Brief: Scheduler–LLM separation

Type: types/write-brief.md

Intent

Develop the theory of why code complements LLM instructions, with system bookkeeping as its central leverage point. The reader should understand when and why moving defined state transitions into code helps more than improving instructions. Make bookkeeping the subject and error correction the explanation. Secondarily, make this the concrete entry point to the two-note argument; accessibility must serve the theory, not replace it.

Must keep

• Explain the leverage, rather than merely recommend separation. Preserve the reliability asymmetry that existing citations rely on.
• Distinguish enforcement of a chosen rule from judgment about whether it is the right rule. Leave room for model-chosen policy and a changing boundary.
• Stand alone as an entry point: establish enough of complementarity to explain bookkeeping's leverage without requiring the companion note. Delegate the general operation-class distinction to that note; a little overlap is useful.

Must exclude

• A taxonomy-led exposition, benchmark survey, or general argument that code is superior to models. Supporting theory and examples must serve the central argument, not become separate destinations.
• Claims of free reliability, guaranteed low bias, or bookkeeping's universal supremacy over other uses of code. Persuasion must not outrun the evidence.
• Defensive disclaimers against hostile readings. Keep conditions that explain where the argument works; do not repeatedly deny claims the note never makes.