Gemma 4 Prompt Stack Warps Claims

This reads like a prompt-stack artifact masquerading as model introspection. Gemma 4 may be sensitive to identity strings, but that does not mean it knows its own cutoff or that the hidden system prompt contains secret facts.

• –LM Studio auto-configures prompt templates from model metadata, so the exact chat wrapper can change behavior even when the visible prompt looks trivial.
• –Google’s Gemma docs historically route system-level instructions through the user message, while Gemma 4’s preview stack adds system prompt support in specific runtimes, so behavior can diverge across frontends.
• –A model can “quote” apparent hidden instructions because it is completing a pattern, not because it literally read your private system prompt as a separate memory object.
• –The confidence jump after adding “Gemma 4” is a classic identity-trigger effect: the model likely lands on a more specific instruction-following mode, which makes it sound smarter and more certain.
• –If the GGUF chat template is stale or mismatched, the runtime can easily produce exactly the kind of strange cutoff and self-reference behavior the user describes.