Engineering notes: Service-level Mixture-of-Experts + test-verified publishing in a self-improvement loop [R]

About this article

We've been running an autonomous problem solver with a closed reasoning loop: observe → hypothesize → write Python transform(grid) → run against training examples → reflect on failures → adapt strategy. It uses retrieval over a "sensei wiki". THese are task-specific notes that get prepended as prompt preambles when the solver attempts the same task again. Standard enough. The notes were hand-written. Last week we added wiki notes for one of the stuck tasks asserting the rule was "output [[2]]...