Karsten Eckhardt Data & AI systems · Teams A field guide should say how it was made. This one is unusual: it was built by the same kind of machinery it describes.
This site is the output of an engineering harness — the one the casework describes. A generator proposes a design round, a blind evaluator scores the rendered result against a written rubric it cannot argue with, and a commit gate refuses the work until a scored round exists. No change ships on the author’s say-so; it ships because it earned a pass. The site is not a claim that the method works — it is an instance of it.
Read the casework: Making AI agents prove their work →The discipline is only worth anything if it can catch its own mistakes. Twice it did: a rubric that scored a round 48/50 while outside readers saw “a high-end contractor, not a leader” — rewritten and re-scored on the identical pixels, it reproduced the human verdict at 29/50 — and a judge that handed out a clean 50/50 nobody believed, which a second blind judge scored 39/50 on the identical frozen evidence. The version you’re reading is a later round that had to beat the ones before it.