Replay of results/efficiency/clean-local-2026-10-29 at 2026-10-29T18:00:32.429Z on a local llama-server router Every language-model call replayed on a warm model (grouped by model, the first call of a model untimed), generating its recorded output length; code, tools and decision calls kept as recorded. case runs 42 success 100% (42/42; the recorded run, minus runs where arch-function-3b answered a delta call differently) latency p50 1.61 s warm (2.22 s recorded, model swaps included) latency mean 1.67 s warm (2.08 s recorded) model calls (warm): qvikhr-3-4b: 60 calls, p50 763 ms, mean 810 ms arch-function-3b: 60 calls, p50 307 ms, mean 336 ms delta answers of arch-function-3b unlike the recorded ones: 0