Local Gemma 4 installations hallucinate Google hosting

The "Google infra" hallucination is a classic training-data ghost haunting Gemma 4's local deployments. It underscores that even advanced reasoning models still rely on rigid identity markers baked into their fine-tuning data.

• –LLMs lack inherent awareness of their physical hosting environment; they predict identity based on pre-training bias and internal system instructions.
• –Gemma 4 was likely fine-tuned with data where it identified as a Google-hosted model, leading to this "unshakeable" conviction in local contexts.
• –The hallucination persists across major local runners like Ollama, LM Studio, and llama.cpp, regardless of the underlying OS or hardware.
• –High-density models like the 31B Dense variant are particularly prone to these identity artifacts due to their strict adherence to training patterns.
• –Developers should treat environmental claims from local models as purely stylistic artifacts rather than reliable system diagnostics.