AI Models — tiers, categories, and allocation profiles

The single reference for how the engine picks a model for each kind of work. Source of truth: src/lib/core/ai-profiles.ts (systems, tiers, categories, profiles) and src/lib/core/model-resolver.ts (modelFor(category) for generation, decisionModelFor(category) for decisions). This doc mirrors that data — if you change the code, update the tables here.

Two systems first

A call is either a decision (System 1 — a typed choice, score, or yes/no probability through callDecide) or a generation (System 2 — text and artifacts through callGenerate*). Decide that before anything below: tiers and profiles allocate capability to System 2; System 1 rides one decision model. The two systems owns the rule and the migration ledger.

Orthogonality is the routing principle

Categories name what kind of work a call performs; profiles allocate how much capability that work receives. Keep those questions separate. This preserves useful model variety, puts capability where it compounds, and gives the system room to develop without turning a generic default into an unexamined bucket. A call that fits no category is a signal to refine the model, not to hide the distinction.