Evidence Requirements for AI-Assisted Structural Diagnostics

1. Introduction

Recent studies in AI-assisted structural inspection, taken together, show that some systems now extend beyond stand-alone image classification or detection. The cited work includes pipelines for metric scene reconstruction from video, structured evidence binding for safety assessment, crack delineation linked to measurement, and probabilistic updating of deterioration state from inspection and monitoring data. This article does not treat that body of work as establishing a settled evidentiary standard. Instead, it addresses a narrower methodological question: what evidence should be required before AI-generated outputs are used in structural diagnostic reasoning about the presence, extent, severity, or progression of a structural condition? [1] [2] [3]

Direct application studies already indicate the relevant components of that burden. A vision-based indoor building inspection framework reconstructs a metric-scale scene from smartphone video, organizes geometry and semantics in a structured state representation, and constrains AI reasoning through evidence-linked tool calls rather than unrestricted text generation. Bridge crack studies likewise show a shift from generic detection toward workflows that must preserve slender, irregular crack morphology and enable subsequent measurement. A hierarchical Bayesian bridge deterioration model further demonstrates that periodic inspection and continuous monitoring reflect complementary aspects of the same structural state and should not be treated as wholly separate evidence streams. [1] [4] [2] [3]

The methodological foundations of a stricter evidence framework also exist outside structural diagnostics. Inductive conformal prediction offers finite-sample marginal coverage guarantees for class-label outputs in object detection, creating an explicit basis for flagging uncertain cases. Separately, a service-reliability framework in materials evaluation argues that evidence should be defined by retained performance after relevant exposure, with comparisons made through structural descriptors and matched test conditions rather than nominal category labels alone. Transferred carefully, these ideas suggest that structural AI should be judged not only by initial benchmark accuracy but by retained diagnostic usefulness under matched structural, environmental, and acquisition conditions. [5] [6]

2. Research question and evidence scope

This article asks: What evidence should be required before AI-assisted vision systems are used for structural diagnostic judgments concerning defect presence, measurement, and deterioration state? The scope is intentionally narrow. It addresses evidence requirements for visual or vision-linked diagnostics in buildings and bridges, especially for crack-related assessment and safety-relevant scene interpretation. It does not examine broad adoption policy, procurement, or regulatory harmonization except where a narrow methodological implication is necessary.

The synthesis is targeted rather than systematic. It draws on six sources for bounded purposes. First, the building inspection study is used as direct evidence that an inspection pipeline can be made evidence-constrained through structured scene representation and tool-mediated verification. Second, the two bridge crack papers are used as direct evidence that detection and segmentation pipelines are increasingly tied to measurement-oriented tasks and must cope with domain variation, complex backgrounds, and fine crack geometry. Third, the Bayesian fusion paper is used as direct evidence that inspection and monitoring data can be integrated probabilistically to update deterioration assessment over time. Fourth, the conformal prediction paper is used as methodological evidence for finite-sample coverage control of vision outputs. Fifth, the retained-performance framework is used as methodological evidence for defining evidence by service exposure, structural descriptors, and matched test conditions. [1] [4] [2] [3] [5] [6]

The manuscript does not claim that these studies together already validate a unified structural diagnostic framework. Instead, it proposes a conceptual methodology that integrates what these studies support directly and leaves empirical confirmation to prospective testing.

3. Prior work and the unresolved methodological gap

The building-inspection framework is important because it formalizes the relation between observation and judgment. It converts a continuous smartphone video into a metric 3D point cloud, parses semantic entities and spatial relations, stores evidence in a hierarchical scene representation, and performs a dual-pathway assessment consisting of an initial scan and follow-up evidence queries. Its pilot validation against a single-assessor manual reference showed variable recall and precision and an explicit judgment closure rate, which is valuable because the study treats incomplete evidence as a reportable outcome rather than hiding it inside a single accuracy score. For present purposes, its strongest contribution is not the reported performance level but the idea that AI inspection should be evidence-constrained and audit-trailed. [1]

The bridge crack detection study focuses on improved object detection under weak contrast, irregular morphology, and complex textures. It reports that preprocessing and architectural modifications contribute complementarily, and it includes repeated-seed and external-dataset evaluation as evidence for stability and generalisation potential. Yet detection outputs are bounding-box oriented and are therefore one step removed from measurement-grade diagnosis. They are appropriate for defect hypothesis generation and spatial prioritization, but not sufficient by themselves for width or length estimation. [4]

The lightweight segmentation study is closer to diagnostic utility because it explicitly connects segmentation to quantitative analysis through an orthogonal skeleton method and reports controlled measurement errors for crack width and length. Its account also distinguishes detection from segmentation and frames UAV inspection as a deployment context with computational constraints. This supports an important boundary condition: diagnostic evidence depends on the downstream task. A model that is adequate for locating likely crack regions may still be inadequate for geometric quantification. [2]

The Bayesian fusion study addresses a different gap: temporality. It argues that inspection and monitoring are complementary observations of the same structural system and couples discrete inspection indicators with continuous crack monitoring in a hierarchical posterior-updating scheme. That formulation is directly relevant because a diagnostic statement about structural condition is often time-indexed. A visually detected crack may indicate a different level of concern depending on whether monitoring suggests stability, intermittent opening, or progressive growth. The study therefore supports using multi-source posterior updating as part of evidence assessment rather than treating each inspection as an isolated image-recognition event. [3]

The unresolved methodological gap is that these strands are rarely joined under an explicit evidence requirement. Existing application papers provide performance metrics for detection or segmentation, and some provide measurement or fusion models, but they do not yet specify a unified rule for when an AI output is sufficiently supported to be used as a diagnostic judgment. Nor do standard benchmark metrics answer whether performance is retained under relevant service exposure, scene complexity, acquisition geometry, or sensor shift. That gap motivates the framework proposed here. [4] [2] [6]

4. Conceptual framework: evidence retention for structural diagnostics

The proposed framework is called evidence retention for AI-assisted structural diagnostics. Its premise is that diagnostic validity depends on retained evidential quality across the full chain from acquisition to action recommendation. Five evidence layers are distinguished.

Layer 1 is task validity: whether the AI output corresponds to the intended diagnostic subtask, such as defect presence screening, defect delineation, metric measurement, or deterioration-state inference. Layer 2 is measurement validity: whether the output can be traced to an auditable geometric or pixel-level basis adequate for the claimed quantity. Layer 3 is uncertainty validity: whether the system provides calibrated uncertainty signals with interpretable escalation rules. Layer 4 is temporal validity: whether present outputs are integrated with prior inspections or continuous monitoring when the target construct concerns progression rather than a single observation. Layer 5 is deployment validity: whether the claimed performance is retained under matched structural descriptors and service-exposure conditions, including surface type, illumination, viewing angle, motion blur, moisture state, and sensing platform.

The key transfer from retained-performance analysis is that evidence should be indexed by descriptor matching, not by nominal task labels alone. Two crack-segmentation systems evaluated on different surface textures, acquisition heights, or moisture conditions should not be treated as equivalent evidence for deployment. Likewise, an indoor scene reasoning system validated in uncluttered, well-lit spaces does not automatically provide evidence for occluded or low-light structural conditions. The relevant question is whether diagnostic usefulness is retained when tested under matched descriptors of the deployment setting. [6] [1]

This article therefore proposes an Evidence Retention Matrix, ERM. Each case c is associated with descriptor vector D(c) = [s, e, a, t, m], where s denotes structural descriptors, e exposure descriptors, a acquisition descriptors, t target-task descriptors, and m metadata quality descriptors. Structural descriptors include component type, material surface class, crack orientation complexity, and expected defect scale. Exposure descriptors include illumination regime, moisture, contamination, thermal state, and traffic-induced vibration if relevant. Acquisition descriptors include sensor type, stand-off distance, view angle, image resolution, motion characteristics, and reconstruction scale control. Target-task descriptors distinguish detection, segmentation, width estimation, length estimation, scene-risk reasoning, and deterioration updating. Metadata quality descriptors capture calibration traceability, annotation confidence, and synchronization across modalities.

Retained evidence for deployment case c* is judged only from prior evidence generated on cases c for which a bounded matching function is computable from comparable descriptors. Let J(c,c*) denote the set of descriptor components observed and comparable in both cases after task-specific admissibility checks. If any descriptor designated as critical for the target task is missing, non-comparable, or internally inconsistent, matching is not computed and the case is routed to evidence insufficiency rather than forced comparison. For admissible cases, define M(c,c*) = (sum_{j in J(c,c*)} w_j m_j(c,c*))/(sum_{j in J(c,c*)} w_j), where each local similarity m_j lies in [0,1], the weights w_j are non-negative, and at least one comparable component must be present. Here m_j = 1 denotes a descriptor match and m_j = 0 the maximum relevant mismatch under the preregistered normalization for that descriptor. Direction-sensitive descriptors are scored asymmetrically when support in one direction does not justify transfer in the other; for example, deployment at poorer effective resolution than calibration may incur a larger penalty than the reverse. Correlated descriptor families are handled by grouped weighting: if several variables represent the same practical source of shift, their combined contribution is capped by a preregistered group weight to avoid double counting. A study may support deployment only if M(c,c*) >= tau_match and all critical descriptors satisfy their componentwise minimums. The threshold tau_match remains hypothetical until prospective calibration, but the rule is interpretable across cases because M is explicitly bounded in [0,1]. Unsupported descriptor shift, missing critical descriptors, or unresolved descriptor conflict is treated as an evidence deficit, not a minor limitation. [6]

5. Proposed methodology and decision sequence

The methodology is designed for AI systems that assist, rather than replace, structural diagnosis. It separates hypothesis generation from decision eligibility. Inputs are: visual data V; optional geometric reconstruction G; optional monitoring data H as time series; inspection history I; and deployment descriptors D. Outputs are: defect hypothesis set Z; measurement set Q; uncertainty set U; posterior deterioration state R; and an action qualification label A.

Step 1 is acquisition qualification. The system records descriptor vector D and checks whether required metadata are present, internally consistent, and comparable with the intended task stratum. If scale-sensitive tasks are requested, such as crack width or clearance measurement, the case is marked measurement-ineligible unless there is a verified scale source, for example calibrated fiducials, known geometry, or validated photogrammetric reconstruction. When multiple scale sources are available, the default rule is conservative: if the sources disagree beyond a preregistered tolerance, the system suppresses automated measurement output and routes the case to manual review unless an external calibration record has predefined priority. Missing metadata for any descriptor declared critical to the task also block automated eligibility rather than being imputed silently. This rule follows directly from the building-inspection study, which used AprilTag fiducials to obtain absolute scale, and from the segmentation study, which tied image outputs to measurement-oriented analysis. [1] [2]

Step 2 is defect hypothesis generation. A detector proposes candidate regions Z = {z1, ..., zn}. Detection models of the type used in bridge crack studies are suitable here because they are fast and robust to complex backgrounds, but their role is limited to proposing candidate zones and classes. They do not finalize diagnosis. [4] [2]

Step 3 is uncertainty-controlled class interpretation. For each candidate zi, the detector produces a class-score vector pi. A conformal module calibrated on a descriptor-matched calibration stratum transforms pi into a prediction set Ci(alpha) with target marginal coverage 1 - alpha for future cases drawn from that stratum. The procedure-level validity condition is therefore external to any single case: deployment may rely on conformal routing only if the calibration design, task definition, and descriptor stratum have been specified in advance and subsequently checked prospectively for acceptable empirical coverage. The case-level routing rule is then: if |Ci(alpha)| = 0 or |Ci(alpha)| > k_max, zi is escalated for human review; if |Ci(alpha)| = 1, automated routing to the next measurement step is permitted; if 1 < |Ci(alpha)| <= k_max, all labels in Ci(alpha) are preserved in downstream reasoning and single-class action statements are forbidden. Here k_max is a hypothetical operational constant chosen in advance by the deployment authority. This step imports finite-sample marginal coverage control from inductive conformal prediction and turns uncertainty into a workflow constraint rather than a confidence number alone. [5]

Step 4 is delineation and metric extraction. For zones that survive step 3, a segmentation model estimates a crack mask Si. A measurement operator then extracts centerline, projected length, and width profile. Width is defined operationally only for crack segments that admit a stable one-dimensional centerline after topology screening. The default rule is to decompose the mask into maximal non-branching segments; for each valid segment, width is the median orthogonal distance between opposing mask boundaries along the skeletonized centerline over skeleton points that satisfy preregistered quality checks for local continuity, boundary visibility, and angle stability. Branch points, fragmented segments below a minimum connected length, and regions with occluded or truncated boundaries are excluded from automated width estimation and flagged for manual review or length-only reporting. This reflects the measurement orientation of the segmentation study, while the exact estimator and exclusion rule are proposed here for prospective testing. Measurement validity requires an auditable chain from mask pixels to reported quantity, including scale source and image quality flags. [2]

Step 5 is evidence binding. Each output is stored in an evidence object Ei = <zi, Ci, Si, Qi, Ui, D, provenance>, where provenance includes source images, tool versions, calibration record, and any reconstruction or cross-check results. This step adopts the evidential logic of structured scene state from the building-inspection framework but applies it to structural diagnostics broadly. [1]

Step 6 is temporal fusion when the target claim concerns progression, severity trend, or intervention urgency. Inspection-derived evidence objects Ei are linked with monitoring stream H and prior inspection history I. A hierarchical Bayesian layer estimates posterior deterioration state R_t from both discrete inspection indicators and continuous or quasi-continuous monitoring variables. The specific probabilistic family may vary by asset and data type, but the methodological requirement is fixed: static image evidence alone cannot justify a progression claim when time-series evidence exists and is materially relevant. [3]

Step 7 is action qualification. The system does not issue a single undifferentiated diagnosis. Instead it assigns one of four labels: A0 insufficient evidence; A1 defect suspected, manual confirmation required; A2 defect confirmed with auditable measurement but no progression claim; A3 defect confirmed with auditable measurement and temporally updated risk estimate. The labels formalize a distinction often left implicit in application studies.

6. Operational definitions, assumptions, comparator, and outcome measures

The primary research construct is diagnostic decision eligibility. A case is decision-eligible for task T only if three conditions hold simultaneously: descriptor matching, validity of the uncertainty-control procedure for the relevant task stratum, and evidence traceability. The operational rule is: Eligible(T) = 1 if and only if M(c,c*) >= tau_match, ConfValid(S,T) = 1, and Trace(T) = auditable; otherwise Eligible(T) = 0. Before this rule is applied, an admissibility screen checks whether all task-critical descriptors are present and comparable and whether provenance records are coherent. If that screen fails, Eligible(T) is set to 0 without attempting numerical matching. ConfValid(S,T) refers to the conformal procedure for task T within calibration stratum S, not to a property of an individual prediction set. It is satisfied only when the nonconformity function, calibration sample definition, and target alpha were prespecified and prospective evaluation shows acceptable empirical coverage for that stratum. Trace(T) is auditable when the reported conclusion can be linked to stored images, masks or geometric measurements, calibration data, and versioned processing steps. For partial descriptor overlap, only non-critical components may be omitted from M, using the renormalized weighted average defined in the ERM; critical components may not be waived by averaging. [5] [6]

Assumptions should be declared explicitly. First, conformal coverage is marginal and finite-sample under calibration conditions; it does not guarantee instance-conditional correctness, and performance may degrade under distribution shift. Second, geometric measurements depend on reliable scale assignment and sufficient image quality. Third, Bayesian deterioration updating assumes that the monitoring variables and inspection indicators are meaningfully related to the deterioration state and that the chosen likelihood structure is not grossly misspecified. Fourth, the framework assumes that human review remains available for escalated cases. Fifth, descriptor comparability is itself an assumption: descriptors must be defined with task-specific normalization rules, and asymmetry should be declared where mismatch is direction-sensitive, such as training support at high resolution being transferred to lower-resolution deployment. If descriptor records, scale sources, or provenance elements conflict and the conflict cannot be resolved by a preregistered precedence rule, the default action is escalation rather than automated decision eligibility. [5] [3]

The natural comparator is current practice in two forms. Comparator C1 is manual visual inspection with conventional measurement tools and narrative judgment. Comparator C2 is AI-only image recognition without explicit uncertainty control, descriptor matching, or temporal fusion. The proposed framework should not be compared only against detector mAP or segmenter mIoU because those metrics capture model discrimination, not diagnostic decision eligibility. The appropriate conceptual comparison is whether the proposed process yields better-defined evidence states and fewer unjustified diagnostic claims than C2, while improving auditability relative to C1. [1] [2]

Outcome measures for prospective evaluation should therefore include at least six quantities. O1 is coverage validity: empirical containment of true class labels within conformal prediction sets at prespecified alpha levels. O2 is set efficiency: mean prediction-set size, because broad sets may satisfy coverage while remaining operationally weak. O3 is measurement agreement: discrepancy between AI-derived and reference crack width or length under matched scale conditions. O4 is escalation rate: proportion of cases routed to human review because evidence is insufficient. O5 is retention score: performance stratified by descriptor mismatch to test whether evidence decays under shift. O6 is posterior risk calibration: agreement between updated deterioration probabilities and subsequent observed structural evolution where longitudinal data exist. These measures align the evaluation with the proposed evidence layers rather than with detection accuracy alone. [5] [2] [3] [6]

7. Prospective validation protocol

Because this article is conceptual, a detailed prospective protocol is offered instead of unperformed results. Validation should be organized by task strata rather than by one aggregate dataset. At minimum, four strata are required: S1 defect detection in still images, S2 crack delineation for measurement, S3 evidence-constrained scene assessment in spatially complex inspections, and S4 temporal deterioration updating using inspection plus monitoring. Cases in each stratum should be indexed by descriptor vector D, and splits should preserve a deployment-relevant test partition with documented descriptor differences from calibration data.

For S1, the primary test is whether conformalized class sets achieve nominal coverage within descriptor-matched groups and how rapidly coverage degrades as descriptor mismatch increases. For S2, the primary tests are measurement agreement and failure detection under small crack widths, blur, oblique view, low contrast, or textured backgrounds. For S3, the primary tests are judgment closure rate, evidence traceability, and agreement with expert assessment on whether the available evidence supports or fails to support a safety-relevant scene claim. For S4, the primary tests are posterior-update stability and whether fused models revise risk estimates coherently when new monitoring information conflicts with earlier inspection impressions. [1] [5] [3]

A crucial design feature is intentional stress testing under service exposure and acquisition variation. Borrowing the retained-performance logic, the protocol should not end with pristine benchmark images. It should include controlled variations in lighting, moisture, stand-off distance, vibration, occlusion, and surface contamination, because diagnostic utility may be lost under such conditions even when baseline accuracy is high. Performance should be reported jointly with descriptor distributions so that support for future deployments can be judged by match quality rather than by headline averages. [6] [4] [2]

Illustrative thresholds may be used operationally, but they must be declared hypothetical until validated. A prospective protocol should preregister one default operating point for each task, including tau_match, the set of critical descriptors, permissible metadata omissions for non-critical fields, scale-consistency tolerance, conformal significance level alpha, and k_max for multi-label escalation. Sensitivity analysis may then vary these parameters over a limited published range while keeping the default rule fixed as the primary analysis. Such thresholds are governance and testing parameters, not empirical findings of this article.

8. Failure cases, conflicting evidence, and limits of generalization

The framework is designed to expose failure rather than conceal it. Several failure modes are foreseeable. First, detector and segmenter outputs may fail on domain shift, especially where surface textures, lighting, or imaging geometry differ materially from training and calibration conditions. The crack detection study itself reports retained capability on an external dataset but also notes performance decrease under external conditions, which is evidence for generalisation potential and for remaining sensitivity to domain variation at the same time. This is not a contradiction; it is a boundary condition. [4]

Second, prediction-set validity may hold marginally while remaining unhelpful for individual decisions if sets become too large in complex scenes. The conformal prediction study reports a balance between coverage and efficiency and notes that some scores produce larger sets under high data complexity and strict coverage targets. Structural diagnostics should therefore treat set size as an operational outcome, not assume that formal coverage alone solves trustworthiness. [5]

Third, accurate segmentation does not automatically imply accurate metrology. Measurement can fail because of poor scale control, perspective distortion, skeleton instability in branching cracks, or partial occlusion. The segmentation paper supports the feasibility of measurement-oriented workflows, but the proposed framework restricts measurement claims to cases with auditable scale and valid image geometry. [2] [1]

Fourth, temporal fusion may amplify errors if inspection coding, monitoring synchronization, or latent deterioration assumptions are inconsistent. The Bayesian fusion study shows the promise of posterior updating, but the present proposal does not assume that any fusion improves diagnosis. Fusion is justified only when the modalities are linked to the same deterioration construct and their provenance is reliable. [3]

Generalization is intentionally limited. The proposal is directed to visible or vision-linked structural diagnostics for buildings and bridges, with special relevance to crack-related assessment and scene-based safety interpretation. It does not establish evidence requirements for hidden defects requiring nondestructive sensing beyond the visible spectrum, nor for autonomous intervention planning. It also does not claim universal thresholds. Descriptor weights, matching thresholds, conformal nonconformity functions, and posterior model families remain context dependent and require empirical tuning.

9. Conclusion

This article established a bounded conceptual claim: AI-assisted structural diagnostics require evidence standards that are stricter than benchmark detection accuracy and should be defined in terms of retained diagnostic usefulness under matched descriptors of structure, exposure, acquisition, and task. By synthesizing evidence-constrained building inspection, bridge crack detection and measurement workflows, hierarchical fusion of inspection with monitoring, finite-sample coverage control for vision outputs, and retained-performance reasoning, it proposed a practical framework centered on descriptor matching, uncertainty-controlled escalation, auditable measurement, and temporal posterior updating. [1] [4] [2] [3] [5] [6]

What has been established here is conceptual and methodological: the relevant evidence layers, an explicit operational rule for decision eligibility, a decision sequence from acquisition to action qualification, and a prospective validation protocol aligned with diagnostic use rather than image-recognition scores alone. What remains untested is the empirical adequacy of the full integrated framework, including optimal descriptor weights, acceptable matching thresholds, real-world coverage under distribution shift, and the operational trade-off between escalation burden and retained safety. Those questions require prospective multi-stratum evaluation before the proposed evidence requirements can be treated as validated practice.

References

1. Jing, Zijian; Zhu, Liyi; Chen, Tianyi; Hovestadt, Ludger; Li, Li (2026). Vision-Based Digital Twin and AI Agent Framework for Low-Cost, Explainable Indoor Building Inspection and Safety Assessment. Sensors. 26(15), 4992. doi:10.3390/s26154992.
2. Dong, Xuwei; Yu, Zhibing; Yuan, Jiashuo; Dai, Jinpeng (2026). Automatic bridge crack detections: A lightweight segmentation model for UAV inspection. PLOS One. 21(9), e0358865. doi:10.1371/journal.pone.0358865.
3. Wang, Benyu; Chen, Ke; Wang, Bingjian (2026). Hierarchical bayesian fusion of inspection and monitoring data for probabilistic bridge deterioration assessment. Scientific Reports. 16(1), 5965. doi:10.1038/s41598-026-36808-4.
4. Zhu, Xianwei; Chao, He; Zhang, Yahui (2026). Yolov8-CA-EYHL: A bridge crack detection algorithm based on improved YOLOv8. PLOS One. 21(9), e0358872. doi:10.1371/journal.pone.0358872.
5. Mohammed, Mohammed Aliy; Anzaku, Esla Timothy; Jonkers, Jef; Krishnamoorthy, Janarthanan; Neve, Wesley De; Van Hoecke, Sofie (2026). Inductive Conformal Prediction for Guaranteed Class-Label Coverage in Object Detection. Journal of Imaging. 12(8), 348. doi:10.3390/jimaging12080348.
6. Ni, Xuepeng; Guan, Jingyu; Dou, Jie; Guo, Jiaming; Cui, Zhe; Wang, Xiaojuan; Ren, Caixia; Hou, Liyin; et al. (2026). Polymer fibrous aerogels for thermal insulation: from molecular building blocks to structural reliability. RSC Advances. 16(46), 58917-58949. doi:10.1039/d6ra08043j.