What I Learned Running The Profiling Pipeline On gpt-oss-120b

Two weeks ago I wrote about splitting the profiling pipeline into an open model, a commercial model and deterministic scripts, and last week I said that work is why I am confident enough to take APIs.io on-premise. What I have not done is write up the measurements themselves. This is that post: which local models I tried, what gpt-oss-120b actually did on real providers, how it compares with Claude on the same inputs, and the lesson underneath all of it, which is not about models.