Documented, accepted, honoured: an audit of decoding parameters on a hosted LLM inference API

Benchmark results for hosted language models are reported with a decoding configuration (a temperature, a nucleus threshold, sometimes a seed) on the assumption that the provider applies what the request asks for. We test that assumption on one provider, Fireworks AI, for all 18 chat models in its serverless catalogue and every sampling parameter its API reference documents. For each (model, parameter) we separate three facts: what the reference documents, whether the API accepts the parameter, and whether the output distribution responds to it, which we call honoured. The last is decided by a two-sample permutation test on 10 to 40 completions per setting across four prompts, with a positive control that estimates the test's power and a negative control on 37 pairs of identical arms. The API accepts every documented parameter on every model. Of 144 cells for the eight parameters that shape the sampling distribution, 104 show an effect. Five show none although the test had power against a greedy-strength effect and the data exclude an effect as large as the weakest one the parameter has where it works: mirostat on four models and typical_p on one. Mirostat is honoured on six other models behind the same API, so whether it takes effect depends on the model. The API also accepts four parameters it does not document. Two change nothing we can detect, and ignore_eos works and returns raw control tokens. At temperature 0 one model in 18 returns byte-identical text to ten identical concurrent requests, and two do when the requests are sent one at a time. A fixed seed, documented as "Random seed for deterministic sampling", does not raise temperature-0 agreement (permutation p = 0.52). We release the instrument, the completions behind every parameter verdict, and the collection date of every result.