The question+
A lecture hall generates an enormous amount of information every minute and throws essentially all of it away. This paper sets out the instrument Diaposi already is: eighteen slide-level signals captured in production sessions without biometrics and without student accounts, separated by whether the student chose to send them, whether they are derived from navigation performed for other reasons, or whether they never leave the device at all. We state the measurement and the intervention questions separately, give the six rules the instrument is built to, map the signals onto all four modes of the ICAP framework1, set out the position the EU AI Act leaves open11, and point to the seven studies the open programme puts to partners.
Keywords learning analytics · cognitive engagement · audience response systems · ICAP · EU AI Act · classroom instrumentation
The question
Teaching remains one of the few skilled professions practised almost entirely without instrumentation.
Whether a concept landed, who fell behind and on which slide, which question half the room was privately holding: all of it evaporates the moment the session ends.
Diaposi exists to capture that information. The research question is whether what it captures is actually worth anything. It has two halves, deliberately. The first is a measurement question, answerable with correlational work against an outcome nobody chose after the fact. The second is a causal question about intervention, and it is not answerable without a trial.
Can volitional, slide-level interaction signals, captured without biometrics and without student accounts, serve as a valid real-time proxy for cognitive engagement? And does exposing those signals to the teacher mid-lecture change what students actually learn?
(1)Both halves are open. We are looking for partners on each.
The first half needs one faculty, one semester and access to assessment outcomes. The second needs sections randomised to a teacher who sees the live pulse and a teacher who does not. The open programme sets out both as scoped studies.
A layer, not a replacement
Most sentences written about measuring classrooms are really sentences about measuring the people in them.
Research pages tend to make teachers nervous, and with good reason. So before the tables and the study designs, the part that matters most to the two groups who have to live with this.
Diaposi does not ask anyone to teach differently. You upload the deck you already made: the same PDF, the same order, the same board work, the same voice, the same pace. It adds a channel back from the room. That is the entire intervention. Nobody is required to gamify anything, adopt a new pedagogy, or restructure a course that has worked for twenty years.
If you have taught the same way for twenty years
- T1Your existing slides work as they are. Nothing is rebuilt inside a tool.
- T2You are never scored, ranked, or compared to a colleague. No number about you leaves your session.
- T3You can ignore every signal on screen and the lecture runs exactly as it always did.
- T4It is not attendance software. It cannot tell you who is in the room, because it does not know who anyone is.
- T5Thirty seconds from upload to a live session. If it costs you a lesson plan, it has already failed.
If you would rather not be watched
- S1No account, no install, no name. You join with a code and you are a participant, not a profile.
- S2Your notes and your ink stay in your own browser. Only how many you made is ever reported, never what they say.
- S3“Not sure yet” is a private answer: it has no server counterpart at all.
- S4Saying nothing is a valid answer. Below a quorum the system reports counts and passes no verdict.
- S5The screen shows the room, never a person. A question is a question, not a name attached to one.
A teacher who ignores every number on the screen is still using Diaposi correctly.
(2)For a research partner this is not a caveat we tolerate. It is a design constraint we chose, and it has methodological consequences. An instrument that teachers resent is an instrument that gets switched off in week three, and a study with no week four is not a study. Consent, low friction and the right to stay silent are what make a multi-site, multi-semester dataset possible at all.
The instrument
Every signal below is captured today, in sessions running in production. None of it is a roadmap.
What matters most about a signal is not what it measures but whether the student had to choose to send it.
Declared signals are semantically rich and biased toward whoever is confident enough to press something, which is the population least likely to be lost. Derived signals are thin on meaning and free of that bias. A third family never reaches a server at all. The distinction is not a taxonomy for its own sake: it is what decides which signal may be trusted for which claim.
Declared
6 signals. The student chose to send it.
Derived
9 signals. Nobody has to opt in.
On device
3 signals. It never leaves the phone.
Figure 1. The three signal families, with the number of signals in each. The dashed rule marks the family whose content never reaches a server: the count of a private note is reported, its text is not.
Table 1 gives all eighteen and what each one reads. Two of them deserve a note here rather than a line in the table. Look-backs count a student scrolling back to an earlier slide after the teacher has moved on; nobody re-reads a slide they understood, and unlike a tap it needs no opt-in. The live pulse computes the same idea on a short rolling window, holds it in memory and never writes it down.
Table 1. Every signal a session records, grouped by whether the student had to choose to send it.
| Signal | What it reads |
|---|---|
| Declared — The student chose to send it | |
| Got it / Review | Self-reported understanding, timestamped at the moment of instruction rather than recalled at the end. |
| Questions | Constructive and interactive engagement, anchored to a specific slide and ranked by the room. |
| Hands raised | Help-seeking behaviour, stripped of its usual social cost. |
| Polls | In-the-moment formative assessment. Free-text replies are counted into a deterministic word cloud that every surface renders identically. |
| Quizzes | A graded knowledge check in which “I don't know” is recorded separately from a wrong answer, so not knowing is distinguishable from guessing. |
| Expectations | A stated learning goal before the session, self-rated after it. One pre and post pair per student, in their own words. |
| Derived — Nobody has to opt in | |
| Dwell time | Time on task, split between following the teacher and reading alone. Students who reached the slide are the denominator, so late slides are not penalised by those who left early. |
| Teacher time | The intended pacing baseline every student signal is measured against. |
| Follow ratio | Attention alignment, measured against the teacher's own pace rather than an absolute standard. |
| Look-backs | A student scrolling back to an earlier slide while the teacher has moved on. Nobody re-reads a slide they understood, and this one needs no opt-in. |
| Pace flags | Slides where the teacher's pace and the room's reading pace disagreed, derived from time already collected. |
| Presence curve | How the room thinned out over the deck, carrying an explicit flag for when the curve cannot honestly be read as attrition. |
| Live pulse | The same two ideas read live on a short rolling window, held in memory and never written down. It exists to be acted on, not stored. |
| Signal strength | How strong a slide's verdict is, and how much of the room actually weighed in on it. |
| Modality | A comparative condition recorded on every session. A natural experiment, already running. |
| On device — It never leaves the phone | |
| Private notes | What a student wrote and when. The text stays in their browser; the report receives the number of notes and nothing else. |
| Slide ink | That a student marked up a slide, never what they drew. The teacher's own board ink is shared, because the room watched it happen. |
| “Not sure yet” | A third answer beside Got it and Review, kept in local storage alone. The server's reaction list has two values, and this is not one of them. |
The field-level schema is not printed here. Partners receive it in full.
Table 1 names what each signal reads, not the field it is stored in. The export schema, the exact thresholds behind the verdicts and the weights behind the score are shared with research partners under agreement, alongside the raw data itself. Nothing in the argument below depends on taking that on trust: what a partner receives is the rows, and the rows can be recomputed.
An earlier version of this page overstated the on-device claim. This is the correction.
Note text and student ink genuinely never leave the device, and there is no endpoint that receives them. Their counts do reach the report: how many notes a student wrote, how many slides they marked up. A count is still a datum, it is listed in Table 1 as one, and the distinction belongs in a data protection review rather than in a footnote.
The model reads the deck. It never reads the room.
A language model now sits beside the instrument. It is worth being exact about where.
It works on slide text, before anyone joins. No student writing, no question, no note and no reaction is ever sent to it. Text is extracted at upload and summarised into topic, level, prerequisites and chapter ranges; a slide reading “6 CO₂ + 6 H₂O → C₆H₁₂O₆” means nothing on its own, and a great deal once the deck is known to be a photosynthesis lesson.
It reads the deck, once.
Text is extracted at upload and summarised into topic, level, prerequisites and chapter ranges. A single slide reading “6 CO₂ + 6 H₂O → C₆H₁₂O₆” means nothing on its own; it means a great deal once the deck is known to be a photosynthesis lesson.
Poll drafts arrive before the lecture does.
Drafting happens ahead of the session, never while a teacher is standing in front of a room. And a question about a given slide may draw only on the material shown up to that point, so a draft cannot give away a slide the class has not reached. That holds by construction rather than by a check downstream.
Quizzes are built on demand.
Prerequisite checks, chapter checks and a final, at a difficulty the teacher picks. Every generated item carries a record of the material it was drawn from, which is what makes it auditable after the fact.
Nothing reaches a student unreviewed.
Every draft is editable and nothing fires without a teacher pressing a button. The path cannot fail an upload: when the model does not answer, the call returns nothing and the session runs exactly as it would have.
Deck text
read once, at upload
Summary
topic, level, structure
Item
provenance recorded
Responses
distribution per option
Figure 2. The generated-item pipeline. Every stage is recorded, so a partner receives the deck text, the item it produced, the slide that item was drawn from, and the distribution of answers it collected.
For a research partner this is an instrumented pipeline, not a black box. Item-level psychometrics on machine-written questions, at classroom scale, with the source slide on every item, is a study nobody has been able to run at this resolution. It is study S7.
Four decisions we would defend in review
Each of these makes the instrument say less than it could. That is the point of them.
The engagement score is decomposed, not delivered.
The score resolves into three weighted components, participation, understanding and interaction intensity, and each is shown beside the raw counts it was computed from. A teacher can follow the arithmetic from the counts on their own screen. There is no model in the middle, and nothing to audit that is not already in front of them.
Below a quorum, the system refuses to judge.
A slide verdict has to clear a floor on both counts: how many students voted, and what share of the students who reached that slide they represent. Under it Diaposi reports the raw counts and passes no verdict at all. Two dislikes out of three in a room of twenty is not “reteach this”. It is three people, and the other seventeen said nothing.
The attrition curve declares when it cannot be read.
Closing the app in a lecture theatre is not leaving the lecture. So the presence curve separates a pressed “Leave session” from a silent drop, and when an in-class session is mostly silent drops it marks itself uninterpretable rather than reporting a drop-off that may not have happened.
The live reading forgets on purpose.
The pulse a teacher glances at reads only the last few minutes, is held in memory and is never persisted. A slide that confused the room twenty minutes ago is a finding for the report, not an instruction for the person currently speaking. The report is what the session leaves behind.
What should be measured, and how
All six are implemented, not aspirational.
Recording something is easy. Deciding what deserves to be recorded, at what grain, and what may never be concluded from it, is the part with consequences.
Measure the act, not the person.
Every row in Table 1 is an event a student produced: a tap, a question, a scroll back to slide 11. None of it is an attribute inferred about who they are. The difference sounds philosophical and is entirely practical. Acts can be counted, aggregated and forgotten. Inferred attributes follow people around.
Measure at the grain of instruction.
The unit is the slide, because the slide is the unit a teacher can act on. A session-level engagement score tells you a lecture went badly. A slide-level one tells you which four minutes to rebuild before next year.
Measure with a quorum, or say nothing.
A verdict has to clear a floor on the number of votes and on the share of the room they represent. Under it, Diaposi prints the counts and refuses to draw a conclusion. Most dashboards in this field will happily render a verdict from n = 2.
Hold the declared and the derived on one timeline.
Dwell traces and look-backs are hard to fake and semantically thin: a slide left open is not a slide being read. Taps are meaningful but self-reported, and only the confident half of the room sends them. Neither channel is sufficient. Carrying both through one session, at one grain, is the methodological point of the whole instrument.
Never measure the body.
No face, no voice, no gaze, no heart rate. Not because it is technically hard, since it is not, but because it changes what a classroom is, and because European law has now agreed. The ceiling on accuracy this imposes is a price worth paying.
Measure the after, not only the during.
The live number is a trigger, not a result. What a teacher changes in the next delivery, and whether the next cohort does better, is the only outcome that settles anything. A study built only on in-session metrics is measuring the thermometer.
The ladder these signals climb
Instrumenting ICAP directly usually requires an observer in the room with a coding sheet.
Chi and Wylie's ICAP framework ranks overt learning behaviours into four modes and predicts that learning increases as students move up them1. It is one of the most cited engagement frameworks in the field, and it is almost never instrumented directly.
Table 2. The four ICAP modes and the signals that stand in for each. The rule at the left of each row deepens as the ladder is climbed.
| Mode | Definition | Instrumented by |
|---|---|---|
| 1. Passive | The learner only receives information. | Dwell time and follow ratio. The slide was on screen, and we know for how long relative to the teacher's own pace. |
| 2. Active | The learner manipulates the material. | A deliberate tap on “Got it” or “Review”, a scroll back to an earlier slide, ink laid over a diagram. |
| 3. Constructive | The learner generates output beyond what was given. | Writing a question, answering an open poll in free text, stating a learning expectation in their own words before the session starts. |
| 4. Interactive | Knowledge is built between people. | Upvoting a peer's question, watching it rise to the top of the teacher's dock, and the teacher re-teaching a slide because the room said so. |
Diaposi instruments all four ICAP modes, live, at slide resolution, in an ordinary lecture hall with no observer and no setup.
(3)The regulatory position
The state of the art just became illegal to deploy.
Multimodal learning analytics is where this field has been heading: cameras, eye trackers, microphones, wearables and heart-rate variability, fused into engagement classifiers reaching roughly 82% accuracy9. It is genuinely impressive work. In a European classroom, most of it can no longer be used.
The EU AI Act classifies learning analytics in education as high-risk, and separately prohibits emotion-inference systems in educational settings outright: any attempt to read a student's emotional state from biometric data11. That is not a compliance detail to work around. It closes the road.
Table 3. Sensor-based engagement detection and the Diaposi approach, compared on the axes a review would ask about.
| Sensor-based detection | Diaposi | |
|---|---|---|
| Signal source | Facial video, gaze, voice, heart-rate variability | A tap, a question, a scroll back to an earlier slide |
| Emotion inference | Central to the method | None anywhere in the system |
| Legal status in education | Prohibited under Article 5 of the AI Act | Outside the biometric category entirely |
| Consent machinery | Per participant, for biometric data | No account, no name, nothing to consent over |
| Achievable scale | One instrumented room, one site, one language | Any room with a browser, ten countries, nine languages |
| When it reports | After the fact, on a dashboard nobody sees live | During the lecture that produced the signal |
We are not claiming the richest instrument. We are claiming the best one that is legal to run at scale in a European classroom, and scale is what this field's evidence base is actually missing.
(4)The gap
The problem with the evidence is methodological, not conceptual.
Audience response systems have been studied for two decades. The meta-analyses report a small-to-moderate effect on cognitive outcomes and a moderate effect on non-cognitive ones4.
Encouraging numbers, except that the systematic reviews built on top of them are unusually blunt about their own foundations: weak variable control, few validated instruments, non-randomised designs that inflate effects, and no consensus on how these tools contribute to performance at all5.
The reviews of multimodal learning analytics say something adjacent: the literature is largely offline dashboards and proofs of concept, with very little work on closed-loop systems running in real university environments10.
Read together, the field is asking for exactly one thing: better designs running at greater scale in genuine classrooms. That is a tooling problem before it is a research problem, and it happens to be the problem Diaposi was already built to solve.
The after
The live pulse is the trigger, and it is deliberately forgotten. The record is what changes the next lecture.
Most of the value arrives once everyone has left the room.
What a session leaves behind is a record: of what landed, what did not, what was asked and by how many people at once. It is also where the two people this is really for finally get something back.
A feeling becomes three slide numbers.
For the teacher who feels they are losing the room
Teachers who sense their engagement slipping almost always already know. What they lack is where. “The second half went badly” is not actionable. “Slides 14, 15 and 22 were flagged by a third of the room, eleven students scrolled back to slide 15, and attention there collapsed to 31% of your pace” is a task for Sunday evening.
It is also, quietly, a defence. A teacher whose course gets questioned has evidence of what they adjusted and why, something the profession has historically had to argue without.
Being heard without being looked at.
For the student who never speaks
An anonymous question that collects twelve upvotes is a student discovering that the thing they were too unsure to ask was the thing half the room needed. That is a different experience of a lecture hall than the one they came in with, and it costs them none of the social risk that kept their hand down.
Afterwards their own recap holds what they contributed, what they flagged as unclear, and the notes they took, in one place, on their device, for them. Not a score. A record of having actually been there.
The interesting question was never whether the engagement number went up. It is whether the teacher changed the next delivery, and whether the cohort that got the changed version did better.
(5)That is where the real dependent variable lives, and it is why the session report exports as rows rather than as a picture. The after is only useful if someone can compare it to the next one.
What we don't know yet
This section exists because the rest of the paper is worthless without it.
Every claim above is about what the instrument records. None of it is a claim about learning outcomes, and we are not going to borrow someone else's effect size and print it next to our logo.
A tap is a claim, not a measurement.
When a student marks a slide “Got it”, that is a statement about their confidence. Deslauriers and colleagues showed that actual learning and the feeling of learning can be strongly anticorrelated: students in active classrooms learn more and believe they learned less2. Our comprehension signal is exposed to exactly that failure mode. Quantifying the gap is study S3, not a footnote.
We have no longitudinal attainment data.
Nothing in our dataset yet connects a session signal to an exam result months later. Until it does, every statement we make about slide-level engagement is a statement about engagement, and stops there.
The derived signals are not unambiguous either.
A look-back is the closest thing we have to an unprompted “I missed something”, and it still is not proof of one: a student may go back to copy a diagram they understood perfectly. Attention time has the same problem in reverse. A slide left open is not a slide being read, which is precisely why the follow ratio never drives a per-student verdict.
Machine-drafted questions shape what gets measured.
An item written from slide text asks about what was on the slide, which is not always what mattered. Generated questions therefore risk measuring deck coverage rather than understanding. The source slide is recorded on every item so that this is testable rather than assumed. That is study S7.
Presence is not attendance.
The curve measures whether the app was open, and closing it in a lecture theatre does not mean leaving the lecture. The report says so itself when the curve cannot be read, but a study that treats it as attrition without checking that flag will reach a wrong conclusion honestly.
Single-institution data does not generalise.
Discipline, cohort size, room layout, language and local teaching culture all plausibly move these signals. This is the strongest argument for running the programme across an alliance rather than a department.
Observation changes the room.
Students who know their reactions are visible on a screen may behave differently for that reason alone. That is an interesting effect in its own right and a confound in everything else. Any serious design has to account for it.
Built in Europe, measured against the goals
For a funding body this alignment is not decorative. It is the frame most European instruments are assessed against.
A live-classroom platform for innovative teaching, built in France and designed to adapt at large scale. Beyond desirability, feasibility and viability, Diaposi is measured against the UN 2030 agenda: impact is produced by the same mechanics that make the product work, and each of the four goals below maps to a mechanism already in the product rather than to an aspiration bolted on afterwards.
Quality Education
The teacher finds out which concept failed during the lecture that failed it, not three weeks later in an exam.
Industry, Innovation & Infrastructure
European research infrastructure for classroom interaction data, built to open standards rather than a private schema.
Reduced Inequalities
No account, no install, any browser on any phone, in nine languages chosen per device. Anonymity removes the social cost that silences the students who most need to ask.
Partnerships for the Goals
Ten institutions across ten countries already reachable through the INGENIUM alliance, which is the reason a multi-site study is realistic at all.
Figure 3. The four Sustainable Development Goals Diaposi is assessed against, each with the mechanism that produces it.
Data and governance
The fastest way to lose an academic partner is to hand them a dashboard screenshot.
Session data exports as CSV and structured JSON today, one row per slide, carrying every derived figure the interface shows.
Raw export
ShippedPer-slide CSV and full session JSON: attention times, votes, questions, look-backs, pace flags, the presence curve, polls, quizzes and expectation ratings. No aggregation you cannot undo.
LTI 1.3, xAPI, Caliper
CommittedLTI 1.3 / Advantage for launching inside Moodle and Canvas without an integration project, an xAPI event profile for research pipelines, and Caliper Analytics 1.2 for institutional reporting12.
References
In order of appearance in our own thinking rather than alphabetically.
- [1]Chi, M. T. H., & Wylie, R. (2014). The ICAP framework: Linking cognitive engagement to active learning outcomes. Educational Psychologist, 49(4). Source
- [2]Deslauriers, L., McCarty, L. S., Miller, K., Callaghan, K., & Kestin, G. (2019). Measuring actual learning versus feeling of learning in response to being actively engaged in the classroom. PNAS, 116(39), 19251–19257. Source
- [3]Freeman, S., et al. (2014). Active learning increases student performance in science, engineering, and mathematics. PNAS, 111(23), 8410–8415. Meta-analysis of 225 studies: 1.5× higher failure rate under traditional lecturing. Source
- [4]Nelson, C., et al. (2012). The effects of audience response systems on learning outcomes in health professions education. BEME Guide No. 21. Medical Teacher. Source
- [5]Do audience response systems truly enhance learning and motivation in higher education? A systematic review (2025). Humanities and Social Sciences Communications. Source
- [6]Castillo-Manzano, J. I., et al. (2016). Measuring the effect of ARS on academic performance: A global meta-analysis. Computers & Education. Source
- [7]Drachsler, H., & Greller, W. (2016). Privacy and analytics: it's a DELICATE issue. A checklist for trusted learning analytics. LAK'16. Source
- [8]Tsai, Y.-S., et al. (2018). The SHEILA framework: Informing institutional strategies and policy processes of learning analytics. Journal of Learning Analytics. Source
- [9]A survey of multimodal learning analytics: Data, methods, systems, and responsible deployment (2026). Future Internet, 18(3), 115. Source
- [10]From heartbeats to actions: Multimodal learning analytics of cognitive and behaviour engagement in real classrooms (2026). Learning and Instruction. Source
- [11]Regulation (EU) 2024/1689 (AI Act). Annex III, high-risk: education and vocational training; Article 5, prohibition on emotion inference in education. Source
- [12]1EdTech. Learning Tools Interoperability (LTI 1.3 / Advantage) and Caliper Analytics 1.2. Source
- [13]SoLAR. LAK26, Bergen: Learning Analytics and AI Synergy. Source
- [14]ANRT. Le dispositif CIFRE. Source
- [15]European Commission. Erasmus+ 2026 call; EIT Higher Education Initiative 2026–2028; Horizon Europe Work Programme 2026–2027. Source