Arms and decisions

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

An arm is one way to answer a question: a rule, or a model with a given setup. We try several arms, score each against the gold maps, and keep what wins. Nothing here is chosen yet.

The families

flowchart LR
  T["passage or<br/>player line"] --> D["A · direct generation<br/>one model writes the map"]
  T --> O["B · oracle pipeline<br/>rules + one small model per question"]
  T --> M["C · oAMF<br/>ARG-tech's mining framework"]
  T --> H["D · hybrids"]
  click D href "#direct"
  click O href "debate-pipeline.html"
  click M href "explainer-argument-mining.html"
  click H href "#hybrids"

• A, direct generation: light on labour, heavy on one model. The spike is its first test.
• B, the oracle pipeline: from pta-atip-algorithm.typ. More reliable and easier to debug, but needs compute and work for each specialised model. Details: the pipeline.
• C, oAMF: already outputs xAIF. Needs research before it becomes a real arm.
• D, hybrids: mix A and B, step by step.

Every arm meets the same gates

flowchart TD
  gA["an arm's output"] --> gV{"parses and lints?"}
  gV -->|no| gX["fix the arm<br/>or drop it"]
  gV -->|yes| gB{"beats the rule<br/>baseline on gold?"}
  gB -->|no| gX
  gB -->|yes| gL{"fits the latency<br/>budget on target CPU?"}
  gL -->|yes| gR["runtime candidate"]
  gL -->|no| gC{"can cost be cut?<br/>caching, fewer examples,<br/>LoRA, GPU layers"}
  gC -->|yes| gL
  gC -->|no| gAU["authoring only"]
  click gB href "debate-gold.html"
  click gC href "#direct"
  click gR href "debate-game.html"

“Beats the baseline” is judged stage by stage, exemplars first. The latency budget isn’t set yet: see open questions.

A · Direct generation

flowchart TD
  D["direct generation"] --> OUT{"output format"}
  OUT --> OX["raw xAIF"]
  OUT --> ON["compact notation<br/>+ deterministic expander"]
  ON --> SZ{"model size"}
  OX --> SZ
  SZ --> S1["0.5–1B"]
  SZ --> S2["1.5–3B (spike)"]
  SZ --> S3["7–8B"]
  SZ --> S4["cloud (authoring only)"]
  S2 --> PR{"prompting"}
  PR --> P0["zero-shot"]
  PR --> P1["fixed few-shot"]
  PR --> P2["retrieved few-shot (spike)"]
  P2 --> CD{"constrained<br/>decoding?"}
  CD --> C0["free text, then parse"]
  CD --> C1["JSON schema / GBNF (spike)"]
  C1 --> LA["later: LoRA, prompt caching,<br/>GPU layer offload"]
  click ON href "debate-spike.html"
  click S2 href "explainer-quantization.html"
  click P2 href "explainer-in-context-learning.html"
  click C1 href "explainer-constrained-decoding.html"
  click LA href "explainer-lora.html"

Choices marked “spike” are the spike’s starting point. Each other leaf is a later comparison, run only if the spike works at all.

D · Hybrids

Some likely mixes, none designed yet:

• Generate, then check. A writes the map; B’s oracles check it. For example, NLI checks each relation, and the editor’s rules derive form and substance.
• Pipeline, with one generative step. B does everything except reconstruction, the hardest step, where a generative model proposes rewordings.
• Generate the skeleton, classify the details. A segments and relates; B’s small classifiers fill in each statement’s fields.

Pick hybrids from where A and B each fail on gold, not in advance.

What links here: Gold maps and measurement, Open questions, Phases, The oracle pipeline, Quell debate: plan map, Argument maps, Debate, Argument mining