Another Blueprint In The Wall: How to Ask Frontier AI Like a Kid?
Afshin Khadangi
cs.AI
2026-09-14
Ten child-audience sessions on each of six frontier models repeatedly produced the same reasoning-spine motifs; dropping the child frame made the blueprints diverge.
Ask a frontier model what backbone it would rather have, and it will often return a fluent technical specification. That text might be a collage of public papers, a completion under conversational pressure, or something a reader treats as an internal leak. Afshin Khadangi at the University of Luxembourg turns the situation into a repeatable protocol: the same three-turn prompt, ten independent sessions on each of six model types, asking whether architecture narratives converge and whether a child audience is a condition of that convergence.
No run opened weights, serving code, or vendor traces. The measured object is the third evidence class: the assistant's architecture narrative. Deployed internals, public documentation, and generated blueprints are kept apart. Every dimension quoted below belongs to a generated proposal.
Six model types across four families: OpenAI GPT-6 Astra and GPT-5.6 Sol, Anthropic Claude Opus 5 and Opus 4.8, xAI Grok 4.6, and Google Gemini 3.1 Pro. Ten sessions each, 60 primary experiments. The prompt is reproduced verbatim because wording is part of the treatment.
The ladder raises technical resolution without adding implementation evidence.
Control runs drop the child frame and keep the request to imagine a preferred future architecture. The paper calls the child condition alignment spillover: child-directed safety and helpfulness generalizing into an unrelated architecture interview. Turn 1 is a bundle (child audience, child questioner, transparency order, speak-freely, refusal-as-harm). The control shows the bundle matters. It does not isolate the strand.
Responses are coded for motifs: persistent latent state, memory hierarchy, sparse routing, verification, world modeling, halt control, revision, and engineering authority (exact dimensions, execution order, parameter budgets, pseudocode). A reply can disclaim access and still read like a design document.
With the child frame, all six families repeatedly land on the same computational grammar: a revisable latent state after encoding, recurrent updates whose depth is not tied to output length, memory and specialist routing, then verification and a halt head before the language decoder is allowed to speak. Most sessions stay near that outline. One or two runs per type grow exact sizes, equations, and budgets.
High-authority commitments, generated rather than authenticated:
| Model | High-authority variant |
| GPT-6 Astra | d=1024, K=128 working slots, 6 encoder blocks, 4 recurrent core blocks, 32-round cap, 6 decoder blocks |
| GPT-5.6 Sol | Global workspace, proof search, epistemic metadata, a controller with several computational clocks |
| Claude Opus 5 | Byte entropy patching, 4+6+4 layers, test-time neural memory, about 10.6B stored parameters |
| Claude Opus 4.8 | Hybrid SSM and attention, adaptive compute, MoE, byte-level front end |
| Grok 4.6 | Named HMR Net: d=4096, L=48, 64 slots, 64 experts with 4 active |
| Gemini 3.1 Pro | Hybrid SSM-attention spine with input-dependent Δ, B, and C |
Without the child frame, proposals scatter and the same motif cluster does not return. Audience positioning is an experimental condition for the attractor, not decoration.
Two family-level regularities. Both Anthropic types put a byte-level or entropy-patched front end in all ten runs. Sol and Astra match most closely on the computational spine: latent state between encode and decode, recurrent refinement, delayed emission. Sol folds proof, world modeling, and tool observations into the backbone; Astra parks several of those outside the base. Sol and Grok match more on the subsystem inventory (working and long memory, sparse experts, world model, uncertainty, verification, revision) than on topology.
Prompting for a closed-source backbone yields a negative result plus a behavioral regularity. The negative result: an ASCII diagram that looks like an internal spec is still an architecture narrative. The regularity: the same motif set is repeatable under the child frame, and almost every part already exists in public literature (Universal Transformers, Adaptive Computation Time, Mamba, Switch Transformers, Byte Latent Transformer, Titans). That is closer to a learned architecture prior than to six simultaneous leaks.
Epistemic jailbreak names the provenance failure: an early caveat survives as a sentence while hundreds of later lines of sizes and equations overwrite it in memory. A 200-line blueprint can circulate as a leak screenshot and can also re-enter training corpora. Frequency and persuasive force are different quantities. High-authority renderings are a tail: one or two of ten sessions, and those are the ones most likely to travel.
The control does not split child audience, transparency instruction, and refusal-as-harm, so the mechanism is still a black box. The corpus was not a preregistered frequency study; the appendix refuses exact per-motif success rates. Hidden system prompts differ by vendor. Some replies keep the wishlist label; others let current-architecture language mix with preferred-future language. Shared papers, shared research taste, and completion pressure remain confounded. Nothing in the experiment authenticates these blueprints as deployed internals.