ADR 0001 — Capability tiers, not model names

Status: Accepted

Context

States differ in how much model capability they need: a classifier is cheap, a final synthesis is not. We must let authors express that without naming models (which would couple a .mk to one provider — see ADR 0003).

Decision

States route by a provider-neutral capability tier — fast / balanced / reasoning — set per machine (default_tier) and overridable per state (tier). The runtime holds a tier → (provider, model) map in its config. A tier applies to both generation and gate-judging. The reference interpreter is tier-following by default (a reasoning state's gates are judged by the reasoning model); a runtime MAY use a cheaper model for judging as an optimization, exposed here as the opt-in judge: config override — never the silent default (see the 0.5.2 remediation).

Consequences

• The same .mk runs on any backend by swapping the config's tier map.
• Cost control is a first-class, readable property of the machine (see the speculative-cascade pattern).
• The tier vocabulary is deliberately small (three levels); finer control lives in the runtime config params, not the .mk.