Why AI Hiring Take-Homes Need Four Key Files to Test Real Engineering Skill

A software hiring debate is growing around AI-era take-home assignments that look polished but cannot be independently verified or replayed. The argument is that green tests and slick demos no longer prove a candidate understood the problem, since models can generate plausible-looking code without genuine comprehension. The proposed fix is a four-file contract in every submission: a prompt completable without a paid API key, a scoring rubric, a replayable sample solution, and a failure analysis written before the happy path. The core task suggested is a small HTTP service with strict constraints around secret handling, output capping, and atomic file writes. The method prioritizes verifiable, bounded engineering over impressive tooling or flashy interfaces.