understand-until-simplelisted

# understand-until-simple The pre-code gate. Operationalizes §0 (SOLID) + §0.1 (decompose to the implementation level) into a checklist you clear BEFORE writing implementation code. Born from the DSv4 concurrency case (2026-06-14): I kept calling batched decode "very hard", proposing it as a multi-day lever on hypothesis. ckl's correction — *"the way you're talking means you don't understand it at all — 'so hard, so tough'; we start work only once you've figured it out and consider it simple"* — was right. Reading the code + measuring collapsed it to: **it's one `for r in 0..n` loop (`dsv4.rs:1872`); batch it → ~2×; pieces exist; MoE won't cap it.** Simple. The "hard" was just un-decomposed. --- ## The law **"Hard" is a confession that you haven't decomposed it yet.** Difficulty is not a property of the problem; it is the distance between you and the atomic mechanism. Close that distance until the problem is simple, then the code writes itself. A plan that still sounds hard is a plan you don't understand. **No code until you can state the fix in one simple sentence with a measured number behind it.** ("Batch the per-row attention loop → ~2× aggregate, MoE stays sub-linear" — that sentence is the license to start.) --- ## The gate (clear every step before coding) ### 1. Catch the hand-wave If your description contains "hard / tough / complex / tricky / multi-day / it's the lever" WITHOUT a line number and a measured number, STOP. You are at hypothesis, not understandi