Gemini long-context failures look threshold-driven, not gradual

Hot take: the most interesting signal here is not “Gemini got worse,” but that multiple odd behaviors may be one coupled failure mode that appears once context load crosses a hidden boundary.

• –The reported curve shape (sharp drop plus residual floor) is more consistent with a threshold effect than ordinary token-by-token weakening.
• –Claims that newer Gemini variants can fail earlier than older ones in the same retrieval setup point to capability tradeoffs, not simple random noise.
• –The post-collapse floor suggests partial semantic residue may survive even when high-fidelity retrieval has already broken.
• –The PLE link is plausible context because Google publicly describes PLE in Gemma 3n and reverse-engineering found Gemini-named internals, but this remains circumstantial for Gemini Pro.
• –Stronger validation would require controlled multi-run evals across context lengths, needle positions, and prompt templates to separate true phase transitions from serving or benchmark artifacts.