Anyone else noticed the same LLM give a completely different answer to the identical prompt, twice?

Ran into something this week that's been bugging me: gave a local model the exact same prompt, replayed it twice with nothing else different, and got a correct answer once and a wrong one the second time. Not a hard edge case either — a pretty basic task it clearly CAN do, it just didn't do it reliably both times.

Made me rethink something: a lot of what gets blamed on "the model isn't good enough" might actually be a consistency problem, not a capability ceiling. Those are very different things to fix — capability needs a bigger/better model, consistency might just need lower temperature, better prompting, or accepting you need a verification step regardless of model size.

Curious if others building with LLMs (local or cloud, doesn't matter) have run into this — do you design around inconsistency (retries, verification passes, structured output constraints), or has this mostly not been an issue for what you're building?