Categories

# Categories

Every modality of LLM work. Categories are intentionally orthogonal: each names one kind of work rather
than a model tier, fallback, or miscellaneous bucket. `impact` is how much dropping the tier costs you —
substrate/reasoning work compounds downstream (protect it), volume/interactive work does not.

The **Systems** column says which boundaries the category's calls cross today and which decision-shaped
work in it is queued for System 1 ([two systems](00-two-systems.md)). A category is not a system: `action`
carries both, and the same is expected of `predict` and `interaction` as their decisions migrate.

| Category | Impact | Systems | What it covers |
|---|---|---|---|
| **generate** | substrate | 2 | Scene generation & reconstruction — committed to the substrate. |
| **plan** | substrate | **1** (scene count) + 2 | Beat plans, reasoning graphs (CRG), phase graphs (PRG) — the causal spine. |
| **extract** | substrate | **1** (topic filing, location parents, findings routing, system invocations, paradigm detection) + 2 | Analysis pipeline — the knowledge graph every force/search reads from. |
| **write** | substrate | 2 | Prose generation — longest, highest-volume outputs. |
| **predict** | reasoning | **1** (stream-prior evidence, log type, attention, and new-outcome gating) + 2 | Variable logits (present/future/rescore) — decides which branches generate. |
| **review** | reasoning | **1** (prose / plan verdicts) + 2 | Branch / prose / plan quality judging — read-only critique. |
| **synthesis** | volume | **1** (research loop/reminder shape) + 2 | Editorial distillation — note-merge, chat-compact, report, search synthesis. |
| **gameTheory** | volume | 2 | Per-scene game-theory decomposition — additive, never mutates substrate. |
| **question** | volume | 2 | Learning question extraction — study aid over committed prose. |
| **interaction** | interactive | 2 | Chat, surveys, interviews, and the Pilot agentic turn. |
| **action** | interactive | **1** (agent seat) + 2 (table talk) | Live Scenario moves — on the play clock, latency-bound. |
| **intuition** | interactive | **1** (stream opening pricing) + 2 | Belief-stream landing distributions, log type, and horizon; creative micro-suggestions and outcome prose remain generated. |

The twelve were reviewed against the split (2026-09-19) and kept: each still names one kind of work, and
the system axis is carried per call, not per category. What the split *does* change is `action`'s reading
— "latency-bound" is now met by System 1 for the seat's move itself, so the tier its profile assigns
governs only the generated remainder of a live turn.

Raw Markdown source