Outline sketch
Prompt
Audience: platform / AI eng.
Steps:
1) State: user request, tools, budget, risk tags.
2) Choice difficulty/risk or Score complexity; Noul needs_strong_model.
3) Code routes cheap vs frontier vs abstain/human.
Expected effect: spend frontier tokens only when justified.
Needs access to: typesafe-sdkllm-gateway
Who it's for
Platform / AI eng owning multi-model stacks
Steps / how it's set up
Sources: LangChain · Building a harness with Jev · confidence-routing · awesome-jev-usecases · Harness.
Treat confidence as a second axis beside category: cheap default model when confidence is high; escalate to frontier / human when low. Community harness numbers are author-reported.
Sources (wave-2 deepen)
Expected effect
Harness pattern: System One decides when System Two is worth paying for.
Unofficial outline for learning. Paraphrased from public docs and tutorials — not a production recipe. Review sources before you automate anything.