A Gemini Deep Research Failure Mode: Refusal, Topic Drift, and Fabricated Charts

I recently ran the same long-form research prompt through four LLM products: ChatGPT Deep Research, Claude with web search, Perplexity Pro, and Gemini Deep Research. Three of them handled it normally. Gemini did not. What followed was not a single bug, but a cascade of failures across multiple pipeline stages — each one revealing a different layer of state desynchronization in Gemini Deep Research. This post documents what I observed, what kinds of failures those observations seem consistent w...