Attempt-Level Logging for Embedded Assessment in a Web-Based Science Learning System for Students with Mild Intellectual Disabilities

2.1 Web-based systems as accessible learning infrastructure

Browser delivery has practical consequences for deployment. Nothing has to be installed on each device, no administrative privileges are required, and a responsive build adapts to whatever screen is available. Reviews of educational technology for learners with disabilities in low- and middle-income countries identify infrastructure and maintenance burden as the main constraint on sustained use [3]. Special education has moved from purpose-built hardware towards general-purpose devices [14], including consumer tablets [15]; browser delivery removes the remaining dependence on a particular device or operating system.

Pairing words with relevant imagery and routing verbal information through the ear spreads processing across two channels [16]. Universal Design for Learning builds on this idea through principles of multiple means of representation and engagement [17]; its benefits are reported but rarely tested rigorously [18]. Most of this literature is conceptual [19] or set in higher education [20], and it covers lessons [21] and teacher preparation [22] far more often than software. Text-to-speech support improves comprehension, but the gain depends on baseline decoding skill [23], which argues for making narration optional.

2.2 Embedded assessment as an interaction data problem

Embedded assessment gathers evidence continuously from ordinary interaction rather than through a separate instrument [12]. Once assessment is embedded, the interaction log becomes the evidence base. This works only if the data model stores individual attempts, each linked to an item, the response given, and the time taken, in addition to outcomes. Cano et al. [8] showed that telemetry such as task-completion time and error frequency validly indicates skill acquisition among users with intellectual disabilities, and game design taxonomies treat assessment as part of the design [7]. The assessment subsystem is thus a data management component whose schema limits later inference: a final score alone cannot support analyses of reattempts or response latency.

2.3 Research gaps

Three gaps remain. Reviews of game-based learning for students with intellectual disabilities find school subjects underrepresented and assessment integration weak [9, 10]. Published systems rarely describe architecture, data models, or technical evaluation, which makes them hard to replicate. Regulatory questions are rarely discussed, even though such systems log data from minors with disabilities, whom Indonesian Law No. 27 of 2022 on Personal Data Protection singles out for stronger protection. In schools, it is the teacher who operates the system, so teacher-facing requirements belong in the specification as well. Perceived usefulness and ease of use predict information system adoption [24], including teachers' classroom technology use [25], and Indonesian teachers' adoption of online assessment depends on facilitating conditions and performance expectancy [26]. These constructs informed only the requirements questionnaire and requirements N1, N2 and R8, not the assessment design.

4.1 Concept: Requirements specification

The teachers rated the need for interactive, web-delivered science media for this population highly (overall mean (M) = 3.89 of 4.00; 97.3%; Table 3). Agreement was near-unanimous on simple navigation (M = 4.00) and immediate feedback (M = 3.95), whereas audio support (M = 3.74) and embedded assessment (M = 3.79) received the lowest means. Audio instructions (94.7%), rewards (89.5%), and large targets (84.2%) were the most often prioritized features. Four themes emerged from the open-ended responses and FGD notes: abstract concepts are hard to grasp from text alone; the interface should be simple and consistent, with large icons and little text; visible rewards sustain effort; and assessment should sit inside the activity. Eight functional and four non-functional requirements were derived from these themes (Table 4). They describe the conditions reported by teachers at one school and are not assumed to characterize Indonesian special schools in general.

Table 3. Requirements questionnaire results (n = 19)

Table 4. Mapping of elicited needs to system requirements

4.2 Design: System architecture

Requirements N1 and N2 call for a client-centric web application, delivered over Hypertext Transfer Protocol Secure (HTTPS) and run entirely in the browser. Keeping presentation, navigation, assessment, and reward logic on the client removes per-device installation and avoids server contact during lessons, which suits the unreliable connectivity reported by the participating teachers. Figure 2 shows the four layers. The presentation layer renders the cards, quiz, progress view, and facilitator page, and holds the accessibility controls (Section 4.5); the application logic layer handles sequencing, answer evaluation, reattempts, star awards, and progress state; the data layer holds the content repository and interaction record (Section 4.3); and the integration layer reaches speech synthesis.

Presentation and application logic are written in React 18.2 and shipped as a pre-built static bundle. The integration layer calls the browser's Web Speech application programming interface (API) [39], which delegates synthesis to the speech engine available on the device (Section 4.6). The data layer uses browser-local storage, and apart from the static file host, there is no server-side component. The system does not use a service worker. Once the bundle has loaded, it remains operable without a network connection for the rest of an open session, but reloading the page requires connectivity. Table 5 lists the specification.

2.png

Figure 2. Layered architecture of the Bintang Sains system

Table 5. System specification

4.3 Design: Conceptual and data model

The data model separates authored content entities, which change rarely, from interaction entities written at runtime, which form the assessment evidence. This separation meets requirement R4 without tying the assessment record to the way content is presented. Figure 3 shows the six entities and their attributes; Learner carries only a pseudonymous session identifier (Section 4.6).

3.png

Figure 3. Entity relationship model of the content and interaction repositories

Table 6. Variables retained and not yet retained in the interaction record

In implementation, an Attempt record is keyed by attemptId (a version 4 UUID) and holds sessionId and itemId as foreign keys. The remaining attributes are selectedOption (string), isCorrect (boolean), attemptOrdinal (integer), and responseTimeMs (integer). Records last as long as browser storage does; they are not transmitted to a server. Since each response is stored separately instead of being collapsed into a topic score, the sequence of reattempts and the latency of each response are preserved. These variables are necessary but not sufficient for competency modelling. The schema does not record narration use, navigation history, interruptions, item difficulty, or teacher assistance, and the last of these proved important during testing (Section 5.2). As Attempt is an independent entity, these variables can be added as new attributes or as a separate Event entity without restructuring existing records (Table 6).

4.4 Design: Process model

Within a topic, the interaction follows a deterministic state machine (Figure 4). A wrong quiz answer produces corrective feedback and returns control to the same item (requirement R3). There is no failure terminal state, and every transition writes to the interaction record before the interface changes, so the stored evidence stays complete if a session ends unexpectedly. Unlimited reattempts do, however, complicate interpretation. A learner can retry until the answer is correct, so final correctness carries little information on its own, and the evidence has to come from first-attempt correctness, the number of attempts and response latency (Section 6).

4.png

Figure 4. State model of the learning and assessment interaction

4.5 Material collection and accessibility implementation

Content covers six IPA topics from the SMPLB curriculum: Living Organisms, Our Body, Ecosystem, The Solar System, States of Matter and Energy. Each topic has eight cards, each with a progress indicator, one short proposition in plain vocabulary with a concrete example, an illustration, and a control that reads the card aloud (Figure 5). Narration is generated from the displayed text, so the spoken and written channels cannot diverge through an authoring error.

Body text has a contrast ratio of 7.2:1 and interactive controls 5.1:1. An automated audit was run to check the interface against the Web Content Accessibility Guidelines (WCAG) 2.1 at Level AA [40]. Targets measure at least 64 × 64 px, font size has three levels (16, 20 and 24 px), and every illustration has plain-language alternative text. These checks establish conformance with general accessibility guidelines; they do not by themselves establish accessibility for learners with intellectual disabilities, which Section 5.2 examines through observation.

5.png

Figure 5. Implemented interface of Bintang Sains: (a) homepage with the three-step visual guide; (b) topic-selection screen with star progress indicators; (c) flashcard screen with the illustrated proposition and the Listen (Dengarkan) control

4.6 Security, privacy and regulatory compliance

Under Indonesian Law No. 27 of 2022, children's data require parental consent, and health-related information requires stronger protection. Accordingly, the system keeps only the data its assessment function needs. The Learner entity stores no names, dates of birth, class, school, or diagnosis, and records are linked to a session rather than a permanent account, so no long-term profile is created. The application bundle is delivered over HTTPS; learner data are not transmitted. Each record stays in the browser storage of the device on which it was created. For the research, the class teacher manually copied a summary of each student's assessment activity from the test device at the end of the session, without identifying information; the research team accessed only these summaries.

Local-only storage also has costs. It rules out longitudinal assessment across sessions, loses records when storage is cleared or a device fails, prevents continuity across devices, and gives teachers no access to interaction histories beyond what they copy by hand. A future version could add a teacher-controlled, encrypted export or synchronization with a school-local store under the same consent and minimization rules.

Speech synthesis raises a less obvious issue. If narration is produced by an online service, card text is sent with each request; the text is not learner data, but request patterns could reveal which topics were studied and when. With an on-device id-ID voice, card text does not leave the device, but some browser–voice combinations route synthesis through the vendor's online service. Bintang Sains therefore cannot guarantee by itself that no text reaches a third party; deployments should prefer voices reported as local (localService = true), and schools should be informed when only an online voice is available.

The requirements came from teachers at a single school and describe that school's conditions. The student evaluations at two other schools show that the design was usable there, but not that the requirements generalize. Architecturally, the design rests on three decisions. Browser delivery removes the installation and maintenance burden identified as the main obstacle in resource-limited schools [3], although offline operation is limited to an open session. Generating narration at runtime from the card text keeps speech and text consistent (Section 4.5). Storing each response as a separate attempt preserves the reattempt sequence and latency on which embedded assessment depends [12], although more variables will be needed before a competency model can be fitted (Table 6).

The Aiken's V values of both expert panels exceeded the tabulated criterion [35] and the conventional benchmark [36]. These indices show agreement on the relevance and suitability of the designed components; they do not show that embedded assessment improves learning or that the attempt data support valid inferences about learners.

With unlimited reattempts, final correctness carries little information on its own, so interpretation has to rest on other indicators. In this session, 299 of 300 responses were correct on the first attempt. Because each topic's quiz follows immediately after its cards, this most likely reflects immediate recall rather than mastery or retention, and the quiz did not differentiate between learners. Future analyses could distinguish four patterns: mastery, indicated by first-attempt correctness that persists on delayed items, such as items repeated in a later session or mixed across topics; guessing, indicated by very short latencies followed by errors; trial-and-error, indicated by attempt counts approaching the number of options; and assisted completion, which the current schema cannot detect. The five students prompted while answering left records identical to independent ones, which is why Table 6 includes an assistance flag.

The positive student ratings and the high rate of independent completion require cautious interpretation. The ratings are close to ceiling, which limits their diagnostic value. Self-report by people with intellectual disability is prone to acquiescence and social desirability, especially when adults are present [34], and the pictorial scale reduces but does not remove this bias. Novelty may also have raised engagement, as reward mechanics can lose effect over time [6]. The evaluation also consisted of one 25-minute session with 20 students, and five needed repeated prompts during the quiz. The evidence therefore supports a narrower claim, namely that the system was usable and well received on first contact by most participants. It does not address durable engagement or learning.

Data storage, retention and third-party exposure are largely architectural decisions, although Section 4.6 shows that speech synthesis also depends on the deployment environment. The reusable contribution is a design pattern made up of content held as data, narration generated from the displayed text, attempt-level assessment records, browser delivery without installation, and a learner entity that holds nothing identifying. Whether this pattern supports learning in under-resourced schools remains to be tested in classroom studies across more schools.

The study was approved by the Research Ethics Committee of LPPM, Universitas Negeri Malang (No. 06.08.02/UN32.14.2.8/LT/2026, valid from August 6, 2026 to August 6, 2027). Parents or guardians provided consent through the schools, including permission to publish the photographs in Figure 6.

The authors used Claude for language editing, translation, formatting, and diagram preparation. All content was reviewed and verified by the authors. Generative AI was not used to generate, analyze, or interpret data.

This work is supported by the RKAT of Universitas Sebelas Maret, Fiscal Year 2026, through the Fundamental Research scheme A (PFA-UNS) (Research Assignment Agreement No. 460/UN27.22/PT.01.03/2026). The authors thank the teachers of SLB YRTRW Surakarta, the two partner special schools in Central Java, the eight media and content experts, and the class teachers who accompanied the students. Above all, we thank the students who used Bintang Sains and shared their responses with us.