The right harness shape is RLM + GEPA + dspy.FLEX, with expensive frontier models driving cheaper models as workhorses.

I am logging this instead of building it. Nobody has monetized a harness, because the labs give them away with API access. Building it would only be something I was proud of.

A bet is only interesting where consensus disagrees. Model-cascade routing may already be less contrarian than it feels from the inside. If that is true, this entry should resolve as uninteresting rather than as a secret I sat on.

The call still stands: expensive models for judgment, cheap models for volume, GEPA for prompt evolution, FLEX for the compile, RLM for the window.