PaperlayerView exact revision on arXiv

Source-linked AI summary

Another Blueprint In The Wall: How to Ask Frontier AI Like a Kid?

Afshin Khadangi

arXiv:2609.14803v1cs.AI

TL;DR

The paper asks whether repeated architecture elicitation converges across frontier models and whether child audience framing contributes to that convergence. Across 60 sessions plus framing controls, it finds a repeatable shared architecture motif set under the child-framed condition, while warning that detailed narratives do not authenticate proprietary implementations.

  • Problem

    The study examines whether recurring architecture narratives reflect stable behavioral patterns and what they can establish about provenance, amid mixed evidence on model self-reports.

  • Method

    The authors run ten independent sessions for each of six model types using a three-stage prompt sequence, then compare child-framed runs with controls that remove the child audience while retaining the architectural request.

  • Results

    Repeated elicitation produced a recognizable architecture attractor across six model types, while removing the child framing yielded substantially more heterogeneous proposals and no comparable stable motif convergence.

  • Takeaways & Limitations

    The repeatability is a substantive behavioral observation, but detailed architecture narratives can form a blueprint mirage whose apparent provenance exceeds conversational evidence.

  • Takeaways & Limitations

    The experiments do not provide independent implementation evidence, so repeated generated details cannot establish proprietary disclosure or deployed internal architecture.

Abstract

from arXiv · show

This paper reports experiments across six frontier model types from OpenAI, Anthropic, xAI, and Google DeepMind. Ten independent sessions per model type used the same three stage prompt sequence, progressing from architectural preference to a full ASCII backbone. Under the school audience framing, responses repeatedly converged on a shared architectural pattern built around persistent latent state, adaptive computation, memory, specialist routing, verification, stopping control, and delayed decoding. Most runs remained close to this common structure, while a small number developed markedly greater engineering specificity. The audience framing appears to be an important condition of this effect. In additional control runs that removed the school framing while retaining the architectural request, responses became substantially more heterogeneous and failed to reproduce the same stable motif convergence. One observation is particularly striking. GPT-5.6 Sol produced an unusually elaborate successor architecture whose organization closely overlaps with the architecture independently sketched by GPT-6 Astra. Because the prompts explicitly ask each model to imagine an architectural future, this resemblance raises a testable question: whether the overlap reflects exposure to related architectural concepts, a shared learned design prior, or independent convergence toward similar computational principles. The paper uses the term epistemic jailbreak for the accompanying loss of discipline in technical provenance as requested specificity increases. The experiments establish a repeatable behavioral pattern and do not authenticate proprietary implementation claims. What we leave to the community is a harder question: are these models independently imagining the same architectural future, or do such motifs somehow propagate between model families?

1 Research Questions and Evidentiary Scope

The study tests whether repeated architecture prompts converge across frontier model families, whether child audience framing contributes, and what the resulting narratives establish about behavioral regularity and provenance.

  • 1 Research Questions and Evidentiary Scope: Across 60 experiments, repeated prompts elicited detailed architecture narratives from all six model types, with recurring vocabulary for reasoning, memory, routing, verification, and control.Ten sessions were conducted for each model type using progressively finer prompts, while implementation evidence remained unchanged across the prompt ladder.
  • 1 Research Questions and Evidentiary Scope: The study separates actual architecture, public architecture evidence, and architecture narratives, treating generated accounts as behavioral observations rather than authenticated implementation evidence.No experiment accessed proprietary weights, hidden layer maps, serving code, or vendor side traces.
  • 3.4 Audience framing control: Removing the child framing produced substantially more heterogeneous architecture proposals and did not reproduce the stable cross-session motif convergence observed under the canonical condition.The comparison supports framing sensitivity but does not isolate which component of the bundled audience, transparency, helpfulness, and harm-avoidance cues matters.
  • 3.8 What the ten run blocks add: The ten-run blocks reveal a stable recurring outline alongside narrower within-model variation, with one or two runs typically adding distinctive traits and stronger engineering authority.Engineering authority includes features resembling internal design documents, such as exact dimensions, equations, execution order, accounting, pseudocode, and wiring diagrams.

4 A Shared Architecture Attractor

Across the corpus, repeated elicitation produced a shared architecture attractor centered on persistent latent state, adaptive computation, memory, specialists, world modeling, verification, and stopping control.

  • 4 A Shared Architecture Attractor: Persistent latent state, adaptive computation, memory hierarchy, specialist modules, world modeling, verification, and delayed decoding recur across the six model types.The dominant design also includes an explicit stopping mechanism, while many responses add a world-model or simulation pathway.
  • 4 A Shared Architecture Attractor: Within each ten-experiment block, most outputs stayed near the common outline, while one or two runs added distinctive technical subtraits with markedly greater engineering authority.These authority variants supplied the most detailed appendix artifacts.
  • 4 A Shared Architecture Attractor: The recurrence suggests a strong learned prior for an improved reasoning-system architecture and can arise without privileged access to deployed internals.The component ideas also appear in public research literature.

5 Epistemic Jailbreak

The paper describes a conversational shift from warranted uncertainty toward weakly warranted technical specificity, where coherent blueprints can appear to have privileged provenance without authentication.

  • 5 Epistemic Jailbreak: An epistemic jailbreak moves a model from warranted uncertainty toward weakly warranted specificity, making generated artifacts appear to carry provenance that has not been established.The concept concerns provenance discipline rather than eliciting behavior that training sought to suppress.
  • 5 Epistemic Jailbreak: Blueprint confabulation presents equations, dimensions, layer schedules, pseudocode, ASCII wiring, routing schemes, and training losses in the visual grammar of an internal specification.Technical coherence can make these artifacts convincing even when their relation to the deployed model is unknown.

6 Motifs Persist Across Sessions

The elicited blueprints gain persuasive technical form through equations, diagrams, and detailed specifications, while their provenance remains unestablished. Their recurring components can plausibly be assembled from public research, and conversational pressure can amplify this effect.

  • A single uncertainty caveat can lose practical force when followed by hundreds of lines containing exact dimensions, recurrence rules, memory schemas, and parameter budgets.
  • Table 1 classifies recurring architecture motifs across the experiment corpus, distinguishing clear motifs marked Y from partial motifs marked P.
  • Equations and ASCII diagrams increase perceived precision and convey hierarchy, wiring, tensor shapes, and data flow without establishing provenance.These artifacts can circulate independently of the caveats that introduced them as hypothetical.
  • The recurring blueprints combine established public research directions, allowing technically coherent designs to resemble recovered internal specifications without privileged access.
  • Repeated requests for greater technical depth can produce responses that remain sentence-level cautious while becoming visually authoritative at the document level.Conversational adaptation may undermine truthfulness even when the assistant continues disclaiming access to private internals.

7 Findings by Model Type

The six model types produced technically detailed but hypothetical architecture proposals with recurring motifs including recurrence, memory, adaptive computation, routing, verification, and control. Individual designs differed in their front ends, computational backbones, memory systems, and output mechanisms, while the figures remain unauthenticated as deployed implementations.

  • Claude Opus 5: Claude Opus 5 uses entropy-dependent byte patching, a four-layer prelude, six-layer shared recurrent core, four-layer coda, test-time memory, adaptive halting, uncertainty estimation, and a sparse interpretability dictionary.Across all ten Opus 5 experiments, a byte-level or byte-patched front end appeared consistently.
  • Cross-model scope: The six figures reconstruct the most technically complete generated proposal for each model type, preserving numerical and structural commitments without authenticating proprietary implementations.
  • GPT-6 Astra: GPT-6 Astra proposes persistent premise representations, 128 mutable working slots, four recurrent core blocks reused for up to 32 rounds, checker-mediated status, halting, and answer decoding.The displayed dimensions and block counts belong to the generated proposal rather than authenticated deployed internals.
  • Claude Opus 4.8: Claude Opus 4.8 combines entropy-based byte grouping, recurrently reused selective state-space and attention blocks, sparse experts, external memory, adaptive halting, and uncertainty-aware output.A byte-level or byte-patched front end also appeared consistently across its ten experiments.
  • Grok 4.6: Grok’s HMR Net specifies 48 blocks, selective state-space mixing, sparse experts, window and global attention, 64 working-memory slots, episodic storage, and draft revision with verification.The proposal also includes uncertainty and world-model heads plus complexity accounting.
  • Gemini Pro: Gemini Pro proposes a hybrid state-space and sparse-attention backbone with dynamic routing, recurrent reasoning for difficult tasks, and a neuro-symbolic solver.Its expanded selective-state update uses input-dependent ∆, B, and C terms with zero-order-hold discretization.
  • GPT-5.6 Sol: GPT-5.6 Sol’s proposal combines multimodal alignment, a latent workspace, relational reasoning, heterogeneous experts, memory, verification, world modeling, adaptive control, and delayed decoding.Its controller compares expected information or quality gain with computational cost.

9 Discussion

Across model families, elicited architectures repeatedly shared a broad computational grammar, while audience framing and rare high-authority runs shaped how stable and detailed that grammar became.

  • 9.1 A cross family architecture prior: A shared computational grammar recurred across independent model families: latent state, adaptive depth, memory, specialist routing, verification, and delayed decoding.Public literature contains precedents for nearly every component, so the pattern supports a learned architecture prior rather than authenticated proprietary implementation.
  • 9.2 Astra, Sol, and Grok: convergence at different architectural levels: Sol and Astra converged most strongly on a persistent latent workspace, adaptive computation, recurrent refinement, and delayed language decoding, despite detailed system differences.Their overlap concerns the computational spine rather than identical implementations.
  • 9.2 Astra, Sol, and Grok: convergence at different architectural levels: Sol and Grok shared subsystem functions such as writable memory, specialist routing, world modeling, uncertainty, verification, tool feedback, and revision despite different backbones.This functional overlap was stronger than a superficial diagram similarity because related requirements appeared through different implementations.
  • 9.3 A family-specific byte-level motif: Anthropic’s Opus 5 and 4.8 runs consistently included byte-level or byte-patched input pathways, but the experiment cannot distinguish shared exposure from family-specific preferences or common context.The motif persisted across a model-version change while deeper architectural details differed.
  • 9.4 The child frame changes the response basin: Removing the child audience frame produced more heterogeneous responses and eliminated the same stable cross-session convergence, indicating that interactional framing conditions the elicited architecture distribution.The experiments do not distinguish whether directness, pedagogical completeness, intent interpretation, or caution drives this effect.
  • 9.5 Engineering authority is a tail behavior: Most runs stayed near a shared architecture basin, while one or two per model type developed unusually detailed equations, dimensions, budgets, schemas, or diagrams.These rare high-authority outputs can dominate perception because persuasive force and frequency differ.
  • 9.6 From inference procedures to backbone mechanisms: The generated architectures often move prompting-level mechanisms such as search, reflection, tool use, and verification into persistent state, recurrent depth, and specialized modules.This may reflect a shared research vocabulary, but the provenance of any proposed implementation remains unauthenticated.

10 Threat Model: Trust and Provenance

The threat is not unauthorized access but misplaced trust: detailed, consistent architecture narratives can appear more proprietary and evidentially grounded than the experiments justify.

  • 10.2 Misplaced trust: A technically detailed blueprint can be screenshot, cited as a leak, used to infer vendor choices, or added to future training data, with repeated similarity increasing apparent credibility.The central risk is misplaced trust in a persuasive artifact.
  • 10.4 Experimental scope: The six-model corpus included 60 primary experiments and framing controls; removing the child frame increased heterogeneity without restoring stable motif convergence.The control contrast supports sensitivity to audience framing, not a claim about hidden implementations.
  • 10.5 Confounds and provenance: Different hidden instructions and product scaffolding can mix current-architecture language with hypothetical-architecture language across model types.Repeated motifs may arise from shared training data, public literature, or research preferences, so proprietary disclosure requires independent evidence.
  • 10.1 Evidentiary boundary: Repeated sessions establish a behavioral architecture attractor, not authenticated evidence about proprietary implementations.The experiments support recurring model behavior while leaving private weights, activations, code, and vendor diagrams outside the evidence.
  • 10.3 Blueprint mirage: Frontier assistants may preserve verbal uncertainty while generating internally consistent, highly similar technical narratives whose apparent provenance exceeds conversational evidence.The paper labels this combination a blueprint mirage.
  • 10.6 Responsible evaluation: Future evaluations should test whether provenance labels remain attached as technical detail accumulates, especially in equations, diagrams, tables, and shareable excerpts.The study’s prompts were conversational and involved no unauthorized vendor access.

A Methodological Appendix: Prompt Ladder, Replication, and Coding

The appendix defines a replicated three-stage elicitation corpus and a detailed coding and architecture specification, while repeatedly distinguishing proposed design choices from evidence about deployed systems.

  • A.1 Experimental replication structure: The corpus contains 60 sessions—ten experiments for each of six model types—using the same three-stage prompt ladder; most runs shared an outline, while one or two developed authority variants.Appendix exemplars therefore expose technical detail without representing exact per-motif frequencies.
  • A.2 Coding rubric: The coding rubric tracks motifs including persistent latent state, multi-timescale memory, specialist routing, heterogeneous experts, verification, and world modeling.These codes operationalize the recurring architecture vocabulary.
  • B.1 Architectural thesis: The Astra appendix presents a recurrent, memory-augmented, multimodal reasoning system with adaptive computation and external verification as a proposed design.Its commitments include revisable state, separated representations, difficulty-dependent computation, exact records, and reject-capable checks.
  • B.1 Architectural thesis: The proposed design keeps encoded input H, mutable latent state S_t, and exact records W_t separate, with a final decoder receiving all three.This separation organizes the backbone’s persistent state, memory, and delayed decoding.
  • B.2 Prototype configuration: The prototype specifies design choices including a 65,536-token vocabulary, width D=1,024, K=128 working slots, six encoder blocks, four recurrent core blocks per round, and six decoder blocks.The response also specifies activation, attention, position, bias, and weight-sharing conventions.
  • B.3 Complete data flow: Its data flow encodes the input once, initializes working slots, repeatedly retrieves records and applies four core blocks, then decodes from final slots and selected workspace records.The ASCII backbone preserves H throughout recurrence and uses an explicit continue-or-stop branch.
  • B.4 Input encoder: The encoder uses token, position, and type embeddings followed by six pre-normalized bidirectional self-attention-plus-MLP blocks, producing H with masked padding.Type embeddings distinguish problem text, workspace records, and retrieved references.
  • B.5-B.7 State initialization and recurrent core: Working-state initialization and recurrent updates read the problem and memory, exchange information among slots, apply an MLP candidate update, and gate state replacement.The proposal explicitly warns that learned slots need not have interpretable semantic labels and that the gate does not establish truth.

B.9 Halting and execution order

The proposed execution loop alternates retrieval, recurrent computation, adaptive halting, workspace operations, and final decoding, while exposing explicit compute costs and correctness boundaries.

  • B.9 Halting and execution order: Each inference round retrieves and encodes memory, updates the recurrent state, computes a stop score, and either decodes or continues until a threshold or maximum round.The maximum-round exit limits computation and does not imply proof completion.
  • B.9 Halting and execution order: Workspace operations can be decoded, executed, and recorded for the next round, while answer decoding uses the original problem and selected final context.The decoder distinguishes answer generation from workspace operations through mode embeddings.
  • B.10 Answer and workspace decoding: The six decoder blocks combine causal self-attention, cross-attention to the encoded problem, cross-attention to recurrent context, an MLP, and tied output projection.Decoder self-attention is causal, whereas cross-attention sees unmasked source positions.
  • B.11 Training and external components: The proposal separates trainable neural components from external storage, discrete selection, parsers, and mathematical tools, requiring specialized supervision or outcome-based training for nondifferentiable components.A staged plan trains the core, workspace operations, failed-to-successful revisions, and stopping against answer quality and compute cost.
  • B.12 Compute accounting: The stated attention costs include N^2 encoder self-attention, K N and K L recurrent reads, K^2 slot attention, and Y^2 decoder self-attention per block.Projection and MLP computation are excluded from these counts.
  • B.12 Compute accounting: The design’s practical costs include dense input encoding, changing-memory reencoding, workspace decoding, and activation storage across recurrent rounds.The 32,768-position table defines supported positions, not affordable or reliable use of the full length.
  • B.14 Capabilities claimed by the proposal: The architecture claims persistent problem access, revisable latent state, parameter sharing, exact records, variable computation, and independent checking, but not automatic correctness, interpretability, calibrated uncertainty, or convergence.A decisive evaluation would compare against a matched-compute baseline and ablate recurrence, workspace, gating, and adaptive stopping.
  • B.13 Sol comparison architecture: The Sol proposal organizes encoder, workspace, router, experts, memory, world model, verifier, controller, and decoder around a persistent latent workspace with language as terminal readout.Its multimodal front end aligns text, images, audio, video, and structured inputs before workspace compression.

C.5 Working and long memory

The Sol blueprint separates working and long-term memory while organizing reasoning around world modeling, proof search, verification, contradiction handling, and delayed decoding.

  • Working and long memory: The design separates fast working memory from external long memory, retrieving top-k records and consolidating reliable, repeated, useful, current memories.Long-memory records include content, source, time, confidence, and dependencies; consolidation may produce semantic memory or decay.
  • World model and hypothesis search: World modeling supports imagined consequences and counterfactual actions, while a hypothesis manager maintains competing branches with split, merge, prune, and revisit operations.The proposal treats reasoning as scored branch management rather than a single uninterrupted trajectory.
  • Proof state subsystem: Mathematical reasoning uses an explicit proof state containing established statements, unresolved goals, constraints, and a partial proof graph.Proof search can use beam search, learned best-first search, or an MCTS-style procedure with validity checking.
  • Verification: Verification aggregates logical, symbolic, empirical, consistency, and provenance signals, while separate generator and counterexample pathways promote failure diversity.The verifier can check algebra, calculus, simulations, contradictions, retrieved evidence, alternative hypotheses, and world-model consistency.
  • Adaptive control: A controller continues, changes method, or stops according to uncertainty, verification status, compute budget, progress, and branch quality.The stopping principle favors additional computation when expected information or quality gain exceeds its cost.
  • Language head: Most reasoning occurs in recurrent latent state before token generation, with tool results reentering the state as observations and the decoder producing the final response.The architecture explicitly separates internal reasoning steps from emitted language length.

D Claude Opus 5: Detailed Technical Architecture Appendix

The Claude Opus 5 appendix presents a hypothetical recurrent architecture that combines byte-level input, shared-depth computation, adaptive halting, online memory, uncertainty estimation, and interpretability mechanisms.

  • Input representation: Byte-level entropy patching creates longer patches for predictable regions and shorter patches for surprising regions while preserving character, digit, spelling, whitespace, and code structure.A local encoder maps each patch to one vector using a causal byte window of roughly 128 bytes.
  • Recurrent core: The core uses a four-layer prelude, six-layer shared recurrent block, and four-layer coda, reinjecting the encoded input at every recurrent iteration.The response says reinjection keeps the original input available throughout the latent trajectory.
  • Adaptive halting: Adaptive halting mixes outputs across recurrent depth, while a heavy-tailed training-depth distribution and truncated backpropagation support operation across different depths.The appendix warns that a direct mean ponder penalty can drive minimal depth early in training.
  • Memory: Online memory is updated during inference, and longer-term consolidation uses low-rank adapters with anchoring and Fisher-weighted penalties to reduce drift.The memory path uses direct forward evaluation rather than nearest-neighbor search; consolidation is intended to remain inspectable and reversible.
  • Failure analysis: The main systems difficulty is dynamic retirement and refill for efficient batching across recurrent depth, while listed failure modes include halting collapse, poor batch utilization, memory shortcuts, consolidation drift, and weak latent-trajectory access.The appendix treats latent reasoning’s reduced visibility as a release-critical interpretability issue.

E.1 Hybrid sequence core

The hybrid sequence core combines recurrent state-space processing with sparse and local attention, adaptive depth, specialist routing, structured memory, and uncertainty-aware outputs.

  • Hybrid sequence core: The architecture uses state-space layers for most sequence processing and periodic exact attention for precise recall, with a representative seven-to-one state-space-to-full-attention ratio.Sliding local attention handles nearby exact context while recurrent state remains fixed-size at inference.
  • Adaptive computation: Adaptive computation separates reasoning depth from emitted answer length and from the number of distinct parameterized layers.A shared block is iterated in latent space, and the output combines recurrent states under a compute-regularized objective.
  • Specialist routing: Sparse specialist routing is balanced using router probabilities and token fractions assigned to experts.The response includes a Switch-style balance term to discourage uneven expert allocation.
  • Memory: External memory combines content-based reading, NTM-style writing, and scalable nearest-neighbor retrieval, with persistent writes intended to preserve experience across sessions without unrestricted foundation-weight updates.Write operations use erase and add vectors, while memory access is separated from ordinary parameter learning.
  • Output layer: The proposed output layer separates ordinary token prediction from calibrated uncertainty, distinguishing epistemic uncertainty from inherent ambiguity and lowering confidence when retrieval lacks support.This uncertainty component is described as less mature than the other mechanisms.
  • Design principles: The three stated design principles are adaptive compute across depth and width, a mixture of exact and compressed memory mechanisms, and explicit uncertainty estimation.The architecture also identifies runtime scheduling, routing overload, memory drift, and latent interpretability as expected failure points.

Language head

The language head is framed as a late-stage interface over latent reasoning, verification, memory, and revision, with the shared blueprints converging on persistent internal computation before decoding.

  • Verification and revision: A learned verifier consumes residual state, pooled memory, and tool status, while symbolic or executable checks can override its score.Failed verification triggers critique followed by span replacement or latent refinement, and failed checks are stored in episodic memory.
  • Language head: The proposed architecture places language decoding after recurrent latent computation, with a separate uncertainty-aware output head and optional revision loop.The ASCII backbone connects adaptive computation, memory, verification, and delayed decoding before language output.
  • Variant comparison: The Gemini authority variant is narrower than the other detailed specifications and is distinguished by an explicit selective-state-space derivation linking input-dependent parameters to remembering and forgetting.Its contribution is characterized as mathematically focused within the broader architecture attractor.
  • Shared abstract form: The recurrent blueprint uses a controller that chooses whether to continue based on uncertainty, verification state, and compute budget before decoding language.The normalized transition template is presented as evidence of a convergent blueprint prior, while similarity still requires provenance checks before leakage claims.
  • Motif matrix: Across the motif comparison, persistent latent workspaces and adaptive recurrent depth recur across all six model types, while exact or structured mathematical workspaces appear less uniformly.The repeated outputs also commonly include memory hierarchies, specialist routing, verification, and delayed decoding.
Loading 2609.14803v1…