Collected with this tool against https://api.deepseek.com/v1, model id
deepseek-flash (API name: DeepSeek-V4.1-Flash), standard preset (8 cells x 25
samples), protocol one-token/v1, reasoning disabled via the openai-effort
adapter. 200/200 valid, 0 errors.
Unlike the bundled Zenodo samples, this reference is protocol-matched to the
package battery, so compare()/verify() against it is a direct comparison.
