Run modes and evidence levels

The three modes answer different questions. Choose the claim first, then the run.

Mode	Use it to answer	Model access	Publish as benchmark?
mock	Is the adapter wired correctly? What framework mechanics remain when responses are fixed?	Deterministic local server	No
codex	Does the end-to-end path work with real model responses on an existing subscription?	Codex sign-in	No
live	How do adapters compare against the same native provider and model?	OpenAI-compatible API key	Yes

Mock mode

Use mock mode for fast, repeatable development and for measurements that remain valid when the model output is held constant: prompt bytes, retry behavior, pause mechanics, and orchestration calls. A high pass rate proves plumbing, not intelligence.

Codex mode

Use Codex mode as a functional bridge when you have a ChatGPT/Codex subscription but no provider API key. It exercises real responses, but the translation layer means latency, token accounting, and cost are not comparable with native APIs. See the Codex bridge guide.

Live mode

Use live mode when you can hold the provider, model, tools, task, and scoring constant across adapters. These scorecards may support native latency, usage, cost, and answer-quality claims. Record the model and framework versions and repeat the run so later readers can judge freshness and stability.

Continue with Methodology for the full comparison contract.