The dial that isn’t connected

A side finding that cost a day and taught more than some full campaigns. I sent sampling parameters (temperature, top_p) to a provider bridge. It accepted them without error and echoed them back on every call. An A/B dispersion test showed they were never applied: response diversity at temperature 0.2 and at 1.1 was statistically indistinguishable, and every response opened with the same phrasing.

The rule this bought: a configuration echo proves nothing. Test the behavior, not the acknowledgment. My campaigns now run a dispersion preflight through the exact client that will make the real calls, on the same day, and the run does not launch unless low and high temperature are measurably different.

A bonus finding from the platform that does honor the dial: temperature shows up as format drift before it shows up as different choices. Higher temperature responses drifted from the requested answer format at a visibly higher rate while still choosing the same answers. If your benchmark uses strict parsing, you are quietly penalizing the higher temperature arm before you measure anything. I now pre-register that correction.