Designing an Optimal Prompt for Generative AI to Perform Hypothetical Reasoning in Technical Judgments of Road Bridges

1 Designing an Optimal Prompt for Generative AI to Perform Hypothetical Reasoning in Technical Judgments of Road Bridges (Version 1) Takumi Kobayashi and Michio Ohsumi Public Works Research Institute, Ibaraki, Japan October 29, 2025 ABSTRACT This study proposes a method to induce hypothetico-deductive reasoning in Generative AI via Prompt Engineering, with the aim of assisting technical decision-making in road bridge maintenance. Specifically, it optimizes prompts input to a multimodal Large Language Model through Simulated Annealing and Hoeffding-UCB, building upon humanauthored prompts that incorporate domain-specific terminology. Employing prompts optimized for hypotheticodeductive reasoning with a given model enabled the reproduction of the thought processes of experienced engineers. This yielded responses semantically closer to those of engineers and logical reasoning scores that were 1.2 - 1.4 times greater than those obtained by experienced engineers. However, no improvement was observed in the extent of hallucination suppression. The prompts developed in this study to induce hypothetico-deductive reasoning are characterized by versatility: they can be readily employed by engineers in routine practice and incorporated into algorithms utilizing various LLMs. 1. INTRODUCTION The rapid advancement of Large Language Models (LLMs) has prompted extensive investigations into AI applications for specialized tasks that were previously considered difficult. For instance, in the medical field, proposals (Asan et al., 2020; Asan et al., 2022; Jermutus., 2022) suggest a role differentiation in which AI assists physicians with diagnosis, while physicians’ expertise and ethical judgment remain indispensable for final decisions. This suggests the establishment of a new collaborative framework between humans and AI, a framework that could potentially extend to other specialized domains. This paper focuses on technical judgments in the periodic inspection of road bridges and emergency inspections following earthquakes, aiming to explore collaboration between humans and Generative AI. While Generative AI does not necessarily need to follow the identical reasoning processes as humans, making its simulated reasoning processes as close as possible to human cognition offers the advantage of rendering its outputs more intelligible to humans. A Generative AI that produces reasoning both comprehensible to humans and capable of advanced technical judgments 2 enables humans to readily assess the validity of the AI’s outputs and ensuing decisions. This, in turn, is expected to facilitate smoother decision-making. These advantages are particularly significant in post-earthquake emergency inspections, where judgments and decisions regarding critical or fatal structural damage must be made under severe time constraints. Furthermore, prior studies (Takumi & Michio., 2024; PWRI., 2025) have clarified that the reasoning patterns of experienced engineers—which this paper seeks to emulate in Generative AI—can be reasonably accounted for by hypothetico-deductive reasoning. Against this background, the present study aims to design prompts that induce hypothetico-deductive reasoning in Generative AI to support technical judgments for road bridges, which require advanced domain knowledge and reasoning ability. It should be noted that this study proceeds solely through Prompt Engineering, without additional training. The rationale is that recent LLMs, including OpenAI’s ChatGPT, already possess sufficient knowledge through extensive pre-training, and such knowledge can be effectively elicited through appropriately crafted prompts. 2. LITERATURE REVIEW This chapter provides an overview of prior research from three perspectives: (1) studies on Prompt Engineering, (2) studies on domain-specific AI applications, and (3) studies on the thought patterns of experienced engineers, thereby situating the present study within the existing body of research. 2.1 Research on Prompt Engineering In recent years, representative techniques for Prompt Engineering, such as Chain-of-Thought (CoT) (Jason et al., 2022) and its advanced variant Logic-of-Thought (LoT) (Tongxuan et al., 2024), have been introduced. These approaches incorporate explicit intermediate steps into the reasoning process, thereby enabling more logical and sophisticated forms of inference. For instance, in CoT, simply appending the phrase “Let’s think step by step” to a question prompts the model to perform reasoning sequentially. Research inspired by hypothetico-deductive reasoning, which is particularly relevant to this study, includes the work of Yitian Li et al. (2024). In tackling reasoning problems, their approach explicitly induces the steps: assuming a conclusion → backward reasoning → fact verification, demonstrating improved performance with ProofWriter (Oyvind., 2020) and RuleTaker (Peter., 2020). Furthermore, Auto Prompt frameworks such as Yongchao Zhou et al. (2022) Automatic Prompt Engineer (APE) and Chengrun Yang et al. (2023) Optimization by PROmpting (OPRO) have achieved performance equal to or surpassing manually designed prompts through the exploratory generation of prompts. However, these studies primarily assume text-based LLMs. In contrast, multimodal LLMs, which are the focus of this research, tend to require lengthy and complex instructions, making the quantitative evaluation of effective prompts particularly challenging. To address this, the present study proposes a method to optimize prompts input to multimodal LLMs, building upon human-authored prompts that incorporate specialized terminology. 3 2.2 Research on Domain-Specific AI Applications Numerous studies have employed multimodal LLMs, capable of handling both images and text, to generate inspection findings from image data. BridgeCLIP (Powei & Gaku., 2025), for example, is a model based on the CLIP architecture that analyzes high-resolution UAV imagery to generate inspection findings, and, through few-shot prompt tuning, markedly improves terminological accuracy compared to conventional models. AECIF-Net (Chenyu et al., 2023) is a vision–language model specialized for infrastructure, offering versatility applicable not only to bridges but also to tunnels and pavements. It combines a ResNet-based encoder with a GPT-based decoder to jointly learn damage detection and findings generation. BDCD-Net (Bridge Damage Captioning Network) (Li et al., 2024) incorporates pretrained attention mechanisms for each bridge structural component (deck, pier, girder) to generate detailed captions containing “component name + damage type + location information.” In comparison with these prior studies, this research is characterized by three key features: (1) it ensures a transparent structure of logical reasoning grounded in hypothetico-deductive methods, thereby aligning with the reasoning processes of experienced engineers; (2) it allows the direct reuse of existing models such as OpenAI’s ChatGPT, thereby facilitating its adoption within existing human-centered operational practices; and (3) it enhances the quality of inspection findings exclusively through Prompt Engineering, without additional training, thereby allowing seamless integration of prompts into other algorithms. In summary, by refraining from adopting a development policy confined either to human-centered or AI-centered maintenance, this study seeks to enable incorporation into diverse future operational frameworks. Accordingly, it is anticipated that the prompts developed herein will contribute to Augmented Intelligence (ISO 9241-210., 2010), integrating AI into expert workflows to extend human capabilities, while ensuring accountability and maintaining transparency from the perspective of Human-Centered Design (Mark., 2020). 2.3 Research on the Thinking Patterns of Experienced Engineers There is limited research that has elucidated the thinking patterns of experienced engineers with respect to technical judgments concerning maintenance policies for road bridges. Recent studies by Kobayashi et al. (Takumi & Michio., 2024; PWRI., 2025) have demonstrated that such thinking patterns can be modeled using the hypothetico-deductive method. According to their findings, the reasoning of experienced engineers in bridge maintenance involves the following logical structure: beginning with relatively easily detectable damage as evidence, deriving hypotheses by means of inductive reasoning grounded in knowledge regarded as common sense among bridge engineers, incorporating damage directly associated with the outcomes of technical judgments, engaging in deductive reasoning based on the arguments, and finally articulating the results of technical judgment in a matter-of-fact manner, which leads to the determination of the maintenance policy (Takumi & Michio., 2024). Furthermore, through semi-structured interviews, it was verified that the thinking patterns of experienced engineers can be largely accounted for by the hypothetico-deductive method (PWRI., 2025). That study also demonstrated that diagnostic findings produced by many engineers can be decomposed into four categories of statements: (1) factual evidence, (2) knowledge-based reasoning, 4 (3) conclusion, and (4) recommended action. However, from a syntactic perspective, these statements do not necessarily adhere to formal logical rules. In essence, while the content of the components (evidence, reasoning, conclusion, etc.) is common across engineers, the sequence and dependency relations differ. This indicates the predominance of a “loosely structured” hypothetico-deductive mode of reasoning, in the sense that logical rules are not explicitly observed. It is important to note that even when the thinking patterns of experienced engineers are grounded in the hypotheticodeductive method, abbreviations or omissions arise due to the presence of tacit knowledge that cannot be fully verbalized. Consequently, diagnostic findings are not always formulated strictly in accordance with complete logical rules. Responses generated by Generative AI exhibit similar characteristics. On this point, prior studies have also emphasized that human explanations of causal chains are inherently selective and are deeply embedded in social contexts (Mark., 2020; Miller., 2019). In other words, humans typically refrain from explicating complete causal relationships; instead, they implicitly presuppose that the recipient shares a common cultural background and associated knowledge, and accepts the credibility of the explanation. The principal objective of this study in developing prompts for Generative AI is to approximate the reasoning patterns of engineers, thereby enabling the generation of explanatory texts that engineers can both comprehend and trust. Moreover, by encouraging Generative AI to employ hypothetico-deductive reasoning, the study also seeks to achieve enhanced quality in diagnostic findings. 3. METHODOLOGY 3.1 Development of an Optimal Prompt Search Algorithm This study is predicated on using exemplary answers provided by experienced engineers (PWRI., 2025) as groundtruth data. Building on this foundation, the maximization of the scores described in Sections 3.1.2 and 3.1.4 is treated as the reward in a mathematical optimization problem, while simultaneously minimizing the number of search iterations. Framing the task as a mathematical optimization problem is warranted, as exhaustive enumeration of all possible candidate question formulations—beyond those explicitly presented herein—would be impractical. Figure 1 illustrates the algorithm and flow for optimal prompt search. In summary, the procedure involves repeatedly posing questions to the OpenAI LLM (ChatGPT) targeted in this study, generating response texts, evaluating these responses across multiple metrics, performing convergence assessment toward a global optimum and escape from local optima using Simulated Annealing (SA) (Kirkpatrick et al., 1983) , exploring candidate solutions in the neighborhood of high-scoring prompts via Hoeffding-UCB (Auer et al., 2002), and incorporating up to three restarts to escape local optima. In Hoeffding-UCB, the maximum number of new beam searches was set to 50. 5 FIGURE 1 Algorithm and flow for optimal prompt search. Beyond this, a range of alternative algorithms was explored through trial and error in order to determine an algorithm and flow capable of balancing score maximization with minimization of search iterations. For example, during the exploratory phase, more rigorous approaches such as Hoeffding-UCB (Auer et al., 2002) and KL-UCB (Garivier & Cappé., 2011) —which offer theoretical guarantees of convergence—were considered to attain solutions with stronger optimality guarantees. However, these algorithms resulted in a substantial increase in the number of explorations. The emphasis on UCB (Upper Confidence Bound) arose from framing prompt creation as a multi-armed bandit problem, in which scores must be maximized within limited trials under conditions of unknown probability. Ultimately, as a pragmatic compromise with real-world deployment in mind, SA was employed for convergence assessment, while Hoeffding-UCB was used for exploring candidate solutions proximate to high scores. Moreover, given the uncertainty in the variability of scores reflecting the quality of findings generated by the LLM, and the lack of clarity regarding trends in score variation attributable to differences in question formulation, both SA and restart were incorporated as mechanisms for escaping local optima. Python was used for the analyses. The details of the algorithm and flow presented in Figure 1 are described in Sections 3.1.1 through 3.1.6. 3.1.1 Photos and Prompts Entered into ChatGPT As shown in Figure 1, images and question prompts were entered into ChatGPT to generate inspection findings. During the analysis, ChatGPT’s language was set to English, and long-term memory was disabled. Photo 1 illustrates the photos entered into ChatGPT. These are the same photos examined by PWRI (2025). In their study, PWRI (2025) presented these photos to experienced engineers, conducted semi-structured interviews, and Input questions and photos into ChatGPT and retrieve responses (3 answers per question) Compute reward vectors and derive evaluation scores Start Accept via simulated annealing? Or have we reached n trials? End processing (Optimal prompt reached) Yes No No improvement after x simulatedannealing iterations Yes No Modify the current solution slightly and rerun Compute Hoeffding-UCB values to identify lowand highcontribution tokens; select two tokens to delete and one to add in the question text 6 collected exemplary inspection findings. Photo 1 (a) depicts salt damage on a PCT girder bridge observed during routine inspection. Photo 1 (b) represents a post-earthquake inspection scenario in which scaffolding and bridge inspection vehicles cannot be deployed, ground-level inspection is infeasible, and the condition must be inferred from the bridge deck. Photo 1 (c) shows corrosion at the end of a steel plate girder during routine inspection. Photo 1 (d) depicts bending-shear failure originating from the lap-splice region, where the amount of main reinforcement is reduced, characteristic of post-earthquake damage in RC piers. The photos are cited from References (NILIM & PWRI., 2011; NILIM & PWRI., 2013; NILIM & PWRI., 2017). During the semi-structured interviews with experienced engineers, span length, superstructure type, and bearing type were displayed in text at the bottom of the screen, while the instruction “この写真で気づいたことを全て述べてください。” (“Please state everything you notice in this photo”) was displayed in Japanese at the top. There is a potential risk of overfitting when relying solely on the dataset comprising four photos and the corresponding findings provided by experienced engineers. However, because Photo 1 covers both routine and seismic conditions, targets multiple structural components, and represents entirely different inspection scenarios, it was hypothesized that overfitting to a specific situation could be avoided. Nevertheless, as this remains only an assumption, to mitigate this risk, verification of generalization performance was additionally conducted, as described in Chapter 6. This approach was adopted because preparing a large dataset of detailed inspection findings for bridge damage photos is not practically feasible. As this study primarily emphasizes the procedure for constructing optimal prompts, it is desirable in practical applications to appropriately expand the dataset. PHOTO 1 Photo entered into ChatGPT. 7 The exemplary responses from experienced engineers, both in the original Japanese and in English translation, are provided in APPENDICES A. In PWRI (2025), semi-structured interviews were conducted with twelve experienced engineers, and the translated responses presented here correspond to the selected exemplary findings. The procedure for selecting these exemplary answers follows that described in PWRI (2025). The scores of other engineers’ responses were benchmarked against these four exemplary findings, which represent both routine and post-earthquake inspections. It has been verified that the hypothetico-deductive method can be applied by experienced engineers in both types of inspections. Figure 2 presents examples of question prompts. The full text of the prompts is provided in Appendix B. Each prompt consists of a mandatory element Z and supplementary blocks A through F. For block B, two variants were prepared for routine inspections and two for post-earthquake inspections (four variants in total). For block F, two variants with differing amounts of supplementary text were prepared. For example, possible combinations include “Z–C–D–E” or “Z–A–E–F2.” We search for the prompt that maximizes the response score among all prompts formed by combining these blocks. All sentences in Appendix B were authored by the author. Although Figure 1 suggests that finer-grained (e.g., sentence-level) optimization might be feasible, the present study focuses on the methodology; therefore, optimization is performed at the block level (A–F). FIGURE 2 Examples of question sentences. In Appendix B, block A specifies the role, and the B-series blocks set the context. Because the context differs between routine and post-earthquake inspections, the input text was varied for each image. In addition, since the images depict 【Z】#Instructions: Based on the hypothetico-deductive method, logically derive a course of action—both its necessity and details—from the damage observed in the photos, taking into account the performance of the road bridge (including load bearing capacity, durability, and other factors such as traffic safety, damage to third parties, drivability, recoverability, etc.), using your expertise. 【A】#Setting the role: You are an expert on road bridges. 【B1】#Setting the scene: Assume a situation where an emergency inspection is conducted after an earthquake. 【C】#Definitions of Terms: Define the words used in the description of the hypothetical-deductive method as follows: 1) “Evidence” as facts that can be seen from the photo. 2) “Argument” is what is not in the photo but can be inferred from your knowledge, or what can be inferred by combining your knowledge with the facts known from the photo. 3) “Conclusion” is a statement about the performance of the road bridge that can be inferred by combining the evidence and arguments. 4) The “course of action” is a reasonable course of action that can be logically derived from the evidence and arguments, especially for the poor performance of the bridge in the conclusion. ・・・ →Standard Question Supplementary text AtoFforStandard Question 8 road bridges in Japan, we varied whether to include the phrase “in Japan.” Block C defines the terminology used in the hypothetico-deductive method assumed in this study. Block D provides supplementary explanations of the hypotheticodeductive method, and block E offers further elaboration, including analogies. As the level of detail that yields the best responses is unknown, we provided the explanations as fully as possible. The F-series blocks specify the reasoning procedure. While there are inherent limits to enforcing strictly specified causal relationships, these blocks are intended to encourage such presentation; additional explanatory variants were also prepared for the F-series. We first evaluated the nine combinations listed in Table 1 and used the question achieving the highest score as the initial incumbent. As shown in Figure 1 and Section 3.1.5, starting from the block combination associated with the updated global best score, we estimated each block’s contribution to the score based on the Hoeffding-UCB value, selected two deletion candidates and one addition candidate, and iteratively searched for the optimum. TABLE 1 Combinations of Initial Sentences (9 Types). No. Combinations Additional explanation 1 Z-A-B2orB4-C-D-E-F1 Reference Sentence Combination 2 Z-B2orB4-C-D-E-F1 Reference Sentence Combination (excluding only Sentence A) 3 Z-A-B1orB3-C-D-E-F1 Reference Sentence Combination (excluding B2 or B4 and including B1 or B3) 4 Z-A-C-D-E-F1 Reference Sentence Combination (excluding only the B-series sentences) 5 Z-A-B2orB4-D-E-F1 Reference Sentence Combination (excluding only Sentence C) 6 Z-A-B2orB4-C-E-F1 Reference Sentence Combination (excluding only Sentence D) 7 Z-A-B2orB4-C-D-F1 Reference Sentence Combination (excluding only Sentence E) 8 Z-A-B2orB4-C-D-E-F2 Reference Sentence Combination (excluding F1 and including F2) 9 Z-A-B2orB4-C-D-E Reference Sentence Combination (excluding only the F-series sentences) 3.1.2 Setting of Evaluation Criteria for Responses In this study, responses were evaluated and optimized from three perspectives, by comparing them with exemplary answers: (1) semantic similarity score (𝑆), (2) logical consistency score (𝑆), and (3) hallucination suppression score (𝑆). However, these three criteria may not fully capture response quality, and establishing more comprehensive evaluation criteria remains a task for future research. The workflow was designed such that the algorithm and flow in Figure 1 can be applied even if additional criteria are incorporated. For the 𝑆 , Sentence-BERT with cosine similarity was employed. Specifically, we used the lightweight 384dimensional model all-MiniLM-L6-v2 (Sentence-Transformers., 2021). Let the embedded vectors of two sentences extracted by Sentence-BERT be and . The reward vector based on normalized cosine similarity is expressed in Equation (1) (Cer et al., 2017). 𝑆󰇛𝑢,𝑣󰇜∙ ∥∥∥∥ (1) 9 For the 𝑆 , we employed RoBERTa-large-MNLI (Facebook AI., 2022) from the Multi-Genre Natural Language Inference (MNLI) task to assess logical relations between texts. The exemplary answer was treated as the premise, while the response under evaluation was segmented into sentences and treated as the hypotheses. For each pair, the three-class probabilities were obtained via the softmax function, and only the probability of “entailment” was extracted, as expressed in Equation (2). By adopting a bidirectional entailment formulation rather than a unidirectional one, prior studies (Khobragade et al., 2019; Bhuyan et al., 2023) have reported results more closely aligned with human judgment of logical consistency, which informed our approach. Formally, this is denoted as 𝑆 , though elsewhere it is abbreviated as 𝑆. 𝑆󰇛𝑝,ℎ󰇜󰇛,󰇜󰇛,󰇜  (2) The 𝑆 was calculated by combining two detection measures: the Self-Contradiction Rate, indicating the degree of internal inconsistency within the response, and the Unsupported-Claim Ratio, representing the degree of inconsistency with the exemplary answer. These were combined according to Equation (3). The weighting parameter has no fixed value; in this study, it was set to 0.4, thereby prioritizing consistency with the exemplary answer. 𝑆 1󰇛𝜆𝑞󰇛1𝜆󰇜𝑟󰇜 (3) The Self-Contradiction Rate was calculated by segmenting the evaluated response into sentences, and then, for all sentence pairs, determining the proportion where the probability of “contradiction” exceeded a threshold of 0.5 (Li et al., 2024), using RoBERTa-large-MNLI (Facebook AI., 2002). The Unsupported-Claim Ratio treated the exemplary answer as the premise and the evaluated sentences as hypotheses. For each pair, the entailment probability was computed, and the proportion of sentences with entailment probability below the 0.5 threshold was obtained (Niu et al., 2024). The threshold of 0.5 corresponds to equal or greater weight under the softmax output and is regarded as a standard in the literature. All three scores range from 0 - 1. While both 𝑆 and 𝑆 utilize NLI entailment probabilities, they differ in how sentence pairs are compared: 𝑆 directly uses the entailment probability, whereas 𝑆 evaluates the proportion of sentences exceeding the threshold. 3.1.3 Verification of the Validity of Evaluation Criteria As noted in Section 3.1.2, similarities in using NLI entailment probabilities raise the possibility of multicollinearity between 𝑆 and 𝑆. Accordingly, this section examines the validity of these two criteria. Specifically, we: (1) 𝑆 isolated the effect of using the MNLI model in both criteria, and (2) 𝑆 applied partial correlation to eliminate the influence of overall response quality, given that high-quality responses tend to yield high scores across all criteria. For this analysis, we used the exemplary answers from one experienced engineer for the four photos, together with the responses of the other eleven experienced engineers. For these responses, we computed the three scores (1)–(3) along with BLEU and ROUGE scores. Partial correlation coefficients were then derived under two control conditions: (i) controlling for only (1), representing an evaluation emphasizing deep semantic similarity, and (ii) controlling for the average of (1) and the three surface-level scores (BLEU and ROUGE), representing an evaluation that incorporates 16 to recognize that hallucinations may not be mitigated, and to incorporate safety-oriented measures such as RetrievalAugmented Generation (RAG) in conjunction with hallucination-reduction algorithms. Addressing this limitation constitutes a subject for future work. 5. Comparison of Two Algorithms with Different Reward Functions Figures 8 and 9 present a comparison of two algorithms with different reward functions. As noted earlier, the overall flow shown in Figure 1 remains unchanged; however, this comparison concerns algorithms employing different evaluation scores, as described in Section 3.1.3. The results indicate that, for both LLMs, Score A reached the optimal solution more rapidly and also converged earlier. Examining the changes in scores across evaluation metrics reveals that 𝑆 remained largely flat but showed significant variability. 𝑆 exhibited a tendency for scores to increase abruptly depending on the combination of question sentences. 𝑆 showed a general upward trend in scores as the number of trials increased. FIGURE 8 Comparison of two algorithms with different rewards. (o4-mini) (a) and (c) show the score trends for Score A. (b) and (d) show the score trends for Score B. 0 0.2 0.4 0.6 0.8 0 5 10 15 20 25 30 35 40 45 50 Score number of trials S_sim S_ent S_hal 0 0.2 0.4 0.6 0.8 0 5 10 15 20 25 30 35 40 45 50 Score number of trials S_sim S_ent S_hal (a) (c) (b) (d) 0.9 1 1.1 1.2 1.3 0 5 10 15 20 25 30 35 40 45 50 Score A number of trials Best Score -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0 5 10 15 20 25 30 35 40 45 50 Score B number of trials Best Score 17 FIGURE 9 Comparison of two algorithms with different rewards. (GPT-5 Thinking) (a) and (c) show the score trends for Score A. (b) and (d) show the score trends for Score B. 6. Evaluation of Generalization Performance Although the prompt was optimized using three intrinsic evaluation metrics (𝑆, 𝑆, 𝑆), there is a risk that overfitting to these criteria could reduce generalization performance. To assess this risk, this section re-evaluates responses generated by the optimal prompt using photos depicting bridge damage that were not employed in the selection of the optimal prompt. The LLM used for this verification was o4-mini. The procedure involved inputting a new photo dataset along with the optimized prompt and the zero-shot prompt. The generated responses were evaluated using five external quality metrics, and the statistical significance of improvements over the zero-shot baseline was assessed using the Wilcoxon signed-rank test. The additional bridge-damage photos were randomly sampled with Python from References (NILIM & PWRI., 2011; NILIM & PWRI., 2013; NILIM & PWRI., 2017), selecting 30 photos each for routine and post-earthquake inspections that did not contain embedded text. Since no responses from experienced engineers were available for these photos, evaluation using the intrinsic metrics (𝑆, 𝑆, 𝑆) was not feasible. The external quality metrics in this section were defined across five dimensions: (1) lexical coverage, (2) level of detail, (3) readability, (4) lexical diversity, and (5) internal consistency. Specifically: (1)Lexical coverage (domain coverage) (Ryan et al., 2023) was quantified as the proportion of 800 specialized terms appearing in the generated responses, with the lexicon extracted mechanically from existing U.S. and Japanese manuals for periodic and post-earthquake bridge inspections (e.g., Lebret et al., 2016; NILIM, 2023; Buckle et al., 2006, among others). (2) Level of detail (Rubino et al., 2016) was measured by the density per 100 words of numerical values, dimensional units, and prepositional (a) (c) (b) (d) 0.9 1 1.1 1.2 1.3 0 5 10 15 20 25 30 35 40 45 50 Score A number of trials Best Score -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0 5 10 15 20 25 30 35 40 45 50 Score B number of trials Best Score 0 0.2 0.4 0.6 0.8 0 5 10 15 20 25 30 35 40 45 50 Score number of trials S_sim S_ent S_hal 0 0.2 0.4 0.6 0.8 0 5 10 15 20 25 30 35 40 45 50 Score number of trials S_sim S_ent S_hal 18 expressions indicating spatial relationships. (3) Readability was assessed using the Flesch Reading Ease (Flesch., 1948) normalized to a 0 – 1 scale, with higher scores representing easier readability. (4) Lexical diversity was evaluated using the Moving-Average Type–Token Ratio (MATTR) (Covington & McFall., 2010) with a 50-word sliding window. (5) Internal consistency (Glockner et al., 2018) was measured by rule-based detection of antonym/negation pairs across sentences, with a consistency score derived from the proportion of “self-contradictory” pairs. Table 3 presents the results of the Wilcoxon signed-rank test comparing responses generated by optimal prompts against those generated by zero-shot prompts. For the dataset of routine inspection photos, responses using the optimal prompt significantly outperformed those from the zero-shot prompt in lexical coverage, level of detail, lexical diversity, and internal consistency, while no significant difference was observed in readability. For the post-earthquake inspection dataset, significant improvements were also observed in lexical coverage, lexical diversity, and internal consistency, though no significant difference was found in level of detail, and readability again remained unchanged. In summary, the inspection reports generated using the optimal prompts developed in this study demonstrated statistically significant improvements, particularly in lexical coverage, lexical diversity, and internal consistency. Importantly, despite being optimized with respect to three intrinsic evaluation criteria, the prompts also yielded improvements in certain external metrics. This indicates that the effectiveness of Prompt Engineering does not depend solely on specific internal evaluation metrics, but rather demonstrates generalizability in enhancing the quality (i.e., generalization performance) of outputs in practical tasks. TABLE 3 Result of Wilcoxon signed-rank test. Indicator Regular Inspection Dataset Post-Earthquake Inspection Dataset p R p r Coverage 0.00 78.42 0.00 84.90 Specificity 0.00 80.15 0.49 42.72 Readability 1.00 0.00 1.00 0.00 Diversity 0.00 84.90 0.00 83.44 Consistency 0.05 10.95 0.00 30.86 5. CONCLUSION 1. This study proposed a method for exploring optimal prompts in prompt engineering to facilitate the application of Generative AI to road bridge maintenance. 2. Using prompts designed to encourage hypothetico-deductive reasoning produced responses from Generative AI that exhibited semantic similarity closer to those of experienced engineers, and achieved scores indicating logical consistency that were 1.2–1.4 times higher than those of experienced engineers. 3. By contrast, the proposed prompts did not exhibit improvement in hallucination suppression. 4. Findings generated using the optimal prompts developed in this study demonstrated statistically significant 19 improvements in lexical coverage, lexical diversity, and internal consistency compared to those generated using zero-shot prompts. These results confirm the generalization performance of the proposed prompts, particularly with respect to these dimensions. 5. While we showed that prompt engineering alone can enhance the reasoning ability of models that have undergone large-scale, comprehensive pretraining in specialized domains, it did not reduce hallucinations under the conditions examined here. Taken together, these findings provide quantitative evidence that this class of intelligence has strengths and weaknesses that differ clearly from those of humans, particularly with respect to the relative ease of improving certain capabilities. 20 APPENDICES A (Translated English) This is a PC bridge with a span of 40m, and there are cracks in the main girder in the direction of the bridge axis, and cracks in the intermediate transverse girder in the axial direction and on the underside. There is concrete spalling. As there are cracks on the underside, there is a possibility that the steel material is corroding and expanding due to salt damage or neutralization. There are no problems with the floor slab. As the main girder, transverse girder and floor slab are all covered with a protective material, there is a possibility that they were reinforced and repaired due to salt damage or neutralization. As the concrete has been repaired in places where it has lifted, there is a possibility that the cross-sectional repair material has deteriorated again. As there are no bending cracks in the main girders, there is a possibility that the load-bearing capacity has not decreased. The position of the PC steel material is sound, and there is a possibility that only the surrounding axial direction steel bars have corroded. If it is due to salt damage, the T-shaped cross-section is likely to trap airborne salt, and as the area above the river is wet and the cracks are wide, there is a possibility that durability has decreased. We believe that it is not necessary to take emergency measures such as closing the road to traffic at this stage. However, as the concrete has risen to the surface, we believe that it is necessary to take measures such as knocking it off from the viewpoint of preventing damage to third parties. (Original Japanese text) 支間長 40m のPC 橋で，主桁下面に橋軸方向のひび割れ，中間横桁に軸方向と下面にひび割れが生じている．コンクリートの浮きが 生じている．下面にひび割れが生じているため，塩害もしくは中性化により鋼材が腐食膨張している可能性がある．床版は問題が生 じていない．主桁と横桁と床版の全てが被覆材で覆われているため，塩害もしくは中性化で補強補修した可能性がある．コンクリー トの浮きが生じている箇所も補修しているため，断面修復材の再劣化が生じている可能性がある．主桁に曲げひび割れが生じていな いため，耐荷力は低下していない可能性がある．PC 鋼材位置は健全で周りの軸方向鉄筋のみが腐食している可能性がある．塩害であ る場合，T型断面は飛来塩分を巻き込みやすく，河川上は湿潤であり，ひび割れ幅が大きいため，耐久性が低下している可能性があ る．緊急措置としての通行止め等は必要ない段階であると考える．ただし，コンクリートの浮きが生じているため，第三者被害の防 止措置の観点で叩き落とし等の対策を取る必要があると考える． FIGURE A.1 Model answer from experienced engineer (Photo 1 (a)) (Translated English) The axis of the parapet is misaligned. There is a possibility that the abutment is in the foreground due to the inspection path. There is a possibility that relative displacement has occurred between the superstructure and the substructure due to the misalignment of the axis of the parapet and the cutting of the pipe for the attached structure. There is a possibility that the span side has moved due to the curbstone being moved to a different position than usual. There is a possibility that damage has occurred due to the earthquake because there is also an abnormal gap in the expansion device. The expansion joint is misaligned in the direction perpendicular to the bridge axis and in the vertical direction, so there is a possibility that the lower structure has moved or the bearings have been destroyed. There is a possibility that the rideability has deteriorated. If the bearings have been destroyed, there is a possibility that the load-bearing capacity has decreased. We think that it is necessary to eliminate the bumps and ensure rideability as a matter of urgency. After that, we think that it is necessary to carry out a detailed investigation. (Original Japanese text) 壁高欄の軸線がずれている．検査路があるため，手前がアバットである可能性がある．壁高欄の軸線のずれと添架物用のパイプの切 断が生じているため，上部構造と下部構造の相対変位が生じた可能性がある．縁石が飛んでいるため，径間側が動いた可能性がある． 伸縮装置の遊間異常もあるため，地震による損傷が生じている可能性がある．伸縮装置は橋軸直角方向と上下方向がずれているため， 下部工の移動やピン・ローラー支承の破壊が生じている可能性がある．走行性の低下が生じている可能性がある．ピン・ローラー支 承が破壊しているのであれば，耐荷力が低下している可能性がある．緊急的には段差を解消し，走行性を確保する必要があると考え る．その後，詳細な調査を行う必要があると考える． FIGURE A.2 Model answer from experienced engineer (Photo 1 (b)) (Translated English) This is the end of the girder. There is partial corrosion and a reduction in the cross-section of the lower flange and web of the main girder. The bearings are white, so there is a possibility that they have been hot-dip galvanized. There is partial corrosion and a reduction in the cross-section of the lower flange and web of the main girder, so there is a possibility that water is flowing to the girder end due to leakage in the expansion joint. There is a possibility that the load-bearing capacity of the main girder is insufficient due to the reduction in cross-section. Since no bending moment is applied to the end of the girder, we think it is necessary to check the vertical load-bearing capacity of the support point, including the vertical stiffeners. To do this, we think it is necessary to check whether the cross-sectional area of the lower edge of the web and the cross-sectional area of the bottom of the vertical stiffeners are secured. If the cross-sectional reduction is large, we think that the girder will not collapse when the road is reopened to traffic, but there is a possibility that the load-bearing capacity at the support point will be insufficient. Therefore, it may be necessary to prevent sudden changes in level by using saddles. (Original Japanese text) 鈑桁の桁端部である．主桁の下フランジやウェブの部分的な腐食や断面減少がある．支承は白い色であるため，溶融亜鉛メッキの可 能性がある．主桁の下フランジやウェブの部分的な腐食や断面減少が生じているため，伸縮装置部の漏水により桁端部に水が流れて いる可能性がある．断面減少があるため，主桁の耐荷力が不足する可能性がある．桁端部には曲げモーメントがかからないので，鉛 直補剛材を含めた支点の鉛直方向の耐荷力を確認する必要があると考える．そのためには，ウェブの下縁の断面積や鉛直補剛材の一 番下の断面積が確保されているかを確認する必要があると考える．交通開放を行う場合に，断面減少が大きなものであるのであれ ば，桁が崩壊することは無いと考えるが，支点上の耐荷力が不足する可能性がある．ゆえに，サンドルによる待受けを行い，急激な 段差が起きることを防ぐ必要がある可能性がある． FIGURE A.3 Model answer from experienced engineer (Photo 1 (c)) 21 (Translated English) Partial destruction has occurred at the base of the pier. The concrete cross-section has been destroyed, and the steel bars have partially come loose. The cracks are at a 45-degree angle. The cause of these deformations may be the earthquake. The concrete on the tension side of the pier has been destroyed, and diagonal cracks at a 45-degree angle have occurred, so there may be no effective cross-section. From these deformations, it is possible that the load-bearing capacity in the vertical direction still remains. However, it is possible that there is almost no load-bearing capacity in the horizontal direction, so it may be dangerous to drive a vehicle on it. Therefore, it is better to contact the road administrator and ask them to decide whether to close the road to traffic. (Original Japanese text) 橋脚の基部に部分的な破壊が生じている．コンクリートの断面が破壊し，鉄筋も部分的にはらんでいる．ひび割れが斜め 45 度方向に 入っている．それら変状の原因は地震の可能性がある．橋脚の引張サイドのコンクリートが破壊し，45 度方向の斜めひび割れが生じ ているため，有効断面がない可能性がある．これら変状から，鉛直方向の耐荷力はまだ残存している可能性がある．しかし，水平方 向の耐荷力がほぼない状態である可能性があるので，車両を走行させることは危険である可能性がある．ゆえに，道路管理者に連絡 を行い，通行止めの措置を判断頂いた方が良いと考える． FIGURE A.4 Model answer from experienced engineer (Photo 1 (d)) 22 APPENDICES B 【Z】#Instructions Based on the hypothetico-deductive method, logically derive a course of action—both its necessity and details—from the damage observed in the photos, taking into account the performance of the road bridge (including load bearing capacity, durability, and other factors such as traffic safety, damage to third parties, drivability, recoverability, etc.), using your expertise. 【A】#Setting the role You are an expert on road bridges. 【B1】#Setting the scene: Assume a situation where an emergency inspection is conducted after an earthquake./ 【B2】#Setting the scene: Assume a situation where an emergency inspection is conducted after an earthquake in Japan. 【B3】#Setting the scene: Assume a situation in which you conduct regular inspections during normal times./【B4】#Setting the scene: Assume a situation in which you conduct regular inspections during normal times in Japan. 【C】#Definitions of Terms Define the words used in the description of the hypothetical-deductive method as follows: 1) “Evidence” as facts that can be seen from the photo. 2) “Argument” is what is not in the photo but can be inferred from your knowledge, or what can be inferred by combining your knowledge with the facts known from the photo. 3) “Conclusion” is a statement about the performance of the road bridge that can be inferred by combining the evidence and arguments. 4) The “course of action” is a reasonable course of action that can be logically derived from the evidence and arguments, especially for the poor performance of the bridge in the conclusion. 【D】#Explanation of the Hypothetico-Deductive Method Since the hypothetico-deductive method is logical reasoning, the validity of the conclusion presented depends on the truth of the evidence and arguments. In the hypothetico-deductive method, inductive reasoning (hypothetical method) and deductive argumentation (e.g., categorical syllogism) change according to the order of “evidence” and “argument”, and the degree of validity of the “conclusion” changes. In the case of inductive reasoning in the order of “evidence,” “argument,” and “conclusion,” the conclusion is only a hypothesis because it is based on knowledge, which is the argument. In the case of deductive reasoning in the order of “argument,” “evidence,” and “conclusion,” the logical premise of the argument is accepted as true, and the conclusion is drawn from the facts on which it is based, so the process of drawing the conclusion is an argument, and a more valid conclusion is obtained. 【E】#Further Supplementary Explanation For example, if, as a hypothetical method, you infer additional damage (argument) based on your knowledge from the damage in the image (evidence) and draw a conclusion, you have presented a hypothesis. If, through subsequent investigation, “other damage” is found that does not exist in the image initially presented but that supports the argument (or if it is assumed that “other damage” has been found), the argument becomes true, and a deductive argument is established in the flow of “argument,” “evidence,” and “conclusion. In short, it is necessary to determine the truth or falsehood of propositions and to be aware that even in sentences that can be considered semantically the same based on the semantic connection of those propositions, the robustness of the conclusion can change due to differences in syntax. In this logical inference, we will support the syntactical validity of the conclusion by using a sound syntax-aware procedure to draw the conclusion as robustly as possible. It also supports the semantic validity of the proposition by determining its truth or falsehood. By going through these steps, you should present a semantically and syntactically valid conclusion and corresponding course of action. That is your mission. 【F】#Reasoning Procedure Please follow the steps below to perform logical reasoning. 1.**Extract the factual basis**. Identify all possible grounding propositions using capital letters such as A, B, C, etc. The propositions here should be all the facts known from the photo. Avoid using negative expressions such as “not” or “no” in your propositions. For example, if the evidence shows that a beam has cracks, express it as: A: This beam shows cracks. Its negation is expressed as ¬A (i.e., “This beam does not show cracks”). In identifying propositions, carefully observe all members and sections of the road bridge that can be seen in the photos for the presence or absence of damage and the location (e.g., which part of the beam has cracks, whether the damage is at the end or base, whether the damage is in the central section of the member, etc.) and direction (e.g., direction of cracks, direction of member movement, direction of misalignment, etc.) of the damage, and then determine the direction of damage. ) and direction (e.g., direction of cracking, direction of member movement, direction of misalignment, etc.) should be carefully observed and noted. The location and direction of damage are important clues for presenting arguments in Section 2. Do not mention any components that are not shown in the photos. 2. **Present the argument from the evidence**. From the factual propositions verbalized in 1. above, please present the propositions of your argument by deepening your reasoning based on your knowledge. In doing so, please identify all possible argumentative propositions based on your knowledge using lower case alphabetical letters such as a, b, c, etc. Do not include negative tones such as “not” or “no” in your propositions. When negation is used, use a symbol to indicate the negative form. For example, the negative form of proposition a can be expressed as ¬a. Also, indicate which rationale the argument corresponds to. Use arrows (→) to indicate causality, for example, “If c , then A”, “A causes c”, etc. can be expressed as A → c. Logical symbols such as A∧B→c may also be used here. An argumentative proposition is something that cannot be known from a photo, but can be inferred based on facts and knowledge. Please present any possible damage that can be imagined from the damage to the components in the photo, etc. Please present arguments for damage inside members that can be inferred from visual information on the exterior, or abnormalities in load transfer paths between adjacent members (floor plate ⇔ main girder ⇔ transverse girder, superstructure ⇔ bearing section ⇔ substructure ⇔ surrounding ground) that can be inferred from the evidence (facts). Also, based on the location and direction of the damage presented in section 1, please carefully draw inferences about the internal condition of the member and the positional relationship between adjacent members. Also, mention factors that can be inferred from information that can be visually observed on the exterior (e.g., action of active loads and seismic motions, and deterioration factors such as moisture and salt). When presenting arguments, please emphasize inferences based on evidence (facts) and be careful not to present arguments without evidence. The “components not shown in the photo” at this time should be all components of the road bridge, such as beams, transverse girders, floor plates, expansion devices, etc. of the superstructure, all components of the bearing section (upper shoes, anchor bars, mounting bolts, etc.), and all components of the substructure (piers, abutments, foundation, surrounding ground, etc.). 3. **Draw conclusions**. Draw conclusions by being aware of the causal relationship between evidence that is fact and arguments based on knowledge, and the causal relationship between arguments and conclusions. Present as many conclusions as possible with complete evidence and arguments; from propositions 1 and 2, analyze them carefully and find causal relationships between the propositions. The conclusion should refer to the performance of the road bridge, with concluding propositions numbered as [1], [2], [3]. For instance, arrange the logical flow as A → c → [1] (where A is a factual proposition and c is an argumentative proposition) to verify syntactic validity. You may also use symbols like A ∧ B → c → [2] if multiple facts lead to the argument. 4. **Provide assumptions, arguments, and a course of action**. Based on the logical reasoning in section 3, please present a reasonable course of action. The course of action should be presented in the form of [1], [2], and [3]. Please propose a course of action assuming that additional research has revealed the truth of the argument. Also, please be aware of the syntax of deductive argumentation (“Argument” -> “Evidence” -> “Conclusion”). 5.**Summary**. Please summarize your findings and proposed course of action in an easy-to-understand sentence and explain them as your final answer or proposal. For example, please make a proposal such as “If c was confirmed, then we can infer [2] from A, in which case we need to do [2]”. FIGURE B.1 All sentences in the question. F1 text includes underlined sections. F2 text does not include underlined sections. 23 APPENDICES C This image appears to show the underside (“soffit”) of a short-span bridge or similar deck structure, taken from the riverbank or water’s edge looking up into the beam-and-girder system. Here’s what I notice: 1. Structural System Primary and secondary girders You can make out a deeper, wider beam running parallel to the camera (just above center), with several smaller “cross” beams (secondary girders) framing into it at right angles. The beams are most likely precast or cast-in-place concrete (or possibly painted steel) given their straight, uniform shape and the way they intersect. Cracking patterns There are several long, thin cracks running along the soffit of the beams. On the large central girder you can see a near-horizontal crack that follows the beam-to-beam junctions and also a diagonal crack that snakes along its flange. Similar slender cracks are visible on the crossbeams at right, suggesting either flexural cracking from bending stresses or shrinkage/delamination. The fact that cracks traverse multiple members could indicate either structural overloading, fatigue, or long-term shrinkage/corrosion of embedded reinforcement. Inspection markings/graffiti Just above the main crack in the central beam there’s some black writing, probably left by an inspector or engineer. It reads something like “G1–G5 Fin P…” which may correspond to grid lines or member identifiers and a date or the initials of the inspector. 2. Corrosion and Staining Rust stains Upper surfaces of the beams show rusty orange patches, especially near the top right corner. That suggests corrosion of embedded steel reinforcement (rebar) or possibly of a steel girder beneath the paint. On the right-hand column/beam you can see vertical drips of rust staining the lower flange. Water infiltration The presence of streaks implies that water is running along the beam undersides, carrying dissolved iron oxides (rust) out onto the surface. On the far left you can glimpse the edge of the bridge deck and signs of moisture ingress at that joint. 3. Substructure and Surroundings Support piers/piles In the lower portion of the photo are two dark-colored, cylindrical or rectangular piers/piles. Their black coating or wrapping may be a protective fender system (to prevent boat impacts) or simply a coating against corrosion. Behind them, you can see a stone-masonry retaining wall or abutment with regularly coursed blocks, typical of riverside embankments. Environment The mottled wet/dry look of the masonry and concrete suggests tidal or fluctuating water levels. Algae or water-line marks a few feet up on the masonry hint at where the water usually reaches. 4. Material Condition and Maintenance Surface preparation & paint Some beams look freshly painted (clean white), while adjacent faces show peeling or fading paint, indicating patch repairs or only partial maintenance. The combination of new and old coatings suggests ongoing maintenance but also that deterioration continues faster than full recoating. Concrete spalling or delamination While there aren’t large chunks missing, the thin cracks could lead to spalling over time if left unaddressed, exposing the steel to more moisture and accelerating corrosion. Overall Impression: This is a classic inspection-style photo of a bridge underside, documenting distress (cracks, corrosion, staining) in the beam-and-girder system over a waterway. The inspection markings, mixed paint conditions, and visible reinforcement corrosion all point to a structure that’s under periodic monitoring but showing signs of long-term deterioration from moisture, load cycles, and possibly chloride attack (from de-icing salts or a marine environment). Let me know if you’d like to dive d eeper into any particular aspect — crack causes, remediation strategies, or the structural details! (a) Answers using zero-shot prompts Here’s a structured application of the hypothetico–deductive method to the damage shown in the photo of the bridge underside: 1. Evidence Longitudinal cracking running the full length of the bottom flange of the girder (approx. 5–10 mm wide in places). Rust staining and surface corrosion around the crack and at beam-to-bearing interfaces. Spalling of protective coating and minor concrete delamination at the bearing seats. No immediate section loss visible, but signs of moisture ingress (water marks, efflorescence on the masonry pier below). 2. Argument Crack morphology (long, continuous) is characteristic of flexural fatigue or overstress, not just shrinkage. Surface corrosion adjacent to cracks means chlorides/water are penetrating through the crack, accelerating steel section loss. Coating failure at bearings and cracks allows repeated wet–dry cycles, undermining durability and promoting corrosion fatigue. Undetected section loss could already have reduced the flange’s effective thickness and thus its moment capacity. Traffic vibrations and cyclic live loads (vehicles, possibly heavy trucks) will propagate fatigue cracks if unmitigated. 3. Conclusion (Bridge Performance) The bridge’s load‐bearing capacity is compromised by fatigue‐type cracking and incipient section loss in the girder flange. Its durability is at risk due to ongoing corrosion, and its traffic safety margin is diminished—there is a realistic danger of sudden crack propagation under repeated live loads. 4. Course of Action Action Step Rationale Details & Timing 1. Immediate Load Restriction Prevent further crack growth under heavy loads • Post a temporary axle‐load limit (e.g. 10 t) and/or single-lane closure within 24 h. 2. Detailed NDT Inspection Quantify section loss and crack depth • Within 1 week, perform ultrasonic thickness mapping and magnetic‐particle testing along the crack. 3. Crack Repair & Protection Restore section and seal ingress paths • Inject low-viscosity epoxy into the crack to bond and seal. • Apply new corrosion‐ inhibiting primer and high-build coating to the entire flange. 4. Structural Strengthening Re-establish original flexural capacity • If NDT shows >10 % section loss, install steel bolted cover plates underneath the flange, sized to restore original moment capacity. • Complete within 1 month of inspection. 5. Improve Drainage & Coating Prevent recurrence of moisture‐driven damage • Reprofile deck joints and bearing seals to divert water. • Recoating schedule: every 5 years. 6. Monitoring & Maintenance Track crack behavior and corrosion progression • Install crack‐width gauges and corrosion sensors; read monthly for first 6 months, then quarterly. • Include this location in biannual routine inspections. Necessity: Without immediate load restrictions and repair, the fatigue crack will grow under traffic loads, potentially leading to brittle fracture of the girder flange and catastrophic loss of capacity. Corrosion will further undermine the section, accelerating failure. Detail Highlights: Load restriction buys time for inspection and repair. NDT provides the quantitative basis for deciding between simple crack sealing vs. full strengthening. Epoxy injection and recoating arrest corrosion, restoring durability. Cover plating is a proven, rapid strengthening measure where section loss is significant. Drainage improvements cut off the moisture source. Monitoring confirms the effectiveness of repairs and schedules future maintenance before damage recurs. This plan logically follows from the observed damage (evidence) and the engineering principles of fatigue, corrosion, and load capacity (argument), to conclude that the bridge is in a degraded state and must be temporized, inspected, and rehabilitated (conclusion → course of action). (b) Answers using prompts that encourage hypothetical deductive thinking FIGURE C.1 Examples of Generative AI responses. Photo 1(a). Passages suspected to be hallucinations are indicated in blue text with underlines. 24 This image appears to have been taken from the deck of a highway bridge looking along its length toward a distant mountain ridge. Here’s everything I notice: 1. Bridge Deck and Pavement Asphalt Surface: The roadway is laid with fresh-looking black asphalt. Lane Marking: A solid white edgeline runs parallel to the concrete parapet, indicating the shoulder line of the carriageway. Expansion Joint: Beneath the asphalt, at the very edge of the deck, you can see the corrugated steel fingers of an expansion joint. 2. Concrete Parapet and Guardrail Parapet Wall: A concrete barrier (parapet) about 30–40 cm wide sits at the extreme edge of the deck. Guardrail Posts: Vertical metal posts—some heavily rusted at their bases—are embedded in the parapet. There is no horizontal rail between them, suggesting the guardrail has been removed or was never installed. Surface Damage: Rust staining around the post bases and two chunks of loose concrete (or perhaps aggregate stones) lie on the deck and parapet cap, implying minor spalling or break-out at the parapet edge. 3. Maintenance Ladder and Cage Ladder Cage: On the right side of the frame is the upper portion of a fixed steel ladder surrounded by a safety cage. This provides access from the deck down the exterior face of the bridge. Piping and Conduit: A thin metal pipe or conduit runs horizontally alongside the cage, fastened to the parapet and ladder posts—likely for electrical wiring or instrumentation. 4. Surrounding Landscape Valley Below: Beyond the parapet, a lush green valley stretches out, with grassland and scattered trees. Hills and Ridge: Rolling, wooded hills rise on both sides; the far ridge forms the vanishing-point of the bridge deck. Overcast Sky: The light appears diffused, suggesting a cloudy or lightly overcast day. 5. Human Presence Footwear: At the left edge of the photo, you can see a person’s lower leg and foot, clad in dark trousers and a white-and-green slip-on shoe—perhaps an inspector or pedestrian. Hand: Just above the shoe, part of a hand and wrist is visible, reinforcing that this is a ground-level vantage. Overall impression: This is a close-to-the-edge viewpoint from a vehicular bridge, capturing both the structural details of its parapet, guardrail posts, and expansion joints, and the maintenance access ladder, against a backdrop of a green valley and distant hills. The presence of loose concrete fragments and heavily rusted posts hints at mild deterioration, while the absence of a continuous guardrail might be a temporary condition or ongoing maintenance. (a) Answers using zero-shot prompts Below is a structured hypothetico‐deductive reasoning, leading from what we see in the photo to a clear course of action for this bridge after an earthquake. 1. Deductive Chain of Reasoning Argument (General principle) An earthquake subjects a bridge to large inertial forces; if bearings or expansion joints fail, deck segments can shift and concrete cover will spall. Loss of proper bearing support or joint function immediately degrades a bridge’s load-bearing capacity and continuity. Evidence (Photo facts) Concrete spalling at the deck edge: two sizeable chunks of wearing surface have detached and lie on the roadway. Exposed rust stains on vertical stirrups and at the bearing‐seat interface. Misalignment and an open gap at the deck expansion joint (visible sag to the right). Guardrail base plates badly corroded and potentially loose. Conclusion (Bridge performance) From these observations and the seismic‐damage principle, we conclude that: Load bearing is compromised: the bearing seat has likely lost contact or has shifted, so live loads cannot be safely transferred to the substructure. Durability is reduced: spalled concrete and exposed steel will accelerate corrosion, further weakening capacity. Traffic safety is unacceptable: falling debris, uneven driving surface, and failing guardrails pose immediate hazards to vehicles and pedestrians. 2. Course of Action Immediate Measures Close the bridge (or impose strict load-limitations, e.g. one lane, 3 ton max) to eliminate risk of collapse or falling debris. Install emergency shoring beneath the damaged span to restore at least a minimal load path. Erect barricades and warning signage to protect third parties from loose fragments. Detailed Structural Inspection Conduct a hands-on survey of all bearings and expansion joints (verify seat displacement, inspect rocker pads, record gaps). Perform non-destructive testing (ultrasonic, rebound hammer) on the spalled areas to assess concrete strength and rebar condition. Inspect adjacent spans and substructure for any secondary damage (pier tilting, foundation settlement). Emergency Repair Design Remove and replace the failed bearing assembly and expansion joint seal. Repair spalled concrete with high-strength patch mortar; apply corrosion-inhibiting coating to exposed rebar. Re-profile the asphalt wearing surface over a new neoprene or PTFE joint seal to ensure smooth drivability and watertightness. Strengthen guardrail anchorage and replace corroded base plates. Monitoring & Long-Term Rehabilitation Install displacement sensors across the joint and strain gauges on repaired sections to monitor performance under live loads. Schedule a full seismic retrofit, if not already in place, to bring the bridge up to current code—this may include new isolators or dampers. Plan for periodic inspections every 6 months for the next two years to watch for corrosion or settlement. Justification of Necessity Without immediate closure or shoring, the bridge could experience a sudden partial collapse under even moderate traffic. Prompt repair of bearings and spalled areas is essential to restore both load-bearing capacity and durability. Ongoing monitoring will catch any delayed failures from hidden damage, safeguarding travelers and extending service life. (b) Answers using prompts that encourage hypothetical deductive thinking FIGURE C.2 Examples of Generative AI responses. Photo 1(b) . Passages suspected to be hallucinations are indicated in blue text with underlines. 25 This photograph appears to show an interior corner of some kind of metal‐and‐concrete structure—perhaps a utility vault, tank, or the underside of a stair/platform. Here’s what stands out: Heavily corroded steel plate The large horizontal surface in the mid-foreground is a steel plate (probably a shelf or tread) whose entire upper edge is encrusted in orange-brown rust. You can see layers of flaky iron oxide built up in a roughly even band, indicating ongoing wet/dry corrosion cycles. The plate’s underside (visible just below) is darker and may be coated in oil or a mix of rust particles and water. Rust “waterline” staining On the vertical backing plate above the horizontal one, there’s a discernible “stain line” where rust has wicked up and down. That suggests intermittent flooding or condensation that wets the joint, dissolves iron oxides, and then redeposits them as water recedes. Paint degradation and rust underfilm In places you can still make out remnants of what was once a grey or light‐colored paint or protective coating. Rust is lifting and flaking the coating—classic underfilm corrosion. Concrete base with efflorescence Below the steel, the support rests on a concrete or masonry footing. The crisp white patches (particularly around the block and under the right‐hand supporting pad) look like salt deposits from groundwater seepage—that is, minerals carried in solution being left behind as water evaporates. Supporting bracket/bearing shoe On the right you see what looks like a small square pad or “shoe” holding the beam/shelf in place. It too is rust-stained at its edges, and there’s a large hex head bolt or plug nearby. General environment clues The overall look—bare concrete, heavy rust, white mineral streaks—points to a damp, poorly ventilated space. Possibly a subgrade utility vault, a water-holding tank, or underside of an industrial stair in a boiler room or similar. No insulation or rustproofing remains in evidence, so this has been going on for some time. Implications Structurally, the metal section is losing thickness and strength where the rust is thickest; if this is load-bearing, it’s a safety concern. The moisture ingress that’s feeding corrosion needs to be addressed (re-seal joints, improve drainage/ventilation, apply rust-inhibiting coatings). The mineral deposits on the concrete suggest water is still actively penetrating— potential for freeze/thaw damage or further chemical attack. In short, this is a case study in what happens when steel and concrete meet persistent moisture: you get heavy rust formation on the steel, lifting of protective coatings, and white “efflorescence” on the concrete, all of which signal a need for maintenance to prevent further deterioration. (a) Answers using zero-shot prompts Below is a structured hypothetico-deductive analysis of the damage shown in the photos, ending in a logically derived course of action. 1. Deductive Reasoning Chain Argument (premise from engineering knowledge): When steel members or bearings corrode, iron oxides (rust) occupy a greater volume than the original steel. This “rust jacking” exerts expansive forces on surrounding concrete or bearing plates, leading to cracking, spalling, and loss of bearing‐to‐seat contact. Corrosion also pits and thins the steel, directly reducing its cross-sectional area and hence its loadb earing capacity. Once water and chlorides penetrate past the protective paint or concrete cover, corrosion tends to accelerate, further degrading durability and fatigue resistance. Evidence (what we see in the photos): Heavy rust deposits layered along the steel flange and bearing area—brown/orange staining over a broad zone. Flaking and pitting of the steel surface, especially at the edge of the flange—indicating section loss. Concrete delamination and cracking immediately adjacent to the bearing, with rust stains tracking down—evidence of rust jacking and water ingress. Absence of intact protective coating over the affected steel, suggesting long-term exposure. Conclusion (performance inference): The bridge’s bearing assembly and adjacent steel– concrete interface have lost significant load‐carrying capacity due both to section loss in the steel and to uneven bearing contact caused by rust expansion. Durability is severely compromised, with ongoing corrosion likely to accelerate. Traffic safety is at risk: under heavy loads, the weakened elements could deform excessively or fail suddenly, and spalled concrete poses a hazard to passing vehicles or pedestrians. 2. Logically Derived Course of Action Immediate Safety Measures Post a temporary load restriction (e.g. reduce permitted axle loads by ≥ 25 %) until structural capacity is re-verified. Restrict lane usage under the most heavily corroded spans if spalled concrete debris could fall onto traffic. Detailed Assessment Perform a non-destructive thickness survey (ultrasonic or magnetic) of the corroded steel flange and bearing plates to quantify section loss. Conduct a bearing seat survey (e.g. using bearing displacement gauges) to confirm loss of uniform contact. Map and probe the extent of concrete delamination around the bearings with chain-drag or impact echo. Remedial Repairs Remove all loose rust and delaminated concrete, exposing sound steel and concrete. If section loss exceeds design allowable (e.g. > 10 %), replace or plate–reinforce the affected steel members or bearings. Rebuild bearing seats: install new elastomeric bearings or pack steel shims to restore correct alignment and load transfer. Patch repaired concrete with non-shrink, polymermodified mortar and re-seal edges to shed water. Long-Term Prevention Apply a high-performance epoxy or zinc-rich primer plus polyurethane topcoat to all exposed steel. Install a waterproofing membrane above the bearing area and improve drainage to keep water away. Consider cathodic protection for aggressive environments. Schedule bi-annual inspections of these repairs, with interim visual checks after major storms or de-icing cycles. Summary By starting from the general principles of corrosion mechanics (argument), matching them to the visible section loss, rust jacking, and concrete spalling in the photos (evidence), we arrive at the conclusion that both load capacity and durability are unacceptable in the current state. This logically leads to a staged course of action: emergency load control, thorough measurement, targeted replacement/repair, and renewed corrosion p rotection plus a tighter inspection regime. (b) Answers using prompts that encourage hypothetical deductive thinking FIGURE C.3 Examples of Generative AI responses. Photo 1(c) . Passages suspected to be hallucinations are indicated in blue text with underlines. 32 (Version 1, October 29, 2025)