crew — The Expensive Model Architects, the Free Ones Type

crew — The Expensive Model Architects, the Free Ones Type An architecture diagram generated by Archify. Task Contract · signature + cases · Architecture component Task Contract signature + cases RAM Sizing · weights + KV cache · local worker tier — zero paid tokens RAM Sizing weights + KV cache Model Catalog · attention geometry · local worker tier — zero paid tokens Model Catalog attention geometry Architect · Claude — expensive, tiny · Architecture component Architect Claude — expensive, tiny crew CLI · zero npm deps · Architecture component crew CLI zero npm deps Contract Prompt · closed, numbered · Architecture component Contract Prompt closed, numbered Worker Dispatch · tiered, parallel · local worker tier — zero paid tokens Worker Dispatch tiered, parallel Local Models · Ollama · any OpenAI API · local worker tier — zero paid tokens Local Models Ollama · any OpenAI API crew validate · spends nothing · deterministic — no model judges another model crew validate spends nothing Gates · typecheck · lint · tests · deterministic — no model judges another model Gates typecheck · lint · tests Patch Guard · write whitelist · deterministic — no model judges another model Patch Guard write whitelist crew report · compressed summary · Architecture component crew report compressed summary closed contract scaffold costs nothing dispatch one task HTTP bytes per token derived num_ctx diff whitelisted only fail → higher tier pass local worker tier — zero paid tokens deterministic — no model judges another model Legend Frontend Backend Database Cloud Security External

Why the contract is closed

  • • Hand a whole feature to a 7B and you get garbage you then have to read
  • • Hand it an exact signature, cases numbered one by one, a write whitelist and a verify command
  • • Now a test can accept or reject it with no human in the loop

Why it escalates instead of retrying

  • • Measured on an M5 with 16 GB: the 7B lands about half of multi-case contracts
  • • It does not converge on feedback — fail once and it usually fails three times
  • • The 14B got it first try on the task the 7B failed three times, so failure goes up a tier

What decides

  • • Typecheck, lint and tests — exact, free, and the only thing that can accept the work
  • • A reviewer model's 'approve' never substitutes for a gate
  • • Zero npm dependencies; CI runs macOS, Linux and Windows on Node 18, 20 and 22