11/7/2026 - Fallacies × Scenarios: The Game as a DatasetNews | Argumentum
Skip to main content
Home  › ... News

Fallacies × Scenarios: The Game as a Dataset

11/7/2026 Evolution
Fallacies × Scenarios: The Game as a Dataset

Draw a scenario, play a fallacy card: every round of Argumentum combines the two, and every combination is a prompt whose answer is known in advance. The “fallacy detection” series of the CoursIA project turns this into a generator of annotated examples that covers the entire taxonomy tree — coverage that, according to the series documentation, no corpus of natural texts provides.

Illustration: docs/publication/articles/illustrations/5b-une-manche-deux-cartes.png, Argumentum repository, commit 981e5233 — montage of the “History” card back (Cards/Scenarii/Assets/Scenario-back/img-histoire.png) and the fallacy card (Cards/Fallacies/Assets/Fallacy-front/7.3.png) — one round of the game, in a single image.

One turn of play, one prompt

The scenario deck describes 167 situations, divided into seven themes — history, mythology, politics, pop culture, intimate relationships, personal life, professional life — identifiable by the back of the cards. The rules sum it up this way, on the card “The Smooth-Talk Mill”:

In each round, starting from a randomly drawn scenario, one of the players finds, within a limited time, as many fallacious arguments as possible illustrating the cards drawn in succession.

In this variant, the scenario stays and the card changes: each card drawn calls for a new argument — fallacious, and fallacious of that particular kind. The player who tries to defend a scenario with the drawn fallacy produces, without knowing it, a labeled example: the situation is the text, the card is the label.

This is exactly the shape of a training dataset for a classifier, a program that sorts a text into a category: texts, each carrying the label of the category it belongs to.

Two scales

The first scale is that of the box. There are 167 scenarios and 175 fallacy cards: the game thus offers 29,225 playable prompts, all different.

The second scale is that of the tree. The taxonomy describes 1,408 nodes — from the major families down to special cases — of which only 175 carry a card: not all of them deserve a printed copy. One mind map per language draws this entire tree. Each scenario is crossed with each node of the tree, and the label of each pair is known in advance; the operation that enumerates “each element on one side with each element on the other” is what mathematicians call a Cartesian product.

CoursIA doesn't stop at fallacies: it also crosses the nodes of argumentative virtues. Each pair is an example whose label — the node — is known by construction: generating the label is trivial, it is the second factor of the pair.

Why the virtues

In the Cartesian product, each (scenario, fallacy node) pair carries a fallacy label: the table drawn from fallacy nodes alone contains, by construction, no example of an argument that holds up.

Yet a classifier that saw only fallacies could not learn what sound reasoning is: it would be able to name diseases without ever having seen a healthy patient. This role falls to the virtues deck — relevant arguments, honest presentation, lexical accuracy… — and the series documentation puts it this way: the virtues give the classifier the “sound argument” class, without which it would see only fallacies.

What CoursIA does with it

The CoursIA project (public repository) devotes a series of notebooks to fallacy detection — the “Fallacy Detector.” A Jupyter notebook is a document that mixes text and code that runs step by step; each step shows its result and can be replayed. The dataset-building notebook, accompanied by a dedicated script, enumerates the 272,043 pairs and then draws three sets from them (measurements on the CoursIA repository pinned at commit 7eea21eb, on November 7, 2026):

Set Rows Coverage
Training 14,000 12,000 fallacies, 2,000 virtues
Validation 1,629 one pair per node
Test 3,258 two pairs per node

The split follows an anti-leakage rule: the three sets use distinct scenarios. Without this, a model could recognize the situation rather than the reasoning: it would learn the scenario, not the fallacy. Validation is used for tuning without touching the test set; the test set measures, only at the end, on never-seen data.

Each draw is reproducible: the random generator's seed, the budgets and the proportions are recorded in a manifest, along with the fingerprints of the source files — Argumentum's CSVs are identified there by their SHA-1 hash, and the regenerable training set by a SHA-256 hash. The copy of the scenarios embedded in CoursIA is deposited with its license and attribution: the corpus texts remain under Creative Commons BY-SA 4.0, credited “Argumentum.”

What remains to be proven

The (scenario, node) pair gives the label, not the text: the argument illustrating each prompt remains to be written. This work is planned through generation by a language model — this must be said plainly. At the cited commit, this generation and the human test meant to check it are explicitly “upcoming.” The announced safeguards echo those of the game — several votes, just as several players judge an argument — and a test by humans outside the project.

No detection performance is claimed here: no measurement has been published yet, and nothing is announced as achieved. What is achieved, however, is the structure: for each node of the tree, in each set, prompts whose answer is certain — and the path linking a draw of cards to a verifiable corpus.


Sources: repository Argumentum, commit 6da19683 (counts from Cards/Scenarii/Argumentum Scenarii - Cards.csv, Cards/Fallacies/Argumentum Fallacies - Taxonomy.csv, the virtues taxonomy and Cards/Rules/Argumentum Rules - Cards.csv, November 7, 2026). CoursIA repository, commit 7eea21eb: series README, build script and splits manifest. Follow-up: Epic CoursIA #10355, external human test CoursIA #17578.

‹ Back to the list

Contact Us

Newsletter Sign-Up

We were unable to confirm your subscription.
Your subscription has been confirmed.

Newsletter

Subscribe to our newsletter to keep up with the latest news from Argumentum.

We use Sendinblue as our marketing platform. By submitting this form, you acknowledge that the information you provide will be transferred to Sendinblue in its capacity as a data processor, in accordance with its terms and conditions of use.