Gemma 4 wrote three summaries in one response. The middle one was a self-disclaimer.

Gemma 4 E2B showed a repeatable three-part output pattern at a 2,048-token context window: summary, self-disclaimer, then revised summary.

The same prompts at a 32,768-token context window did not reproduce the behavior, even under control inputs.

This suggests the odd formatting was driven by context-window size, not by prompt damage or a broader truthfulness issue.

The finding matters for evaluating model calibration, reliability, and how output behavior shifts with context length.