Self-Improving Engineering Program
Turn language-model behavior into an experimental variable: compiled programs, independent metrics and promotion rules that justify replacing what already works.
Problem
Prompt wording is craft; there is no disciplined way to say what may change, what evidence guides the change, or what justifies a replacement.
Outcome
A program whose behavior can be varied, measured and compared — with optimization kept separate from claims of improvement.
Research base
DSPy From First Principles
This solution is the applied destination of DSPy From First Principles: moving from handwritten prompts to language-model programs with contracts, datasets, metrics, optimizers, tool boundaries, context strategies, experimental firewalls and promotion policies.
The book’s empirical question — how optimization changes when correctness moves from a proxy for judgment to evidence produced by the environment — becomes the acceptance criteria here. This page will grow into the inspectable implementation once the experimental work lands.