efficient-frontier

config/model-routing.json is source of truth; never copy subjective scores into prompts.

1. Choose the best qualified primary owner/runtime.
2. Default GPT-5.6 Sol xhigh for UI, code, plans, review, computer use.
3. Use max only when context-ablation evidence or user selection supports it.
4. Terra/Luna require versioned behavioral eval promotion for product code/review.
5. Fable/Opus may own when qualified; the user explicitly authorizes a different-family pass.
6. ultra is multi-agent and needs explicit delegation or /swarm. Pro mode, persisted reasoning, programmatic tools, explicit cache are API-only unless exposed.

One owner implements. Without delegation, run useful lanes inline. Authorized lanes each get one bounded objective, inputs, exclusions, evidence, stop. Coordinator retains architecture, priority, risk, synthesis, acceptance.

Capacity

Use explicit /stay-within-limits host meter for Claude capacity; otherwise say unknown. Never infer capacity from tokens/cost. Capacity removes routes, never lowers quality.

Promotion

Before changing defaults run agent-evals/context-ablation/: vary one context group, hold tasks/scoring constant, and prefer lower cost only among quality-equivalent results. Record winner in routing config.