BLITZMETRICS · DEFINITIVE / SOP
Model Judgment & Delegation
Route each task to the cheapest model tier that clears the quality bar. Delegate the bulk down. Escalate the hard calls up. Let the agent decide, by judgment, every time.
The one idea: Stop asking “which model is best?” Ask, for each piece of work, what is the cheapest tier that still clears the quality bar — and route it there. Cheap for the grunt work, the ceiling for the judgment calls. Let the agent make that call itself, task by task.
This is the affordable half of boil the ocean: full coverage only stays sustainable when each run is cheap.
Where this came from
Two ideas a month apart that turn out to be mirror images.
Delegate down — the Claude Code team, and Simon Willison’s write-up (“Fable’s judgement,” July 3, 2026): it’s better to let a top model use its own judgment about how to work — including which lower-power model to run a subtask in — than to hand it rigid rules. His exact instruction: “For all tasks, use your judgment to decide an appropriate lower-power model and run that in a subagent.”
Escalate up — Anthropic’s Advisor Strategy (April 9, 2026): the inverse. A cheap executor (Sonnet or Haiku) drives the whole task and consults a top model (Opus) only when stuck. Haiku’s hard-research score more than doubled — 19.7% to 41.2% — with an Opus advisor riding along, at a fraction of the cost of running Opus outright.
Willison’s example was coding. Ours isn’t — and it fits us better. Anthropic’s own advisor docs name research pipelines, data classification, and document extraction as the sweet spot. That is our Content Factory: audits, mention harvests, scoring, article waves, site builds.
The tier map — the whole ladder
| Tier | The work | Engine | Cost |
|---|---|---|---|
| 0 · mechanical | transcript pulls, REST publishing, PDFs, dedup, enrichment, file wrangling | plain scripts — no model | ~$0 |
| 1 · bulk / draft | mention harvesting, first-draft articles, boilerplate pages, restyles, bulk social | Haiku / Sonnet 5 / GLM in subagents | cheap, flat |
| 2 · judgment | honest scoring, the “earned but illegible” insight, disambiguation, strategy, writing in the subject’s voice | Opus 4.8 → Fable 5 for crown jewels | metered, tiny |
| 3 · computer-use | logged-in publishing, firewall-blocked writes, browser fixes | Cowork / Claude in Chrome | included |
Most of the token volume lives in tiers 0 and 1 and does not need a frontier model. Only the thin tier-2 slice — the honest number, the real insight, the actual voice — earns the ceiling.
The decision, in one breath
Does getting this wrong require taste, honesty, or a person’s real voice?
No → push it down to a cheaper model in a subagent.
Yes → keep it on the top model, and let the cheap tiers escalate up when they stall.
How to actually do it
In Cowork or Claude Code. The agent spawns a subagent and sets the model per call — Haiku, Sonnet, Opus, Fable. This works in Cowork through the Agent tool’s model override; it is not a coding-only trick, and it applies to research, writing, scoring, and site builds exactly as it applies to code.
The one-line standing rule — paste it into a Project, a memory file, or a skill pack:
“For any substantial build, research, or writing task, use your own judgment to route it to the cheapest model tier that clears the quality bar. Delegate bulk and first-draft work down to a smaller model in a subagent; reserve the top model for scoring, strategy, disambiguation, and voice; escalate a stuck cheap tier up. Verify subagent output before shipping. The bar stays ‘holy shit, that’s done.'”
In an API harness you run on a schedule. Set the executor model per tier in config, and turn on the advisor tool so the cheap tier can consult the top model when stuck. Our reference build is the Audit Factory: scripts collect (no model), a cheap model drafts, the frontier model judges, and an opt-in escalate-up path lets the draft consult Opus when it’s stuck.
Worked examples
- Harvesting every good mention of a client → tier 1. Cheap models sweep the sources in parallel; the top model only ranks and interprets the shortlist.
- A 400–650-article wave for a topic cluster → tier-1 drafting, tier-2 editing. Drafts on a cheap model; the canonical hub and the voice pass on the top model.
- Scoring a personal brand honestly → tier 2, top model only — and put the honesty rules in code, not the prompt, so no tier can hand out a flattering number.
- Writing in the subject’s voice → tier 2, Fable. This is the one thing that reads as “AI” the moment you cheap out. Don’t.
The guardrails
- Verify before shipping. A cheap subagent that loops without review just automates its own mistakes.
- Honesty rules go in code, not prompts. Bands, gates, and clamps run identically on every tier.
- Never cheap out on voice or the final judgment. Those are the crown jewels.
- The advisor is Anthropic-backend-only. Keep the executor on Anthropic when using it — Haiku executor plus Opus advisor.
- One shared file spine. Every tier reads and writes the same files, so models stay swappable. Bridge through files, never one vendor’s memory.
Route by tier and you get roughly 3× the throughput at a fraction of the cost, with no quality loss where quality matters. Cheap where it’s mechanical, expensive where it’s judgment, complete always.
Read how we built it
Get the skill pack
Sources: Simon Willison, “Fable’s judgement” (Jul 3, 2026) · Anthropic, the Advisor Strategy (Apr 9, 2026) · Anthropic, Create custom subagents. Part of the BlitzMetrics Content Factory.