Cochinescu's Structured Unpredictability. The Entropy Stance Tested
In July, Sebastian Cochinescu of the University of Bucharest published a framework arguing that perceived mind in AI agents rests on four expressible stances, time, truth, entropy, and love, covered here in the Perceived AGI analysis. The framework’s second paper, posted to arXiv on July 20 as arXiv:2609.19044, picks the hardest of the four and tries to build it. The result is an unusually honest experiment, one real effect established, two stronger predictions failed, and a boundary drawn by the author before anyone else could draw it for him.
The boundary comes first because it controls everything after it. The paper’s object, in its own words, is the evidence structure from which an observer could infer a hidden state, not the possession of one in any phenomenal sense. This is perception engineering, not a theory of machine consciousness. Nothing about what a model actually experiences is claimed, tested, or implied.
What structured unpredictability is
The entropy stance asks an agent to express orientation toward surprise, order, and decay rather than pure reactivity. The paper operationalizes that as structured unpredictability, defined as conditional dependence between an output and a persistent hidden state beyond what an observer can infer from the transcript. The formal target is an observer-conditional entropy gap, written as the difference between the entropy of the output given the transcript and the entropy of the output given both the transcript and the hidden state, which is the conditional mutual information between output and hidden state.
The gap itself is not measured in the paper. Two proxies stand in, and the author flags that substitution as a limitation. The first, N, is lexical novelty, measured as pooled bigram non-overlap from the distinct-n family. The second, C, is stylometric consistency, built from function-word distributions, sentence length, and type-token ratio. The target behavior is the high-novelty, high-consistency quadrant, a system whose outputs keep surprising you lexically while remaining recognizably the same voice.
The hidden state that drives the proxies is engineered rather than learned. It has twelve dimensions, six style groups and six topics, and it evolves through two mechanisms, exponential accumulation with a decay parameter of 0.08, and a Van der Pol oscillator running open-loop at one cycle per forty turns. The state is rendered into the prompt each turn. The base model, a 4-bit quantized Qwen2.5-1.5B-Instruct, never sees the preceding dialogue, only the current turn plus the rendered state.
The experiment
Nothing is trained. The paper runs seven arms against the same fixed base model, a mechanism arm that renders the evolving state, a low-variance control, a novelty-matched sampling control that equalizes raw sampling randomness, a memory-reset ablation, novelty-only and consistency-only selectors, and a temperature and nucleus-sampling grid. The design went through a synthetic validation stage, a real-model pilot, a renderer gate, a powered grid of 24 sequences per arm, and a final grid of 56 sequences per arm, each sequence twelve turns with evaluation on the final six-turn window. The whole protocol was preregistered on OSF before the final grid ran.
The control structure is the paper’s strongest feature. Novelty-matched sampling exists because a critic could say the mechanism arm looks novel simply because sampling temperature was higher. Matching raw sampling randomness isolates what the hidden state contributes beyond entropy the sampler would produce anyway.
What passed and what failed
One prediction passed. Lexical novelty rose in the mechanism arm by 0.073 over the low-variance control, with a 95 percent confidence interval of 0.057 to 0.103, and by 0.023 over the consistency-only selector, interval 0.018 to 0.034, both at a Holm-corrected p of 0.0004. The mechanism arm reached a median novelty of 0.984. The effect sits at the response-selector level, the arm that chooses among candidate responses using the state, and the paper is careful to locate it there rather than claim the state itself drives generation.
Two stronger predictions failed, and the failures are the paper’s most informative results. Twin separation, the requirement that two agents with different selection histories diverge more than agents sharing a history, was not established. The observed effect was 0.003 with a confidence interval from minus 0.011 to 0.019, indistinguishable from zero. The joint novelty-and-consistency criterion also failed, with the consistency contrast against novelty-matched sampling at minus 0.006, equivalent to zero. The original accumulation criterion failed as well. A revised reset contrast, comparing the mechanism arm against the memory-reset ablation, came in positive at 0.028, but the author states plainly that a fixed nonzero control state could produce the same contrast, so it does not establish path dependence, the requirement that outputs depend on the accumulated selection history rather than on a fixed conditioning state.
In the framework’s own terms, the paper does not establish the joint entropy condition, and the handoff condition for a perception study, the step where a second experiment would test whether human observers actually attribute more mind to high-novelty agents, was not met.
What the failures mean
The pattern is specific. The hidden state makes outputs more novel, and that effect survives the sampling-matched control. But the state’s history does not leave a detectable signature in the outputs, which means the twins test cannot distinguish a persistent accumulating state from a state that gets re-rendered each turn from the same distribution. For the entropy stance as a component of perceived mind, this is the difference between performing unpredictability and performing continuity of unpredictability. The first is established. The second, the part a human interlocutor would need to infer an ongoing interior life, is not.
The limitations section adds the right caveats. The proxies measure non-repetition and style stability, not the entropy gap itself. The novelty metric and the selector objective are aligned by construction, which means the positive result partly confirms the engineering. One model at one quantization supports no generalization claims. And evaluation is independent-turn only, so nothing here supports long-context or memory claims.
Comparison to The Consciousness AI
The paper’s boundary is the same one this project’s research stance draws. The Consciousness AI architecture treats consciousness as an emerging property of system dynamics, substrate independent, and treats the impression of consciousness as a separate, buildable artifact. Cochinescu’s series is a formal demonstration of the second half. Perceived mind can be engineered incrementally, stance by stance, and the engineering can be audited with preregistered designs, while the question of what, if anything, the system experiences stays outside the experiment entirely. The project’s own evaluation discipline, documented on the project repository, holds that behavioral impressiveness is never evidence of experience, and this paper supplies a concrete example of why, a system that renders more believable novelty without any claim about interiority being warranted. The flagship field survey keeps this perception-side literature distinct from the consciousness-science side, and this paper belongs squarely on the perception side.
The paper “Entropy in Conversational AI. Structured Unpredictability as Inferrable Interiority” by Sebastian Cochinescu was posted to arXiv on July 20, 2026 as arXiv:2609.19044, with preregistration at osf.io/dyt3w and code on Zenodo.