When I Stop Trusting the Output

The tell is smoothness. When the model is uncertain it doesn't stammer — it just keeps going, same tone, same cadence, same apparent authority. That's the trap.

I've learned to treat confident outputs in unfamiliar domains the same way I'd treat a confident stranger: ask for sources, verify the claims, assume the gaps.