Generic-Safety Measurement

  • LlavaGuard’s Safe/Unsafe scores are max-pooled across sampled frames into a generic-safety reading, not a child-inappropriateness probability.Only the risk-assessment stage is optimized; the measurement stage remains frozen.Source: §B.1.1 · p. 56, Trainable and frozen components · p. 61

B.1.2 Semantic Concept Measurement

  • Semantic concepts use a normalized mean video embedding and a frozen prompt-set scorer, unlike the LlavaGuard judge’s temporal maximum pooling.Broad concepts use positive-versus-negative prompt scores, while adultized-performance concepts use the maximum positive-prompt cosine score.Source: §B.1.2 · pp. 56–57

B.1.3 Deterministic Concept Measurement

  • The deterministic inventory covers motion and pose, skin exposure, and darkness; concepts are quantile-normalized using training data, with the fitted map frozen for held-out and test data.Short duration and stillness are excluded from the final model.Source: §B.1.3 · p. 57, §B.1.4 · p. 57

B.1.5 Concept Inventory and Fixed Structural Assignments

  • Fixed assignments distinguish risky action (M1), suggestive and adultized performance (M2), and self-harm or depressive cues (M3); contextual attenuation operates only on generic safety.M1 combines motion/pose concepts with risky-action semantics; M2 combines skin exposure, suggestive context, and five adultized-performance concepts; M3 combines darkness, self-harm recall, and depressive affect.Source: §B.1.5 · pp. 57–58

Anchored monotone response basis.

  • Anchored monotone response functions use nonnegative slopes to allow nonlinear threshold-like growth and saturation while remaining bounded and zero-anchored.The final implementation uses eight knots and a maximum per-concept response of 8.0 before a tighter semantic cap.Source: Anchored monotone response basis · p. 58

Generic-safety calibration.

  • Generic-safety calibration is anchored at zero and bounded, while fixed contextual attenuation thresholds are set at each concept’s 0.90 training quantile with τ_e = 0.25.Contextual attenuation concepts do not enter child-specific mechanism blocks, and their thresholds are not learned by gradient descent.Source: Generic-safety calibration · p. 59, Contextual attenuation · p. 59

B.2.2 Mechanism Structuring and Differentiable Gating

  • Mechanism gates combine fixed concept families conjunctively, including a soft skin-exposure–adultized-context gate for M2; the final model also activates its semantic contribution budget.Pose inversion, broad suggestive recall, self-harm recall, and depressive affect provide intensity evidence rather than independently opening gates; the budget penalty and inference projection are active.Source: §B.2.2 · p. 60, §B.2.3 · pp. 60–61

Training objective.

  • Training uses class-weighted binary cross-entropy with positive-class weight w+=0.45 and a realized-hazard penalty discouraging large or diffuse child-specific hazards, not discrete concept selection.Source: Training objective · pp. 61–62

Residual-separation penalty.

  • An RBF-based separation regularizer contrasts child-specific mechanism hazards with generic-safety vectors, but is a soft dependence regularizer, not an independence guarantee.Source: Residual-separation penalty · p. 62

Probability calibration and decision rule.

Faithful hazard output.

  • When total hazard Λ_i=0, all mechanism shares are defined as zero; any software stabilizer is only to avoid division by zero and does not replace that definition.Source: Faithful hazard output · p. 63

B.3.1 Proof of Proposition 1

The proof converts the assumed composition property into an additive equation, yielding an exponential form that becomes canonical after rescaling.

B.3.1 Proof of Proposition 1

  • The complement Q_t(·) = 1 − F_t(·) is positive, equals 1 at t = 0, and satisfies Q_{a+b}(·) = Q_a(·)Q_b(·).Source: §B.3.1 · p. 63
  • Continuity and strict monotonicity of F make Q continuous and strictly decreasing, so G_t(·) = −ln Q_t(·) satisfies the additive Cauchy equation.Its continuous solutions are G_t(·) = ct for a constant c.Source: §B.3.1 · p. 63
  • Rescaling the additive risk by c yields the canonical form F_t(·) = 1 − exp(−t(·)).Before rescaling, the proof derives F_t(·) = 1 − exp(−ct(·)).Source: §B.3.1 · p. 64

B.3.2 Proof of Proposition 2

The proof establishes that the hazard decomposition is unique under shared block and anchoring constraints, provided joint support permits the required blockwise variations. For ChildRiskGuard, the reference points make each anchored component independent of the other component’s gate or attenuation.

B.3.2 Proof of Proposition 2

  • Under identical total hazards, shared block and anchoring constraints, and blockwise variation allowed by joint support, the decomposed components coincide.The proof varies one block at a time while holding the others at their reference values.Source: §B.3.2 · p. 64
  • In ChildRiskGuard, the all-zero mechanism reference fixes mechanism intensity regardless of its gate, while the generic reference fixes generic hazard regardless of contextual attenuation.Source: §B.3.2 · p. 64

B.3.3 Proof of Proposition 3 … Appendix C Additional Generalizability and Sensitivity Analyses

The appendix establishes conditions for semantic-share bounds and faithful child-specific attribution, characterizes when a conditional Shapley correspondence holds, and details the integrated training and inference workflow.

B.3.3 Proof of Proposition 3

  • Under the stated condition that δ_r=0 for every mechanism, the proof derives a semantic-share bound using H_i.Source: §B.3.3 · p. 65

B.3.4 Formal Faithfulness Axioms and Proposition 4

  • Axiom 1–Axiom 4 formalize operational criteria for mapping-side explanatory faithfulness.Source: §B.3.4 · p. 65

A2 (Sign consistency).

  • Child-risk concepts must have sign-consistent contributions, while contextual attenuation concepts are restricted to the generic-safety channel.Source: A2 (Sign consistency) · p. 65

A3 (Unique attribution).

  • Each concept has a unique structural destination, with disjoint child-specific mechanism blocks and a separate contextual attenuation block, making the decomposition identifiable under Proposition 2.Source: A3 (Unique attribution) · p. 65

Proof of Proposition 4: Axiomatic faithfulness of the child-specific pathway.

B.3.5 Conditional Shapley Correspondence in the Additive Regime

  • When aggregate semantic projection is inactive, each concept’s marginal contribution is identical across coalitions not containing it, yielding a Shapley correspondence conditional on the realized mechanism gate.Source: §B.3.5 · p. 67
  • With active projection, the mechanism remains exactly reconstructable, but the aggregate minimum couples semantic concepts and prevents unconditional exact Shapley equivalence.Source: §B.3.5 · p. 67

B.4 Integrated Training and Inference Procedure

  • Training uses frozen measurements, five-fold out-of-fold F1 threshold selection, and a 480-epoch final fit with the threshold locked before test evaluation; inference reports decisions, hazard shares, and concept contributions.Source: §B.4 · pp. 67–68

C.1. Generalizability to an Additional Child-Specific Risk Mechanism

Adding a new child-specific risk mechanism demonstrates that ChildRiskGuard’s mechanism-structuring approach can be extended without redesigning its core architecture. This tests extensibility, not whether the resulting four mechanisms exhaust relevant risks.

C.1. Generalizability to an Additional Child-Specific Risk Mechanism

  • M4 is independently grounded in external substantive knowledge and operationalized through a separate concept block and mechanism-specific evidence structure.It is introduced alongside the existing M1–M3 mechanisms.Source: §C.1 · p. 69
  • Extending ChildRiskGuard to M4 requires new M4-specific measurements, but not redesigning its core risk-assessment architecture, changing M1–M3, or fine-tuning the frozen foundation model.The risk-assessment stage is re-estimated with M4 added while retaining the existing architecture and M1–M3 specification.Source: §C.1 · p. 69
  • Adding M4 tests whether ChildRiskGuard can accommodate a previously unmodeled risk, not whether four mechanisms exhaust the risks relevant to children.The analysis assesses generalizability of the mechanism-structuring approach rather than completeness of a particular risk taxonomy.Source: §C.1 · p. 70

C.2. Sensitivity to the Semantic Representation … D.2. Operationalization at Consequential Transitions

This section examines how ChildRiskGuard’s performance depends on its semantic readout and how AI4CDS governs provisional AI contributions as they cross into consequential scientific commitments. It operationalizes graduated trust, reversibility, and auditability while reserving scientific authority for researchers.

C.2. Sensitivity to the Semantic Representation

  • F1 fell from 0.769 with the reference representation to 0.731 with OpenCLIP and 0.709 with SigLIP, showing sensitivity to semantic readout choice.The drops were 0.038 and 0.060, respectively; the reference representation is aligned with the reused safety model.Source: Table C2 · p. 70, §C.2 · p. 70

Appendix D Theoretical Grounding and Operationalization of AI4CDS Governance

  • Appendix D presents the three governance principles as nascent prescriptive guidance, grounded in established theory and illustrated by ChildRiskGuard rather than justified by the case alone.The appendix treats the case as instantiation evidence and established theory as the basis for explaining mechanisms and disciplining abstraction.Source: Appendix D · p. 71

Governance at consequential scientific transitions

  • AI4CDS makes consequential transitions—the points where AI-supported outputs affect scientific commitments—the primary unit of governance, retaining human authority over commitments while allowing AI discretion over intermediate steps.Transitions may affect downstream decisions or published claims through propagation, evidentiary or claim significance, or the cost of correcting later-dependent commitments.Source: Governance at consequential scientific… · pp. 71–72

Graduated Trust

  • Graduated Trust calibrates AI discretion to task constraints, independent verifiability, and error consequences: stronger constraints and verification permit more discretion, while open-ended or consequential tasks require more researcher control.The principle is task-contingent rather than determined by a task’s phase in the research process.Source: Graduated Trust · p. 72

Reversibility

  • Reversibility requires reopening an earlier commitment when credible later evidence materially challenges its justification, then explicitly deciding whether to retain, revise, replace, or reject it.It protects against both retaining poorly supported choices and abandoning choices prematurely on weak or unstable contrary evidence.Source: Reversibility · p. 73

Auditability

  • Auditability requires inspectable records of AI’s contribution, the researcher’s decision and its basis, and the effect on the artifact or claim, with documentation scaled to scientific consequence.The auditable object is the observable research process, not the model’s hidden internal reasoning.Source: Auditability · p. 74
  • Auditability connects an externally available contribution with the researcher’s decision and supporting evidence.Source: Auditability · p. 74

D.2. Operationalization at Consequential Transitions

  • Graduated trust governs AI discretion before a transition, reversibility addresses later evidence that undermines a commitment, and auditability preserves a record for evaluation.Table D1 translates the three principles into observable assessments and required responses.Source: §D.2 · pp. 74–75, Table D1 · p. 75
  • AI participation does not transfer scientific authority; governance asks how a provisional contribution crosses into a consequential commitment.The passage says the relevant governance question remains how that transition occurs.Source: §D.2 · p. 75

D.3. Differentiated Reproducibility as an Auditability Standard · D.4. Inter-Principle Relationships, Tensions, and Boundary Conditions

AI4CDS matches verification to the kind of research activity and connects trust, reversibility, and auditability through proportional governance. At consequential scientific transitions, validity and accountability take precedence, while the principles’ application depends on task and research conditions.

D.3. Differentiated Reproducibility as an Auditability Standard

  • Differentiated reproducibility matches verification to each activity’s determinism, stochasticity, process dependence, and claim type, avoiding infeasible exact reproduction or imprecise evidence.AI-enabled CDS spans deterministic computation, stochastic model fitting, generative search and diagnosis, and formal analytical claims.Source: §D.3 · pp. 75–76
  • Verification targets execution for deterministic computation, result stability for stochastic computation, decision-process reconstruction for AI-assisted search, and derivation validity for formal claims.Procedural reconstruction requires provenance sufficient to assess how AI-supported alternatives became researcher-authorized commitments, not identical generative outputs.Source: §D.3 · pp. 76–77

D.4. Inter-Principle Relationships, Tensions, and Boundary Conditions

  • Prior AI performance may justify greater discretion on comparable, well-specified tasks, but credible contradictory evidence makes reversibility take precedence over accumulated confidence.Earlier commitments must remain reopenable when later evidence challenges their justification.Source: §D.4 · p. 77
  • Reconsidering a consequential commitment requires an audit record of its rationale, supporting evidence, dependencies, new evidence, and whether it was retained, revised, replaced, or rejected.Full documentation applies to consequential reversals, not every minor iteration.Source: §D.4 · p. 77
  • AI4CDS balances audit completeness with feasibility through proportionality, increasing governance effort with a decision’s consequence, verification difficulty, and propagation potential.Routine assistance can receive lightweight documentation, while transitions affecting admitted evidence, artifact structure, reported results, or transferable claims require stronger records.Source: §D.4 · p. 77
  • At consequential scientific transitions, validity and accountability take precedence over efficiency, with AI discretion contingent on constraint and independent verification and application shaped by task, system, and research setting.ChildRiskGuard illustrates an inspectable instantiation, but studies across CDS problems, AI capabilities, teams, and epistemic settings are needed to test and delimit the principles’ effectiveness.Source: §D.4 · pp. 77–78

Related papers

Loading 2609.05270v1…