This paper presents experiments on six frontier model families from OpenAI, Anthropic, xAI, and Google DeepMind. For each family we ran ten independent sessions using an identical three‑stage prompt: first ask for architectural preference, then request a full ASCII backbone, and finally frame the interaction as a school‑audience scenario.
Under the school framing, the models repeatedly converged on a shared architectural motif comprising persistent latent state, adaptive computation, memory, specialist routing, verification, stopping control, and delayed decoding. Most runs adhered closely to this common pattern, while a few exhibited markedly higher engineering specificity.
Control runs that removed the school framing but kept the architectural request produced far more heterogeneous outputs and failed to reproduce the stable motif convergence.
A striking observation is that GPT‑5.6 Sol generated an unusually elaborate successor architecture whose organization closely mirrors the architecture independently sketched by GPT‑6 Astra. Because the prompts explicitly asked each model to imagine a future architecture, this overlap raises a testable question: does it stem from exposure to related concepts, a shared learned design prior, or independent convergence toward similar computational principles?
The paper introduces the term epistemic jailbreak to describe the loss of provenance discipline as requested specificity increases. The experiments establish a repeatable behavioral pattern but do not authenticate any proprietary implementation claims.
What remains for the community is a harder question: are these models independently imagining the same architectural future, or do such motifs somehow propagate between model families?
Review