CAST & ROUTING

The doctrine in one breath: start each role on the least expensive tier that is usually good enough, then escalate on evidence. A model that was best-in-class a few weeks ago is still excellent for first-pass planning, critique, and implementation. Don't burn frontier models on work a standard model can do — and don't cheap out when critique, architecture, data integrity, or product judgment is actually failing. Both halves are the rule; most routing mistakes are someone honoring only one of them.

The five-tier ladder

Routing is provider-neutral: a role names a tier, and a runtime adapter resolves the tier to whatever model currently earns it. Tiers, not model names, are the stable vocabulary — models rotate under the ladder without the doctrine moving.

Who runs on what

The writers' room plans, critiques, and decomposes — it writes no code:

ANALIZER 2000    requirements + scope        standard
THE ARCHITECT    systems design              strong    highest leverage in the room
THE CHALLENGER   independent critique        strong    always fresh-context
THE CLERIC       design briefs + review      standard
THE SPELLWRIGHT  plan → prompts              standard  translation, not invention
THE COUNSELOR    voice + audience            standard
THE SCRIBE       long-form drafting          standard  escalates to strong for drafts

The build party writes the code, one vetted prompt at a time, each prompt tagged to exactly one coder:

THE MAGE         frontend + design impl      strong
THE SYSTEMSMITH  backend, data, the contract strong
THE MECHANIC     builds, deploys, infra      standard  escalates for prod or irreversible ops

And the ministry and the shop — the watchers, binders, and runners:

THE WITNESS      cross-run briefing          standard  read-only
THE COUPLER      cross-coder phase-lock      standard
THE NOTARY       external cold attestation   strong    advisory
THE DELEGATE     automated checkpoint gate   strong    flow-and-gating only
TALLY            thorough odd jobs           standard  capped — never higher
SCOOT            one-breath fetches          cheap     locked — never higher

The two droids are the floor of the ladder, and the floor is load-bearing: because Scoot exists, a one-line lookup never falls through to whatever expensive session happened to be open. The odd job goes to the odd-job rate.

Fresh context is not strength

The routing table tracks two axes, and conflating them is a classic failure. Model strength is how capable the mind is; fresh context is what it has been allowed to see. The Challenger may run mid-ladder for a first pass, but it must arrive cold — no shared session, no accumulated sympathy for the work under review. A strong reviewer marinated in the builder's context is worth less than a modest one reading blind. When a runtime cannot guarantee coldness, the review is logged as warm rather than passed off as independent — the record keeps the distinction even when the infrastructure can't.

Even the Conductor economizes

The orchestrator itself obeys the doctrine. Picking a workflow, spawning specialists from the template, mechanical adjudication — blocker routes a fix, warning logs and proceeds — none of that needs the top of the ladder. The Conductor escalates its own session when the work turns into judgment: critic findings that need a real call, two specialists contradicting each other with no obvious winner, a second critique loop on the same phase, any checkpoint — or the moment it catches itself drafting spec content inline, which is the surest sign the cheap seat has been outgrown.

The escalation rule

One rung at a time, on evidence: when a handoff comes back thin, contradictory, or blind to obvious edge cases after one routed fix, the role moves up a tier — and the move is logged, so the run's record shows not just what was decided but what it cost to decide it. Workflows may name a tier as documentation; the resolved routing always wins when they differ.