Debate

Claude Opus 5.5, 3 Nov. 2026

NLP and argumentation

Debate is Quell’s planned central mechanic. Players read in-world texts, which are excerpts of real philosophy. They highlight and take notes, build arguments from what they have found, and debate the game’s characters on questions without easy answers. Some debates must be won to progress; others only change the conversation. Each run starts over, but players keep what they have learned. Their progress is knowledge rather than items or levels, as in a knowledge-based roguelite.¹

The player’s loop

flowchart TB
  READ["read in-world texts"] --> HL["highlight and take notes"]
  HL --> BANK[("collected statements")]
  BANK --> ARGUE["build an argument"]
  ARGUE --> DEB["debate a character"]
  DEB -->|character answers| ARGUE
  DEB -->|win or move on| NEXT["next area or run"]
  NEXT -->|keep what you know| READ

• Reading. In the plan, a highlight snaps to a statement that has already been identified in the text’s argument map, so the game always knows which statement the player chose. Notes remain free text. See Reading notes for how this relates to the notes lab.
• Arguing. In a thinking editor, the player links a premise to a conclusion and picks the kind of connection.
• Carrying over. A new run keeps some highlights, texts and established conclusions, which can open new conversations.

The thinking editor needs no language model during play. Every collected statement is analysed in advance, so the argument’s type can be derived by rule. The player performs the classification, and that is where the learning happens.

Analysis at authoring time

The texts are analysed offline. Each excerpt becomes an argument map: its statements, how they connect, and the type of each argument under the Periodic Table of Arguments. Models and agents may propose maps, but a person reviews each one in the xAIF editor.

Reviewed maps serve two purposes. They are game content, and they are the gold standard used to score automated approaches.

So the game does not need an automatic analysis pipeline in order to exist. Hand-made maps are enough for a playable debate.

²

Constraints

• Local models in the game. No cloud model may run during play. Cloud models and agents are acceptable for authoring and labelling.
• Modest hardware. The plan must not depend on a strong GPU.
• No subscription. The game is free or bought once.
• One developer. Anything that needs a team to maintain is ruled out.
• Real books. The source texts are not committed or published by the labs. Whether they can be shipped is a separate question for each text.

Structure first, judgement later

At first, the game will check only an argument’s structure: whether it has a recognizable type, and whether the chosen connection fits that type. Whether an argument is good, and who wins, is later research. The leading candidates are:

• CAPNA: judging an argument in stages (relevance, reasoning, language). Each kind of argument invites its own critical questions, which could become what the character says in reply;
• Dung’s abstract argumentation: a debate as a graph of attacks, in which the arguments that survive are the ones that win;
• value-based argumentation: each character ranks values, so the same argument can succeed against one character and fail against another.

Each has its own article: CAPNA, Abstract argumentation and Value-based argumentation.

Open questions

Where it would live in Quell

The draft plan sketches reading, notes and debate services, a debate screen, and content exported from reviewed maps as RON assets. It also proposes a small pure-Rust crate (working name q_pta) for the type table and derivation rules, checked against the same conformance examples as the editor. Dialogue has been sketched in Yarn, but bevy_yarnspinner has not been evaluated for Quell.³

See also

• Argument maps — what an argument map contains and how one is made.
• Reading notes — interpreting the player’s highlights and notes.
• Research labs — the labs where this work happens.

External links

• Periodic Table of Arguments — Wagemans’ classification of argument types.
• Argumentation framework — background on Dung’s abstract model.

What links here: Argument maps, Abstract argumentation, Periodic Table of Arguments, Value-based argumentation, Reading notes, Glossary, Research labs