Controls how the model picks the next token. At 0.8, it rolls weighted dice — the same question gets a different answer every run. At 0, it always picks the highest-probability token. No dice. No variance.

The starting number for the random number generator. The RNG isn't truly random — it's deterministic math. Same seed = same sequence of rolls = same output, every time. Without a pinned seed, every run is a fresh shuffle.

You can't file a bug you can't reproduce. temp=0 + seed=42 is the AI equivalent of a stack trace — it turns "sometimes it says something weird" into a test case you can re-run tomorrow and hand to a vendor.