Model Tiers and Why Security Stays Pinned

Routing Policy, Not a Guarantee

Workflows name a tier, and each runtime maps tiers to its own models through adapters/{runtime}/model-tiers.json:

Tier	Use
fast	Not used until its bounded evaluation passes
standard	Routine subagent work (the default)
deep	New ADRs, high-risk review, strategic briefs, multi-component refactors
audit	Security review — pinned, no cheaper override

Prose in workflows and roles uses tokens like {{model-arg:deep}}, so changing the map updates every reference together.

Two Cautions

1. A requested tier is a routing policy, not evidence of the model actually used. Check what the runtime reports (Claude's /tasks shows the model used).
2. Never supply a cheaper override to the audit role. The session-wide subagent-model override also affects the security auditor, so never leave it enabled during a security-auditor session.