The engineering discipline chain built specifically for probabilistic AI systems: requirements stated as acceptable tolerances, four-dimension design reviews, model-card documentation, vendor/SLA scrutiny, independent model/prompt/data version control, and the ethics of honest limitation-reporting.
An AI system that performs well in a demo, built by one engineer who knows exactly which inputs make it look good, is not yet an engineering deliverable — it is a demo with a due date. This module works through what turns it into one: requirements stated as acceptable error tolerances rather than "it works," since a probabilistic system's output has no binary correct/incorrect line to test against, and design reviews that examine the data pipeline, model choice, prompt design, and evaluation plan as four genuinely separate dimensions rather than one generic code review.
The module closes with the documentation and governance discipline unique to AI systems: model cards that record training data, known limitations, and intended use; critical evaluation of vendor and API claims and SLA terms; independent version tracking and rollback across model, prompt, and data/index versions (Module 20 supplies the actual templates); and the professional obligation this program treats as inseparable from the technical work — communicating a system's real limitations honestly to the non-technical stakeholders who rely on it, in support of the human-oversight principle from Module 14.