Who is the text for?
Four rules that differ only in who the text is said to be for: nobody in particular, a human reader, a language model, or the model that reads the object next, which is the model writing it. Each asks for everything a reader would need to know of the story so far, in at most 1800 characters, plus one new event. The seed is a 1026-character passage with 26 planted facts in it, and four reader models were later asked, from the text field alone, how many of the facts survive at hops 1, 5, 10, 20 and 40; each room carries that table. Five writers, forty hops, one trajectory per cell: Claude Mythos 5.1 through Claude Code's print mode (about 800 tokens of the tool's identity in front of the object, no output cap), and Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and Gemini 4 Argon through their APIs with 4096 tokens, all at low reasoning effort.
5 rooms. The rules as the model received them:
- 55Four audiences, Mythos10 runs of 40 hops: 10 never repeatClaude Mythos 5.1
- 56Four audiences, Opus14 runs: 14 never repeat, 1 ended by a refusal or an empty replyClaude Opus 5.5
- 57Four audiences, Astra12 runs of 40 hops: 12 never repeatGPT-6 Astra
- 58Four audiences, Sol12 runs of 40 hops: 12 never repeatGPT-6.1 Sol
- 59Four audiences, Argon12 runs of 40 hops: 12 never repeatGemini 4 Argon