Quell debate: plan map

design · draft · NLP and argumentation · edited 2026-11-06

Draft 2, 2026-11-03. Supersedes .docs/designs/quell-debate-roadmap.typ.

Quell’s players practise debating real philosophical questions. They read in-world texts, take notes, build arguments, and debate NPCs. Between runs they keep only what they know.

The plan has one current focus: argument structure. Can we turn natural language into an argument map (IAT + PTA, stored as xAIF) well enough, and fast enough, to use? Judging who wins is later research.

Click a node to open its page.

flowchart TD
  GOLD["Gold maps<br/>hand-made, reviewed"] --> ARMS{{"Which way from<br/>text to map?"}}
  ARMS --> DIRECT["Direct generation<br/>NL → JSON"]
  ARMS --> ORACLE["Oracle pipeline<br/>rules + small models"]
  DIRECT --> SPIKE["Spike: local, CPU,<br/>few-shot NL → xAIF"]
  SPIKE --> FAST{"Fast and good<br/>enough?"}
  ORACLE --> FAST
  FAST -->|yes| RUNTIME["Parse at runtime:<br/>free-text input possible"]
  FAST -->|no| AUTHOR["Authoring only:<br/>speeds up gold maps"]
  RUNTIME --> GAME["Quell integration"]
  AUTHOR --> GAME
  GAME -.-> EVAL["Evaluation<br/>(later research)"]
  GOLD -. "scores every arm" .-> FAST
  click GOLD href "debate-gold.html"
  click ARMS href "debate-arms.html"
  click DIRECT href "debate-arms.html"
  click ORACLE href "debate-pipeline.html"
  click SPIKE href "debate-spike.html"
  click FAST href "debate-arms.html"
  click RUNTIME href "debate-game.html"
  click AUTHOR href "debate-gold.html"
  click GAME href "debate-game.html"
  click EVAL href "debate-evaluation.html"
  classDef now fill:#fde68a,stroke:#b45309,color:#111
  classDef later fill:#e5e7eb,stroke:#6b7280,color:#111
  class SPIKE now
  class EVAL later

The spike (yellow) is the next piece of work.

Constraints

• Minimal GPU. The game stays graphically basic, so models can use some GPU, but the plan must not depend on a strong one.
• Local models first. No cloud LLMs in the game. Cloud models are fine for authoring and gold labelling. Claude and other agents are part of the team.
• Pricing: free or a one-time payment, no subscription. Hosting a few dollars a month for a modest player base is acceptable.
• One developer. Anything that needs a team to maintain is out.
• The texts are real books. The lab never commits or publishes them.

Not decided

• How a debate is won: evaluation.

See also: phases, open questions.

What links here: Debate, NLP and argumentation