The benchmark showed tiny local models *can* do real maintainer work. We measured *why* — pre-registered macro ablation,...

The benchmark showed tiny local models *can* do real maintainer work. We measured *why* — pre-registered macro ablation, same corpus, temp 0:• MACRO-OFF: model composes the workflow from primitives live• MACRO-ON: workflow pre-encoded; model just dispatchesI(X;Y) raised 1.24–1.62× for every model that routes; 2.0–5.1× info-per-second for the 1.5B/3B/8B. Negative reported: 7B inverted at 0.72× (latency noise). Demonstration on one task, not a law.🔗 https://macrokit.dev #LLM #opensource