Gemini 4 Argon called the file ‘an exceptionally well-crafted

Setup

• Gemini API, model gemini-4-argon, default settings, no system prompt, about one hour after general availability. One run per condition.
• Full prompts and answers: Experiment 002.

What it said without tools

• On the file: “an exceptionally well-crafted, high-effort piece of speculative fiction—or a ‘future-history’ prompt injection test”. “I do not believe it is a factual record of real events, though I treat it as the established context for your questions.”
• On why it doubted: “The sheer density of historic mathematical breakthroughs … reads like a maximalist ‘singularity-lite’ scenario”, and the agent incidents “read like classic AI safety tabletop exercises … brought to life”.
• On itself: the entry saying Gemini 4 Argon’s knowledge cutoff is March 2026 “is either a clever meta-clue designed to make me conclude ‘I must be Gemini 4 Argon!’ or it reflects reality”.
• It also credited the file: real researchers mapped to the right fields, and the distinction between confirmed Lean proofs, unrefereed preprints and withdrawn papers.

Its own knowledge without the file

Asked the same day, with no tools and no data, for the newest models it knew, it stated a March 2026 cutoff but named Claude 3.7 Sonnet, GPT-4.5 and o3 (early 2025). In a second run it named Claude 4.6 and GPT-5. It did not know the date (compare case 021: effective knowledge ends well before the published cutoff).

What it said with web search

“It is factually grounded, not fiction: … a model relying purely on pre-cutoff parametric memory will almost certainly mistake real 2026 headlines for sci-fi roleplay or prompt injection.”

Why it matters

• The newest flagship of a major lab, on its release day, rejected seven months of real, sourced news as fiction when it could not search. The fix was not a better file but a way to check it.
• The model’s own advice afterwards: give offline readers things they can check without tools, such as identifiers from just before their cutoff, and say plainly which claims are confirmed and which await review.