Distributional thinking for AI quality

Lesson 1.4 builds on the previous three lessons. In 1.1, non-determinism was quantified: the same query run five times produces different outputs. In 1.2, the evaluation surface was mapped, covering every pipeline stage where failures could happen. In 1.3, actual failure categories were discovered from traces, bottom-up, with no predefined checklist. Now comes the question that matters for shipping. Is the evidence good enough to decide? A single success rate is not evidence. It is a point estimate from one sample. This lesson introduces thinking about quality as a distribution: error bars on every metric, breakdown by segment, decomposition of variance sources, and a decision framework that connects distributional evidence to ship-or-hold calls. This is the paradigm shift from "did it work?" to "how often does it work, under what conditions, and with what variance?"