Stephen Deslate

Engineer and applied researcher. I started this project to answer a question most of the industry treats as settled: can AI-generated code be systematically reliable?

So far, the research has produced 93 controlled trials across four experiments, tracking 218 failure modes and iterating until convergence. Claude is the primary engineering tool — it builds the code, designs the scoring frameworks, and implements the methodology. My role is defining what to test, evaluating the results, and deciding what to do next. The result so far is Convergence Engineering Development (CED) — a specification-first methodology tested across 10 full-stack layers, 5 progressive phases, and 40 ISO/IEC 25010 quality dimensions. The most recent experiment surfaced a hard question about Goodhart's Law — when the scorer and the code co-evolve, does convergence prove quality?

The research data is public — every trial, every score, every convergence curve. The full methodology is available on request. This site documents what I ship, how it's validated, and what the data shows. The program is ongoing.