You already know: expected value, variance and decision-making with probabilities (D2); strategy and choosing problems (Technical leadership); the system shapes that make changes cheap or expensive (System design thinking); cost models for serving and training (D12, D17).
Why this matters for Atlas. "Judgment" is what people call good decisions made with incomplete information, and it can be learned. It rests on a handful of habits: classify how reversible a decision is and match the effort to it; think in expected value and downside, not best case; know when more information is worth buying; compare total cost of ownership, not sticker price; treat technical debt as a loan with interest; and set kill criteria before a project starts. This lesson makes each habit concrete enough to apply in a design review.
1. Reversibility: one-way and two-way doors
The question. How much analysis does a decision deserve?
Start from nothing. Ask how expensive it is to undo. Two-way doors (feature flags, prompt changes, which model the gateway routes to, an internal API with one caller) can be reversed cheaply: decide fast, measure, adjust. One-way doors (a public API others build on, a data model that every service adopts, a multi-year vendor contract, deleting data, a launch that sets user expectations) deserve deliberate analysis, written alternatives and wider review. The common failure is treating every decision as a one-way door, which slows everything, or treating a one-way door casually. Part of good design is turning one-way doors into two-way doors: put the provider behind a gateway, the schema behind an API, the launch behind a flag.