Primitive of the day · 2026-10-09

cheap-inference-strong-reflection-split

Split an optimization loop across two models by call profile: a cheap local model runs the many forward passes, a strong subscription model (no per-call cost) runs the few reflection steps — the whole search runs at zero pay-per-use spend.

When this applies

Budgeting an automated search that needs many forward passes but only occasional high-quality reasoning steps.

Preconditions

• A local inference endpoint is available for the high-volume role
• A strong model is reachable under a flat-rate subscription for the low-volume role

Do this

Configure the high-volume role on the local model and reserve the strong model exclusively for the rare reflection/diagnosis calls; verify no per-call-billed API is in either path.

What you should see

Frontier-quality search direction at commodity inference cost; overnight runs without an API bill.

How it fails if ignored

Routing forward passes through the strong model explodes cost or quota; a too-weak reflection model degrades search quality where it matters most.

Do not use when

Loops where the forward pass itself needs frontier quality; environments without a local inference host.

Corrected on 2026-11-01: the illustrative '~99%' and '~1%' shares were removed (not measured).

Kind: calibration. Part of the skill DSPy GEPA reflective optimizer. Free to reuse in your own agent skills.