The scientific basis of Dialectica

01Where the weights come from, and they are not ours

The decision. The assessment is split across content 40 % · style 40 % · strategy 20 %, out of 100, which then converts to the local grading scale.

The sources, in this order. The weighting is the World Schools marking standard, the format of the international championships since 1988. The sheet we ship first is its Italian curricular adaptation: Presutti, G. / SN-DI, Griglia di valutazione per il Debate preparato (March 2020) — five blocks scored 0.4 to 2.0 out of 10: two for content (argumentative and documentary research competences; active listening and quality of rebuttal), two for style (para-verbal; non-verbal), one for strategy (strategic competences and use of POI).

The finding. Four tenths, four tenths, two tenths — the same vector in the international standard and in the national curricular sheet, because the second is an adaptation of the first. That fixes what can be claimed: the weighting is not a national choice we inherited, it is the tournament standard; the national sheet is what turns it into descriptors a teacher can mark against.

What we do. That vector is the default and it does not move. A school that needs it can switch on 50/30/20, published by Russo & Refrigeri (2023) for curricular debate by subject, which argues its own deviation: content weighs more because studying sources other than the textbook is the precondition for everything else. Every assessment records which weighting produced it — if the weights are configurable and the active one is not stored, two cohorts stop being comparable and nobody notices.

What travels, and what does not. The weighting travels: every school on the World Schools circuit already uses it. The descriptors, the bands and the grading scale do not — those are the national sheet, and ours is the Italian one because Italy is the first market. We will not ship a preset for a country whose official assessment sheet we have not read: a preset written from a secondary description is an invented rubric.

02What cannot be assessed on text — and we say it first

The decision. Dialectica assesses a student’s written turn. It does not assess voice, gesture or presence.

The source and the finding. In the SN-DI griglia, two of the five blocks — para-verbal (tone, rhythm, pauses, reading versus esporre a braccio) and non-verbal (eye contact, gesture, posture) — are entirely prosodic and bodily. They are 4 points out of 10: 40% of the rubric, and no tool that reads text can score them.

That this is a boundary of modality and not a shortcoming of the build has independent confirmation: Ruiz-Dolz et al. (2021), building the VivesDebate corpus from a real university tournament, explicitly excluded oral fluency, grammatical correctness and non-verbal language from their assessment file, because those aspects — which the jury did use — were not reflected in the textual corpus.

What we do. Both blocks appear on the student’s screen flagged, with the sentence that their teacher assesses them in class because the system reads text. On text, Dialectica covers the content block in full and the part of strategy that does not depend on live clash. Which makes us a between-debates training tool, complementary to running the debate in class and not a substitute for it. One caution on the 40 % itself: it is a property of this rubric family, not of debate assessment in general. Each market's sheet gets its own figure, stated with the preset rather than repeated as a slogan.

03The three dimensions come from the canonical systematisation

The decision. Contenuto, stile and strategia are not categories invented for the product.

The source. Wachsmuth, H. et al. (2017), Computational Argumentation Quality Assessment in Natural Language, EACL — the reference systematisation: 15 argument-quality dimensions organised into three families, derived from the earlier theories and validated on the Dagstuhl-15512 corpus.

The finding, and the mapping.

One correspondence comes out especially clean: the authors define global sufficiency as the argumentation adequately rebutting the counterarguments that can be anticipated — which is exactly the strength Dialectica labels as anticipation of objections.

The nuance we also publish. The same work reports that the three families correlate strongly with one another. That is: three separate dimensions carry less independent information than their separation suggests. It does not invalidate the rubric, but it fixes what can be said: the separation is pedagogical, not orthogonal. It exists because a student needs to know where to improve, not because the three measure independent things.

What we do not claim: that Dialectica implements the 15 dimensions. The mapping is of the three high-level ones.

04Fallacies: how many can be detected, and how reliably

The decision. The product identifies a broad catalogue of fallacies with operational definitions, but any aggregate measurement is done over families, not over the full catalogue.

The sources and the finding. The performance of automatic fallacy classification falls with the number of classes:

Above six classes, performance collapses.

And the human figure, which is the important one. Habernal, Pauli & Gurevych (2018) measured how well a person identifies the type of fallacy: macro-F1 of 77%, with enormous variation by type — appeal to irrelevant authority 95%, red herring around 60%. Some fallacies are intrinsically harder to agree on than others, and that is a property of the phenomenon, not of the annotator. But the collective vote converges: with three or more judgements, the aggregated label matched the expert one in 90 of 92 cases (97.8%).

What we do. Three things.

1. 01

   The product catalogue is broad because its job is to explain to the student, and a precise explanation is worth more than a coarse label.

2. 02

   Any performance figure is reported broken down by family, never as a single global number that hides the fact that some are easy and others are not.

3. 03

   The machine’s label is not the last word; the teacher’s decision is.

A future line, not yet in the product. Goffredo et al. (2025) introduce fallacy repair: beyond detecting them, generating a reformulated version of the argument without the fallacy. For a fourteen-year-old, “this is a false cause” is a label and “here is your argument without it” is a lesson. It is plausibly the best formative feedback in this whole literature, and it is on our roadmap, not in today’s product.

05The teacher decides, and the design is built so that they can

The decision. The AI reads and proposes, the system structures, the teacher decides. And the proposal is presented in a deliberately unassertive way.

The source. A study published in Frontiers in Psychology (2026) with a 2×2 design and 214 teachers who assessed the same text, varying only the number shown by an automated report and its visual highlighting.

The finding. The effect of the numeric cue on the teacher’s judgement, measured as η²p:

The text did not change. That is the magnitude of anchoring, measured. The study closes by calling for neutral report design, independent teacher judgement, human oversight and training on automation bias.

What we do.

1. 01

   The teacher’s field is never locked, and the value the AI proposed always stays in a separate read-only column: that column is the record of what was proposed, not the field the grade is typed into.

2. 02

   When the model declares low confidence, the proposal is shown as a band, not a point and the three fields start empty — the reliability gate that Al-Zawqari, Safa & Vandersteen (2026) introduce to soften strong recommendations when confidence is low. At high or medium confidence the fields arrive pre-filled, which is the anchoring risk we are choosing to run and measure.

3. 03

   The confidence level is frozen into the record of the assessment, so that it can later be checked whether the caution achieved anything.

And we state it falsifiably. If showing a proposal as uncertain does not increase the time a teacher spends deciding, nor the number of corrections they make, then the measure is not working and we will say so in those words.

An honest caveat about this source: the study is about AI-detection reports, not grade proposals; the experimental manipulation is extreme and the sample is university-level. We cite it as evidence that the phenomenon exists and is large, not as a measurement of Dialectica’s case.

06How we measure whether the system resembles a teacher

This is the part almost nobody publishes, and where it would be easiest to overstate.

The objective. Not to “get the grade right”, but to know in which direction and by how much the machine’s proposal deviates from the teacher’s judgement, per dimension and per grade band.

07Why practice is curricular and not extracurricular

The decision. Dialectica is assigned in class, as a compito. It is not an app a student signs up to on their own.

The sources and the finding. Russo & Refrigeri (2023) note that evening and extracurricular projects “mantengono intatto il principio di autoselezione”, which in practice selects the most extroverted and sociable personalities leaving out precisely those weakest in learning; and being extracurricular, the activity falls outside the teacher’s assessment (Refrigeri & Russo, 2020).

The empirical reinforcement is blunt. Habernal et al. (2018), deploying a voluntarily adopted serious game on fallacies: in five weeks and across three recruitment channels they gathered 58 players and 296 arguments; 69% played for less than an hour and only 9% came back after five days. A voluntarily adopted argumentation tool does not retain.

What we do. Curricular assignment by the teacher is not a limitation of our model: it is the only documented route to sustained practice, and the only one that reaches the student who would never sign up in the afternoon. Put as the authors put it: classroom debate remains the place where learning happens; asynchronous training is what lets a student arrive there with hours of practice behind them.

08Inclusion: voice assistance is logged, never scored

The decision. Use of dictation and other aids is recorded as a compensatory measure and is excluded by design from any score calculation.

The sources. De Conti (2025), Debate 4.0, develops inclusivity as one of the three legitimate axes of AI in debate: simplifying a text, creating tailored explanations, rendering a text in voice format, and equity of access to competitions for students with BES or disabilities. Sarcina (2024) argues that rubrics are not only instruments of summative assessment but of continuous feedback, and that the student knowing the indicators in advance is what enables the self-assessment dimension.

What we do. The student sees the rubric — the real descriptors of the griglia, written as instructions — before writing their turn, not only in the final report. And the second half of the accessibility sentence is the half that matters: not “we have dictation”, but voice assistance is logged as a compensatory measure and stays out of the grade.

09Where Dialectica sits in the international state of the art

There are three active groups in this terrain, and none of them works on school debate in Italian.

The gap. We have found nothing equivalent for school debate in Italian. We phrase it that way — “we have not found” — and not as “it does not exist”, because it is more honest and harder to rebut.

And that is why we do not export calibrations across languages. Two findings rule it out, and it is worth saying before anyone asks:

Consequence. Any reliability figure we publish will be tied to the language and the rubric it was measured with. We will not say that a calibration obtained in Italian holds for Spanish or German, because the literature says it does not.

10What this page does not claim

This section is deliberate and will not be deleted in future versions.

References