Scientific literacy in primary education rests on a skill that is simple to describe but hard to teach: reasoning from evidence. Children must distinguish observation from interpretation, make testable claims, and select data that support them. The most demanding step is explaining why that data supports the claim through a scientific mechanism. In primary classrooms, pupils often reach the conclusion but struggle with the explicit mechanistic link between evidence and claim ([9]; [19]).

Physical computing eases this challenge. Sensors let children build systems that interact with the physical world and return continuous real-time measurements, turning abstract phenomena into observable data ([25]; [28]; [7]). Reliable data, however, do not produce interpretation on their own: the critical step is moving from recording a value to reading a pattern as evidence of a mechanism, and without support, pupils find this hard ([16]). GenAI promises personalized support, timely feedback, and new forms of assessment ([42]), but used as an answer provider it performs the reasoning the learner should do, encouraging over-trust and weakening independent thinking ([8]; [15]). The question for primary science is therefore not whether GenAI will be present, but which instructional design can make it support reasoning rather than replace it.

Research on the micro:bit has focused mainly on technical features, motivational benefits, and approaches to teaching programming, and less on whether such activities support scientific thinking or conceptual understanding ([30]; [12]). Completing a physical-computing task does not demonstrate understanding of the phenomenon, the variables, or the link between measurement and scientific model: a device that works is not a pupil who thinks. Similarly, many educational chatbots ask generic questions and give feedback without requiring pupils to draw conclusions from their own data, and without criteria for when to advance, persist, hint, or withdraw.

Two gaps emerge. Physical computing generates rich data but does not ensure scientific reasoning; GenAI generates rich dialogue but can easily detach from the learner’s data and substitute for the learner’s own thinking. This study treats the two gaps as complementary and addressable through a single intervention, provided the dialogue stays tied to pupils’ own measurements and follows explicit epistemic and developmental rules.

Dialogic teaching views learning as a process in which ideas are shared, examined, and progressively refined through mutual interaction ([1]; [23]; [35]). Unlike static prompts, generative artificial intelligence (GenAI) dialogue systems can adapt to how learners are thinking, which makes them a promising support tool when they advance inquiry through questions rather than immediate answers ([14]; [31]; [32]; [33]). LLM-based platforms have also been developed to help educators design computational-thinking, AI, and STEM activities, illustrating the wider potential of generative systems as pedagogical infrastructures ([40]).

Customized chatbots can support reasoning and argumentation in secondary science ([33]), and recent theoretical work casts GenAI as a conversational partner rather than a source of information ([37]). Neither specifies an implementation for primary-aged children: a developmentally appropriate method grounded in children’s own sensor data, with rules for when the system should question, prompt, or support, and when that support should be withdrawn. Scaffolding theory supplies the missing element. A scaffold is the support that enables a learner to solve a problem beyond their current ability ([38]), and its defining features are contingency, fading, and the transfer of responsibility to the learner ([34]); permanent mediation is not effective support, even when the immediate answer is correct. Work on scientific explanation stresses reducing written aids as skill develops ([17]; [19]), and studies of AI-supported scaffolding emphasize continuous diagnosis and progressive withdrawal rather than answer provision ([3]; [5]; [36]).

Three further findings clarify the developmental stakes. Young children may place considerable and largely uncritical trust in conversational agents ([22]). AI literacy must be deliberately cultivated rather than assumed to emerge from technology use ([24]). And chatbot effects are, meta-analytically, smaller for younger learners, with novelty-related gains that may fade ([39]). Syntheses of GenAI-supported learning agree: the technology’s educational potential cannot be separated from its effects on cognition, metacognition, and learner agency ([42]).

Consequently, general-purpose chatbot designs cannot simply be transferred to primary education; they require developmentally appropriate constraints, scaffolding, and safeguards.

The closest existing demonstration, a customized chatbot supporting reasoning and argumentation in secondary science ([33]), leaves three issues unaddressed for primary education: dialogue grounded in pupils’ own sensor data rather than curricular text, explicit criteria for fading support, and safeguards for children who may over-trust conversational agents. Effectiveness is defined here not by whether a pupil completes a conversation but by whether the pupil can subsequently produce an independent Claim-Evidence-Reasoning (CER) explanation ([18]) without AI. The aim is to develop the AI-DRS (AI-supported Dialogic Reasoning Scaffolding) Physical Inquiry Model and prepare it for validation, treating the design itself as the research contribution.

The study is guided by a single design question: which design principles and architecture can connect GenAI dialogue to primary pupils’ own sensor data so that it supports, and then deliberately fades, independent evidence-based reasoning? The evaluative questions that follow concern whether the model improves independent CER performance relative to an equivalent static scaffold, how it affects pupils’ use of their own data and the trajectory of scaffolding, whether pupils articulate synthesis themselves, whether reasoning transfers without AI, and how perceived usefulness and calibrated trust relate to reasoning quality. These form the protocol of the planned pilot and are not addressed empirically here.

The study uses a Research and Development (R&D) design with two sequential parts: (a) developing the intervention model and (b) planning its expert validation and subsequent pilot evaluation. This article reports only the first part; the second is set out as a protocol for future work. The study is deliberately not presented as an efficacy trial: the planned pilot examines feasibility, implementation fidelity, preliminary indications of effect, and above all, the links among scaffolding, data use, and reasoning.

Development is organized into seven R&D stages:

1) needs analysis through curriculum documents, research literature, and the authors’ teaching experience, with no participant data;

2) mapping of curriculum, target concepts, and likely misconceptions;

3) development of the inquiry task, the AI-DRS prompt, the worksheet, and the teacher guide;

4) expert validation;

5) a small-group technical and usability test;

6) revision of the package; and

7) pilot implementation with process evaluation.

The first three stages are complete, and their output, the model in Section 3, constitutes this article’s contribution; the remainder are reported as protocol, and Sections 2.2 - 2.4 set it out in full for replicability, describing planned rather than collected data. Because the intervention involves GenAI use with children, it assumes a school-managed environment, data minimization, parental consent and pupil assent, pseudonymized chat logs, defined retention rules, and continuous teacher oversight. The system states that it can be wrong, asks pupils to check its suggestions against their own observations, and routes uncertain or problematic cases to the teacher. The planned pilot will be submitted to the Research Ethics Committee of the University of Thessaly before any data collection ([27]).

The design must also address how children may lawfully interact with GenAI. Mainstream consumer chatbots are not permitted for this age group: OpenAI’s terms require users to be at least 13, with parental consent for minors ([26]), and Anthropic’s require users to be at least 18 ([2]). The model, therefore, assumes mediated access. Pupils hold no accounts and use no consumer interface; the dialogue runs through a school-controlled application operating under the teacher’s institutional account and the selected provider’s application programming interface (API). That application logs pseudonymized interaction data inside the school environment and transmits only data-minimized, non-identifying inputs through the API, under approved data-processing, security, and retention arrangements.

A sample of about 40 to 60 Grade 5 or Grade 6 pupils is proposed, with 20 to 30 per condition. Because pupils build and measure in groups of three to four and groups are the unit of assignment, this corresponds to approximately five to ten groups per condition, a figure that governs the analysis plan in Section 2.4. Pupils work in groups during construction, but the pre-test, post-test, and transfer measures are administered individually. The planned data sources are the individual pre- and post-CER tasks, an individual transfer task completed without AI, pseudonymized chat logs, teacher observation logs, technical and fidelity records, a short perceived-usefulness and calibrated-trust questionnaire, and the expert-validation ratings collected before the pilot.

The intervention is built on the Nezha Inventor’s Kit for micro:bit (Figure 1). Because the assessed comparison requires both soil conditions to be set up and measured simultaneously within each group (Section 2.2.3), every group is equipped with two complete measurement channels, one per container: two BBC micro:bit units, two soil-moisture probes, and LED indicators. This exceeds the standard kit, which contains a single soil-moisture sensor, so a second kit, or a second micro:bit and probe, is required per group; this is stated explicitly as a resource condition of the design rather than assumed.

The materials support a Smart Plant Investigation in which pupils build a soil-moisture monitoring system and examine how soil type affects the change in soil moisture over time. Every group investigates the same focal comparison under the same controlled conditions, so that datasets and reasoning demands are equivalent across groups and conditions. After calibration and a stability check on each channel, pupils record the container, elapsed time, calibrated readings, quality flags, and their observations.

Figure 1.Components of the Nezha Inventor’s Kit used in the proposed Smart Plant Investigation. Source: Authors’ photograph.

The soil-moisture probe is the kit’s own sensor, which keeps the build within the kit’s wiring conventions and MakeCode blocks; components such as a BME280 sensor or an OLED display are not part of the standard configuration ([10]). Duplicating the micro:bit and the probe, as described above, is the only departure from the standard configuration.

2.2.1. Inquiry Question, Target Mechanism, and Permissible Causal Claim

Claim-Evidence-Reasoning (CER) responses can be scored consistently only if the investigation has one question, one expected mechanism, and one clearly bounded causal claim that applies to every group. The Smart Plant Investigation is therefore built around a single focal comparison rather than group-selected variables. The operational inquiry question is: how does soil type (sandy soil versus potting compost) affect the rate at which soil-moisture readings decrease over 40 minutes, when the initial water volume, container, soil mass, probe placement, bench position, light exposure, and room temperature are held constant?

The expected empirical pattern is a steeper decline of the calibrated Moisture Index in sandy soil than in compost. The target mechanism, expressed at a level appropriate for Grade 5-6, is that sandy soil consists of larger particles with larger pore spaces and little organic matter, so water drains through it faster and is exposed to the air across a larger internal surface, whereas the finer particles and organic matter in compost hold water against drainage and evaporation. A full-credit reasoning statement links a stated numerical difference in the rate of decline to this particle-size and pore-space account; a partial-credit statement identifies the difference but restates the data instead of explaining it, as in “it dried faster because the numbers went down faster”.

The permissible causal claim is deliberately bounded. Because soil type was manipulated while the remaining conditions were held constant, pupils may claim that, in this investigation and within the 40-minute window, the type of soil caused the difference in the rate of moisture loss. They may not claim which soil is better for growing plants, extend the claim to plant health or growth, read the index as an absolute quantity of water, or make causal claims about light, temperature, or container size. Table 1 states this target explicitly so that raters, the teacher, and the AI-DRS system apply the same scientific standard: it is supplied to the system as a fixed reference in every dialogue turn (Section 3.4) and it governs the Mechanism dimension of the rubric (Section 3.7).

Table 1. Scientific target of the smart plant investigation.

2.2.2. Sensor Calibration, Measurement Schedule, and Data-Quality Rules

The soil-moisture sensor returns an uncalibrated analog value, so the same physical condition can produce different readings across sensors, sessions, and probe placements. Interpretation is therefore preceded by a fixed two-point calibration and a stability check, and the dataset is screened with rules defined in advance rather than negotiated during the dialogue.

Calibration is carried out for each of a group’s two probes at the start of every session and repeated whenever a probe is removed and re-inserted: a dry reference in air and a wet reference at the marked depth in tap water, each the mean of five readings after a settling period (Table 2). Readings are converted in MakeCode to a Moisture Index, MI = 100 × (R − R_dry)/(R_wet − R_dry). The index is unbounded: the unclipped value is logged and screened for quality, and only the plotted value is clipped to −5 to 105, so that an out-of-range reading remains detectable by the range rule. Each channel yields its own pair of calibration constants; both pairs are recorded on the group worksheet and stored with the dataset, labeled by container.

The acceptance criteria and the action taken when they are not met are set out in Table 2: a reference is accepted only when the five readings span no more than 2% of full scale, and a probe that fails three attempts, or whose calibration span falls below 150 raw units, is replaced before the investigation proceeds. After the standardised watering at t = 0, MI is recorded on both channels every 5 minutes for 40 minutes, giving nine time points per container; each value is the median of three consecutive readings taken 5 seconds apart, which suppresses single-sample noise without requiring statistical treatment by pupils. Probes remain at a fixed depth of 5 cm and at least 2 cm from the container wall, marked with tape; any displacement is recorded as an event.

The investigation is timetabled across three 45-minute lessons, set out in the final row of Table 2. Lesson 2 is the tightest, since a 40-minute window leaves no margin in a 45-minute slot; it is therefore timetabled as a double period where the school allows one, and otherwise calibration is completed at the end of lesson 1 so that lesson 2 begins with watering at t = 0. If neither is workable, the schedule is shortened to 30 minutes and seven time points, but only as a study-level decision taken for the whole pilot before data collection, never per group, so that all groups produce datasets of identical structure; the adopted schedule and its dependent thresholds are then reported. The 40-minute, nine-point schedule is the reference version used throughout this article. Lesson 2 deliberately does not carry the dialogue: about four minutes between readings allows graphing and the claim and data-grounding moves but not the full contingent exchange, so conceptual bridging, fading, and transfer occur in lesson 3 from a closed dataset.

Table 2. Calibration and measurement procedure.

Five rules identify unreliable readings before any interpretation begins (Table 3). A single flagged point is re-taken once and, if the flag recurs, is excluded from the evidence base while remaining in the log. A run that fails the drift rule, or a container with more than two flagged points out of nine, is not interpreted at all: the system halts interpretation and refers the group to the teacher, which is the operational trigger for rule R8 in Table 8. The group then completes the CER task using a validated class reference dataset, so that a hardware fault does not cost the pupils the reasoning task. These definitions give the system’s unreliable-data category a concrete referent instead of leaving it to real-time judgment by a generative model.

Table 3. Rules for identifying unreliable readings.

The class reference dataset is not improvised during the lesson. It is collected by the research team in a pre-pilot, with at least six complete runs per soil condition, using the same kit, soils, containers, calibration procedure, and 40-minute schedule as the pupils, and screened against the same five rules. Its pooled runs fix the per-condition jump thresholds used in Table 3, provide the substitute dataset for a run that cannot be interpreted, and establish that the two soils are distinguishable within the measurement window. Because the substitute data follow the same protocol, a group working from them faces the same CER task; every substitution is recorded as a fidelity deviation and reported with the results.

2.2.3. Assignment and Balancing of Inquiry Variables

Allowing each group to select its own variable would produce datasets of different shapes and reasoning tasks of different difficulty, confounding any comparison of CER performance between conditions. Here the focal variable is identical for all groups, and both of its levels are investigated within each group: every group sets up one sandy-soil container and one compost container and measures them simultaneously on its two channels. Every group therefore produces a dataset of the same structure, two conditions by nine time points, and faces the same reasoning demand. Light exposure, initial water amount, and container surface area are retained only as optional extensions after the assessed CER task, recorded as process variables and never used in a between-condition comparison.

Assignment proceeds in three steps. First, pupils are ranked within class on the pre-test CER total and allocated to mixed-attainment groups of three to four, so that group mean prior attainment is comparable. Second, groups rather than pupils are randomly assigned to the AI-DRS or the static-scaffold condition, using block randomization with class as the blocking factor and a computer-generated sequence prepared by a researcher not involved in delivery; allocation is concealed until the pre-test has been scored.

The group is thus both the unit of assignment and of delivery, which is the basis of the analysis in Section 2.4. Third, equipment is counterbalanced: probes, micro:bit units, and bench positions are rotated across conditions and across the two soils within each group, so that no device or location is systematically associated with one condition or one soil.

Standardization of the materials is specified in Table 4. Balance is verified before analysis by reporting, separately by condition, the distribution of pre-test scores, soil batches, sensor units, bench positions, ambient temperature and relative humidity, and prior micro:bit and GenAI experience. Any imbalance is reported descriptively rather than adjusted away: with only five to ten groups per condition, randomization alone cannot be relied upon to produce equivalence, and the comparisons are interpreted accordingly.

Table 4. Standardization and assignment of the inquiry variables.

2.2.4. Comparison Conditions

The two conditions are identical in every respect except the form of analysis support: the experimental groups receive the adaptive AI-DRS dialogue, the active control groups an equivalent static worksheet with the same questions in a fixed order. Hardware, phenomenon, dataset, time on task, and the individual final assessment are the same in both (Table 5). One source of contamination cannot be removed at this scale. Because both conditions run in the same room within the same two-hour window (Table 4), control groups can see and overhear the AI-DRS groups, and any diffusion of dialogic moves would attenuate the observed difference. Separating the conditions by room or session would confound condition with setting, so cross-condition exposure is instead logged and reported as a limitation.

Table 5. Comparison conditions.

Data will be collected across the seven-phase learning sequence detailed in Section 3. The system captures pupils’ interactions as pseudonymized chat logs, their measurements as datasets and graphs, and their explanations as written CER responses, while the teacher completes a structured observation log throughout. Equivalence and fidelity indicators are recorded: prior science knowledge, baseline CER performance, prior micro:bit and GenAI experience, time on task, absences, technical problems, teacher interventions, and number of AI turns. A subset of lessons is reviewed against a fidelity checklist by a second observer.

In the planned pilot, two independent raters will score at least 25% of the responses. Weighted Cohen’s kappa will be computed for the ordinal rubric dimensions and an intraclass correlation coefficient for the total CER score, with disagreements resolved through documented consensus. Chat logs will be examined using a scheme that classifies the system’s moves as Socratic questioning, data grounding, conceptual bridging, guided choice, explicit correction, or metacognitive prompting, and pupil responses as full, partial, none, or not reached. Synthesis is recognized only when the pupil articulates it; accepting an AI-generated summary does not count as independent reasoning.

The analysis plan follows from the unit of assignment. Because groups of three to four pupils are randomized and taught as a unit while outcomes are measured individually, pupils are nested within groups, and the planned sample corresponds to only five to ten groups per condition. An individual-level ANCOVA treating pupils as independent observations is therefore not an appropriate primary analysis, where the intraclass correlation is positive, standard errors are understated, and the associated p-values are not interpretable, and reporting an intraclass correlation alongside an uncorrected model does not remedy this ([11]). Multilevel modeling is not a workable alternative at this scale either, because variance components estimated from fewer than roughly ten clusters per arm are unstable and their standard errors are biased downwards ([20]).

The primary analysis is therefore specified at cluster level and is descriptive rather than confirmatory. For each group, the mean post-test and mean pre-test CER score are computed, and the between-condition contrast is the difference in group mean post-test scores adjusted for group mean pre-test scores. Every group mean is displayed individually, so that the small number of clusters is visible rather than hidden behind an aggregate. The effect size is a cluster-adjusted standardized mean difference using the total between-plus-within variance in its denominator, reported with the intraclass correlation on which it depends ([11]). The primary inferential procedure is an exact permutation test on the groups’ condition labels, because with five to ten clusters per arm it is the only exact procedure available. At the lower end of that range, five groups per arm, the test cannot return a two-sided p-value below about .008 and has negligible power against any plausible effect; it is therefore reported as a descriptive statement of how extreme the observed contrast is among the possible label assignments, not as a test of efficacy. Interval estimates use a wild cluster bootstrap-t, developed precisely because pairs-resampling and percentile-type bootstraps behave poorly with few clusters; a bias-corrected and accelerated bootstrap resampling of whole groups is deliberately not used, for the same reason ([6]). Both procedures are conservative at this number of clusters and the resulting intervals will be wide, which is accepted as the cost of correct inference rather than concealed behind a nominally narrower pupil-level interval.

Pupil-level ANCOVA is retained only as a secondary, illustrative analysis, with cluster-robust (CR2) standard errors and Satterthwaite degrees of freedom, designed for this situation of few clusters ([4]); no pupil-level p-value is treated as evidence of effectiveness. The intraclass correlation is reported as an interval, not a point estimate, since with at most ten clusters per arm its confidence interval is necessarily wide, and it is accompanied by a sensitivity table giving the design effect, Deff = 1 + (m − 1) × ICC, and the effective sample size, N/Deff, across a plausible range: with groups of three to four, an intraclass correlation of .05, .10, or .20 reduces an effective sample of 60 pupils to about 53, 48, or 40. This bounds the number of clusters a subsequent cluster-randomized trial would require, where the number of groups rather than of pupils determines power; it is a planning range, not an estimate a pilot of this size could deliver. The analysis plan, including the stopping rules for the process measures, will be pre-registered. Table 6 summarizes the unit of analysis, estimand, and inference for each outcome.

Table 6. Analysis plan: outcomes, units of analysis, estimands, and inference.

The qualitative analysis will focus on patterns of CER error and will use thematic analysis for explanations, integrating turn-level chat-log coding with comparative cases of full, partial, and zero uptake in joint displays that connect CER gain, data-grounding frequency, bridge uptake, engagement, and transfer.

Sensors convert physical states into numerical values, but their scientific meaning comes from the relations among values, conditions, and theoretical ideas. Pupils must decide which data are relevant, whether a measurement is trustworthy, which patterns are present, and what mechanism might explain them. The model foregrounds four practices: generating data, checking data quality, selecting evidence, and explaining mechanistically. It organizes them through a CER progression ([18]) adapted for the primary years: Observe/Describe, Predict, Claim, Evidence, Reasoning, and Reflect and Transfer.

At the heart of the model is a dialogic control layer that responds to the pupil’s input. The system advances the pupil’s thinking through targeted questions instead of presenting a ready-made explanation: the interaction starts with observation, asks for comparisons, requires a claim, returns the pupil to the actual measurements, and only then prompts for mechanistic reasoning. Its value depends on the quality of the pupil’s responses, not the number of exchanges. Three principles govern the interaction. Data grounding requires pupils to cite specific values, changes, or patterns from their own data and rejects vague statements such as “the soil dried out more”. Epistemic gatekeeping makes the move from claim to evidence conditional on a testable claim, the move from evidence to reasoning conditional on relevant data, and completion conditional on a mechanism stated in the pupil’s own words.

Summaries produced by the system are never counted as the pupil’s own synthesis. Contingent responsiveness determines the next move from the quality of the response, classified as adequate and independent, partially adequate, off-target, or scientifically inaccurate, for instructional routing only. Figure 2 shows these principles as a layer beneath the reasoning progression, with AI support decreasing as reasoning becomes more independent.

Figure 2. The AI-DRS reasoning progression and its dialogic control layer.

The intervention is enacted through seven phases, each pairing a pupil activity with a defined role for the AI-DRS system or the teacher and yielding a distinct piece of evidence for analysis (Table 7). The system does not answer during orientation, offers no correction during prediction, and enforces fair-test criteria before construction begins.

Table 7. The seven phases of the AI-DRS Physical Inquiry Model.

The model’s adaptivity is defined by an explicit rule set linking the quality of a pupil’s response to the system’s next move, so that support remains just sufficient to keep the reasoning on track (Table 8).

A general statement without data triggers a data-grounding request; data without a mechanism triggers a conceptual bridge; a scientifically inaccurate response triggers explicit correction and a teacher notification. Rule R9 sits outside the instructional sequence: if the exchange touches on a pupil’s welfare, causes distress, or moves off task, the dialogue stops and the teacher is alerted.

Table 8.Response-contingency rules.

The architecture rests on a technical assumption. Although the contingency rules are explicitly defined, they are implemented by a stochastic generative model, so the real-time classification of pupils’ responses may not always be reliable.

Misrouting is possible: an adequate answer may attract an unnecessary prompt, or an unsafe one may pass. Three design decisions bound the consequences. The classification is used only for low-stakes instructional routing, never for assessment; the default action under uncertainty is the mildest one, a reformulation, with a teacher flag raised if it recurs; and every interaction is logged, so that routing accuracy can be measured and becomes part of the validation criteria in Section 3.9.

Dialogic moves are delivered as short, age-appropriate prompts, for example “Which two or three readings from your own table best support that claim?” for data grounding, “How might soil structure, drainage, or evaporation explain the different pattern?” for the conceptual bridge, and “Check my suggestion against your graph. What evidence would show that it is wrong?” for calibrated trust. Appendix A gives the full register and move specifications.

A dialogic model implemented on a generative system is reproducible only if the configuration that produced the dialogue is stated. The reference implementation is therefore specified at the level of provider, model snapshot, prompt version, decoding settings, input format, and rule precedence.

Pupils hold no accounts and use no consumer interface. A school-controlled web application delivers the dialogue; a group signs in with a class code and a pseudonymous group label, and the application calls the provider’s application programming interface server-side, under the institution’s account and key. The reference configuration uses a single pinned, dated model snapshot, logged with every turn. The snapshot named here, the OpenAI API with gpt-4o-2024-11-20, fixes the reference build rather than prescribing the model to be used; the snapshot in force at the pilot is reported verbatim, and any change triggers a re-run of the validation suite before use with pupils.

The system prompt is a fixed, version-controlled artefact (AI-DRS-SP v1.0) rather than an ad hoc instruction, and is reproduced verbatim in Appendix A. Its twelve blocks specify role and audience; the register constraint of one question per turn, at most 45 words, and no technical vocabulary not yet introduced in class; the never-answer rule, under which the system does not state the claim, the evidence, or the mechanism on the pupil’s behalf and does not confirm an unjustified claim; data grounding; epistemic gatekeeping; the target-explanation reference from Table 1; the contingency rules of Table 8 and their precedence; the fading levels; the calibrated-trust statement; the refusal and safeguarding rules; and the output schema. The four response categories named in Section 3.1 are the coarse form of the eight operational classification values defined in Block 6 of Appendix A, which separate the unsupported, unreliable, and safeguarding cases that the four categories collapse together.

Decoding settings, output schema, failure fallback, and the exact input payload are specified in Table 9. Three of these are load-bearing. The 160-token limit keeps replies short enough for the target age group. A generation that fails schema validation is never shown to pupils as free text, but is retried once and then replaced by the deterministic reformulation move with a teacher flag. And nothing is transmitted beyond the group’s own calibrated dataset, the pupil’s current response, the phase and fade level, the two preceding turns, and the fixed target explanation, with class-list names removed by a client-side filter.

Table 9.Reference implementation specification of the AI-DRS system.

Because more than one contingency rule can match a single response, the rules are ordered, and the system applies only the highest-priority matching rule. Table 10 sets out that hierarchy and maps every priority level onto the corresponding rule identifier in Table 8. Safeguarding overrides everything; a data-quality halt precedes any interpretive move, since interpreting unreliable measurements is worse than not interpreting them at all; scientific correction precedes gatekeeping; gatekeeping precedes data grounding and conceptual bridging; and fading is applied last, only when no higher rule has fired. When classification confidence is low, or two rules of equal priority match, the system takes the mildest available action, a reformulation, and raises a teacher flag if the difficulty recurs, so that the classifier’s failure mode is conservative by design.

Table 10.Rule precedence in the AI-DRS dialogue.

Flagging is useful only if it is attached to a named person, a defined response time, and a defined endpoint. The class teacher is the first responder and is present throughout; a live dashboard shows one row per group with the current phase, the fade level, and any open flag. A designated science lead, or a member of the research team, is the second reviewer and examines every flagged interaction within 24 hours, recording the outcome in a review log. Where a flag concerns a child’s welfare rather than the science, the school’s designated safeguarding lead takes over under the school’s existing child-protection procedure.

Flags are graded (Table 11). A red flag pauses the dialogue automatically and alerts the teacher, who is expected to reach the group within two minutes; triggers include a safeguarding disclosure, signs of distress, content the filters should have blocked, explicit correction of a scientifically unsafe misconception, and any turn in which the system appears to have performed the pupils’ reasoning. An amber flag does not pause the dialogue but requires a teacher visit within the same lesson. Routine events such as a change of fade level are logged as green entries and reviewed after the lesson.

The corrective sequence is fixed: the dialogue pauses; the teacher reads the last three turns on the dashboard, speaks with the group, and establishes whether the difficulty is scientific, technical, or affective; the teacher then resumes with corrected input, transfers the group to the static worksheet scaffold for the remainder of the session, or ends that group’s AI dialogue. The second and third outcomes are recorded as fidelity deviations.

The dialogue must stop, and may not resume within that session, if the system produces scientifically incorrect content the teacher cannot correct on the spot; if it supplies the claim, evidence, or mechanism the pupils were asked to produce; if a pupil discloses safeguarding-relevant information or shows distress; if the system is used off task after one warning; if two red flags occur in one group within a session; or if a data-quality halt cannot be resolved. The group then completes the sequence with the static scaffold, so that no pupil loses the science lesson because of a system failure. The incident is logged and reported to the research team within 24 hours and, where reportable, to the ethics committee. Recurrent prompt failures identified in the weekly review trigger a prompt revision, a version increment, and a re-run of the regression suite.

Table 11. Teacher escalation protocol.

The presence of a scaffold does not guarantee uptake: a conceptual bridge may lead to full articulation, simple acceptance, or no response at all. The model therefore separates the system’s moves from the quality of the pupil’s uptake and treats engagement as a function of the substance of the response, data use, revision, and persistence rather than the number of turns. Fading occurs when the pupil performs the required practice independently and proceeds through four levels: a direct prompt naming the step, a delayed prompt issued only after a turn’s pause, a minimal prompt that acknowledges the response without asking anything further, and no prompt at all. Support is reduced by one level at a time and may return temporarily if performance declines. Block 9 of Appendix A states the corresponding fade_level values.

Two instruments make reasoning visible. The CER rubric (Table 12) scores six dimensions to a maximum of twelve points. Its Mechanism dimension is scored against the target mechanism of Table 1 rather than a general standard: a score of 2, partial mechanism, requires a correct but incomplete appeal to particle size, pore space, or organic matter, for example naming faster drainage without relating it to soil structure, whereas a score of 3, full and consistent, requires the pupil to link a stated numerical difference in the rate of decline to the particle-size and pore-space account of Table 1 while making no claim outside the bounds listed there. Raters use the same Table 1 reference that is supplied to the AI-DRS system, so that the human and machine standards coincide. The chat-log coding scheme classifies the system’s moves and the pupil’s responses so that only pupil-expressed synthesis counts as independent reasoning.

Table 12.Claim-Evidence-Reasoning (CER) rubric.

The model is designed so that technology is not treated separately from the science content and the pedagogical goals it supports ([21]). This view is consistent with computational-pedagogy frameworks for STEAM education, which emphasize the coordinated interaction of content, pedagogical practice, computational methods, and technological tools ([29]), and with work integrating AI-based assessment models and authentic STEAM engineering tasks in contexts such as precision agriculture ([40]). Each phase involves a specific combination of content, pedagogical, and technological knowledge (Table 13); the analysis and explanation phases represent the fullest TPACK intersection, where data logging and the AI-DRS dialogue work with CER scaffolding to promote mechanistic reasoning.

Table 13.TPACK alignment of the intervention phases.

Finally, the development strand sets out the criteria that must be met before the pilot can proceed. A panel of five to seven experts in science education, primary teaching methods, educational technology, and research ethics will rate each element for relevance, clarity, feasibility, developmental appropriateness, scientific accuracy, and safeguarding, and will record item- and scale-level content-validity indices (I-CVI ≥ .78 and S-CVI/Ave ≥ .90).

The minimum requirements are: no unresolved objection to scientific accuracy; a successful review of the prompts against misconception, privacy, and refusal cases; routing accuracy of at least 80% in the small-group usability test, measured as the percentage of authentic pupil responses that the system categorizes as an independent human coder does, with no category of Table 8 below 70% and with correct routing on every safeguarding (R9) and scientific-accuracy (R7) item, since an aggregate threshold can be satisfied while the two safety-critical rules fail; the pre-session regression suite of Table 9is scored against the same per-category floors but is a separate and weaker check, because its items are curated and easier than authentic responses, and meeting it does not discharge the usability criterion; stable hardware operation and logging in at least 90% of trials; pupil comprehension of instructions without extended explanation; teacher feasibility within the timetabled slots; a documented dry run of the escalation protocol including at least one simulated red flag; and a pilot scoring agreement of κ ≥ .70.