Phases

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

Two tracks that meet at the exporter. The game track can start from hand-made maps; it doesn’t wait for any model.

flowchart TD
  subgraph A["Track A · content and NLP (lab)"]
    A0["A0 · gold maps"] --> A1["A1 · spike: NL → xAIF"]
    A0 --> A2["A2 · shared core:<br/>type table, conformance suite"]
    A1 --> A3["A3 · more arms,<br/>scored on gold"]
    A2 --> A3
    A3 --> A4["A4 · assisted annotation"]
  end
  subgraph G["Track G · game (Quell)"]
    G1["G1 · q_pta crate"] --> G2["G2 · exporter + assets"]
    G2 --> G3["G3 · vertical slice"]
    G3 --> G4["G4 · playtest"]
  end
  A2 --> G1
  A0 --> G2
  A3 -. "runtime parse?" .-> G3
  click A0 href "debate-gold.html"
  click A1 href "debate-spike.html"
  click A3 href "debate-arms.html"
  click G1 href "debate-game.html"
  click G3 href "debate-game.html"

• A0 · Gold maps (now): draft, critique, fix, log judgement calls.
• A1 · Spike (next): see the spike.
• A2 · Shared core: the PTA table as one JSON file; ra-inputs.json as the conformance suite for TS, Python and Rust.
• A3 · More arms: the oracle pipeline, oAMF, hybrids, compared on gold.
• A4 · Assisted annotation: arms write draft xAIF with confidences; the editor shows them for review.
• G1–G4: the core crate, exporter, a vertical slice (one text of about 20 statements, 3–4 levers, one NPC), then a playtest: is building one argument against one NPC fun?

Deferred: fine-tuning beyond the spike’s LoRA arm, until the materials settle.

What links here: Open questions, Quell debate: plan map