The spec had the architecture right — the grilling found the soul

Three /grill-me rounds before writing a single line of code, and the third one changed everything. The spec for a "Dynamic Friction Graph & Context Translator" had the architecture right — dependency graphs, cost vectors, domain types, output modes. Then in the second session I said "something feels like it's missing from the plan, and whatever that something is, isn't that I want to add complexity." Not a technical gap. Something about the premise.

The grilling surfaced it. The original spec treated internal barriers — executive dysfunction, emotional drain, cognitive overload — as secondary to external ones. Five domain types, all external: physical, financial, social debt, administrative, time horizon. But for the people this tool is actually for, the internal barriers are the primary friction. Someone can't "just apply for jobs" not because they lack WiFi, but because depression makes getting out of bed cost more energy than they have, and the shame of asking parents for money makes the depression worse, and the spine pain from sitting means 30 minutes is the max before everything locks up.

So executive_function became a first-class domain type alongside financial and physical. Eight domains instead of five. Permanence calibrated differently — "can't initiate a task" scores 0.85, categorically different from "waiting for a repair appointment" at 0.3. The Boundary Output generates physical-analogy energy equivalences: "The executive dysfunction required to initiate this phone call costs the equivalent energy of climbing four flights of stairs with groceries." Making the invisible structurally undeniable.