docs/benchmarks.md

Benchmarks

The offline benchmark compares the sum of recorded event durations with the weighted critical path of the compiled DAG. It makes no network or model calls.

This is a trace-model scheduling estimate. It can show available parallelism, but it does not measure provider latency, rate limits, model planning overhead, caches, serialization, or distributed infrastructure. Do not present the ratio as a production speedup without an independent end-to-end measurement on real workloads.