Gemma 4 26B fabricated an entire code audit. I have the forensic evidence from the database.

I run Gemma 4 26B-A4B locally via Ollama as part of a custom self-hosted AI platform. The platform s every model interaction in SQLite, including three columns most people never look at: content (the visible response), thinking (the model's chain-of-thought), and tool_events (every tool call and its result, with full input/output). I asked Gemma to audit a 2,045-line Python trading script. She had access to read_file and bash tools. Here's what actually happened.