Benchmarks

The method, local replication path, current calibration, results, and limits.

All benchmarking is local. There is no public Auwai benchmark endpoint.

Contents

• local-replication.md: Exact local commands, mock server, environment variables, and troubleshooting
• methodology.md: The protocol, controls, and gates
• results.md: The current measured suite and release evidence
• calibration.md: Local model throughput and turn estimates
• quality/README.md: Fixed rubrics and the answer scorer

Reading order

Read the local replication page first if you want to run a benchmark. Read the methodology and results pages if you want to interpret one.

Release rule

A benchmark result is releasable only when it was produced by a command in local-replication.md against an explicit local endpoint. Historical hosted-proxy runs are evidence of past behavior, not a runnable default and not a release gate.