Decoded #3: The dial that isn't a safety dial

Ask an AI the same question twice and you might get two different answers. Most people figure there must be a setting to fix that. Turn the randomness off and the AI becomes reliable.

The setting exists, and it's called temperature. Temperature controls how the model picks its next word. Low temperature means always pick the most likely word. Higher temperature means sometimes roll the dice. Intuition says low for accuracy, high for creativity. Turn it to zero for clinical work and you're safe.

First surprise: temperature zero is not deterministic. Different runs can still produce different outputs.

Second surprise: even if temperature zero were perfectly deterministic, it wouldn't make the answer correct. A confidently wrong answer at temperature zero is still wrong. Temperature controls variance. It does not control truth.

We run our router cold because we want consistent classification. We run our narrative layer warmer because explanation benefits from variation. Neither dial is what makes the output clinically correct. That comes from the deterministic logic that owns the decision before the AI ever opens its mouth.

A useful frame: temperature is a UX choice, not a safety control. Determinism lives in your architecture. Not in a setting.