The 2025 edition of Epita’s symbolic intelligence project built a system of agents that analyzes an argumentative text. Its Argumentum plugin detects fallacies there by walking down the game’s taxonomy level by level, instead of picking in one shot from more than a thousand labels. The shape of the tree means each step offers only a few branches — and the answer arrives together with the path that leads to it.
Illustration: docs/publication/articles/illustrations/3b-du-chemin-a-la-carte.png, Argumentum repository, commit 981e5233 — a montage of the repository's engravings (Cards/Fallacies/Assets/Fallacy-front/): levels 7, 7.3, 7.3.1 and 7.3.1.2, with the path number above and the card name below.
The project
It is a public repository: software agents — analysis, logic, fallacy detection — are orchestrated there by Semantic Kernel, a framework that lets a language model call tools, and they rely on a formal reasoning engine written in Java. The platform is educational: each building block is an exercise in symbolic artificial intelligence architecture, where you manipulate symbols and rules rather than statistics.
One thousand four hundred and eight labels
Argumentum's fallacy taxonomy has 1,408 nodes, from the root node down to the special cases, reaching up to ten levels of depth. Asking a model “which fallacy is this?” in front of the entire list means making it draw one label out of more than a thousand, with no structure to guide the choice.
Yet the shape of the tree says otherwise: each node has at most 9 children, and half of the nodes that have any have no more than two. A choice guided by structure replaces the gamble.
The descent, step by step
The plugin loads the taxonomy as CSV, the format of the game's repository. At each step, it shows the model the current node and two levels of its descendants; the model either confirms the node — this is it — or asks to descend toward a child, sometimes skipping a level. Each step costs one call to the model. A confirmation too high in the tree is refused: at fewer than two levels of depth, the label is too general to be an answer.
An example, built on a card from the game (no analyzed text is shown here). The “Straw man” card lives at path 7.3.1.2: Obstruction, then Ad hominem, then Accusation of inconsistency, then Straw man. A descent that reaches it looks like this: the text seeks to disqualify the opponent rather than answer them — descend toward Obstruction; it targets the person — descend toward Ad hominem; it reproaches them for having contradicted themselves — descend toward Accusation of inconsistency; the very object of the attack is distorted — confirm Straw man. Four steps, each among few children.
What bounds the descent
There is no depth ceiling: the leaves of the tree are the only ceiling, and the agent can descend all the way to them. The safeguard against runaway behavior is a budget of calls to the model: by default, 18 per branch and 240 for an entire analysis. The descent starts with a wide net that retains candidates, of which at most four branches descend in parallel; at a fork, if several children look promising, the following ones are explored in parallel, within a limit of twelve sub-branches. When a branch confirms a deep node, its sibling branches that were exploring an ancestor give up: the most specific answer wins. And if the descent fails, the plugin does not give up — it retries the choice in one shot, without guidance.
The trace
Each descent leaves its navigation trace: the path followed, the confirmations, the abandonments. The answer can be explained by its itinerary — the system can show why it proposed this node rather than another. On a deep taxonomy, this is perhaps the main product of the method: not only a label, but the reason for the journey.
After the project
The taxonomic funnel was carried over as is into CoursIA, a public repository of course materials on artificial intelligence, which integrates it into its argument analysis series. The orchestration of plugin-based agents, for its part, is still a work in progress there.
Sources: repository Argumentum, commit 32408fdd (counts established on November 8, 2026 from Cards/Fallacies/Argumentum Fallacies - Taxonomy.csv, path column). Repository Epita 2025, pinned to commit acaf670e9 (November 6, 2026): fallacy_workflow_plugin.py and taxonomy_navigator.py — budgets and mechanics cited from these files. Carried over into CoursIA: issue #19287 (done) and issue #18390 (in progress).