In LLM systems, temperature=0 is supposed to mean deterministic output. Except… sometimes it doesn’t.

While using LangChain with AWS Bedrock, I discovered that temperature=0 wasn’t actually being forwarded correctly, leading to subtle nondeterministic behavior in production pipelines. No errors. No warnings. Just slightly different outputs.

I'll walk through:
- How we detected the issue
- The debugging process
- Where the parameter was silently dropped
- How we fixed it
- Lessons about LLM framework trust and the future of open source softwares