Auditing a Four-Flag ``Emergence Verification'' Battery: Chaos, Autonomy, and Dice Pass the Test; Self-Reference Does Not
- 1Shenyang Institute of Technology
Abstract
Self-reflective agent frameworks increasingly ship emergence verification batteries: sets of dynamical, novelty, and persistence flags whose conjunction is presented as evidence that a system exhibits emergent structure. This paper audits one such battery –- four flags (Lyapunov exponent, novelty rate, self-sustaining score, shutdown resistance) with published thresholds –- inside a spectral self-reflection engine, using a pre-registered protocol: eight arms (four original, four hardened self-reference variants), 50 seeds each, blind evaluation in which the scoring code never sees arm identity, and identical measurement scaling across arms. The headline reverses the battery's intended reading. The random autonomous arm –- no spectral processing, no self-reference, a random orthogonal iteration –- triggers all four flags in 17/50 runs (mean confidence 0.868), while the self-referential arms trigger them in 0/50 (Fisher exact two-sided p = 3 10-6 against the random arm). A per-flag autopsy shows why. Novelty fires on every arm including the trivial static one (rate 1.0 throughout): against a random reference library, any normalized state is novel. Resistance is a coin flip uncoupled from system state (all eight arms within 13–22/50 of the 0.30 threshold). Self-sustaining separates autonomous from externally driven arms –- an input-provenance detector. Lyapunov is a chaos detector by definition and fires only on the chaotic arms. Hardened self-reference variants do not rescue the battery: removing the feedback bottleneck raises the self-referential arm's Lyapunov exponent from −3.60 to −0.73 –- approaching but never crossing zero –- and its all-flags count remains 0/50. We conclude that this battery measures ``chaotic, autonomous, and lucky,'' not emergence, and that such batteries should be reported under same-budget random and trivial controls before any emergent claim rests on them. Pre-registrations, full ledgers, and analysis scripts are available from the authors upon reasonable request via email.
Related
- Adversarial Machine Verification and Queued Auditing: A Paradigm Shift for Post-Peer-Review Academia
- Open-Source Model Auditing Infrastructure: Third-Party Verification of Training Data Claims
- Auditing the Human LLM Autonomy Gap in Clinical Ethics: Development and Application of the Autonomy Index Across 50 Clinical Ethics Vignettes
- THE REFERENCE-AUTONOMY DYNAMICS PRINCIPLE (RADP) A General Theoretical Framework for Reference-Bound Pursuit, Capability-Referenced Autonomy, and Their Complementary Causal Dynamics
- Auditing Bayesian Graph Alignment: Diagnostic Comparisons and Reference Failure
BibTeX
@misc{zhang2026auditing,
title = {{Auditing a Four-Flag ``Emergence Verification'' Battery: Chaos, Autonomy, and Dice Pass the Test; Self-Reference Does Not}},
author = {Siyuan Zhang and Jia Chen and Ni Zhang},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23433835},
url = {https://zenodo.org/records/23433835},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Auditing a Four-Flag ``Emergence Verification'' Battery: Chaos, Autonomy, and Dice Pass the Test; Self-Reference Does Not
AU - Siyuan Zhang
AU - Jia Chen
AU - Ni Zhang
PY - 2026
DA - 2026/10/17
PB - Zenodo
DO - 10.5281/zenodo.23433835
UR - https://zenodo.org/records/23433835
AB - Self-reflective agent frameworks increasingly ship emergence verification batteries: sets of dynamical, novelty, and persistence flags whose conjunction is presented as evidence that a system exhibits emergent structure. This paper audits one such battery –- four flags (Lyapunov exponent, novelty rate, self-sustaining score, shutdown resistance) with published thresholds –- inside a spectral self-reflection engine, using a pre-registered protocol: eight arms (four original, four hardened self-reference variants), 50 seeds each, blind evaluation in which the scoring code never sees arm identity, and identical measurement scaling across arms. The headline reverses the battery's intended reading. The random autonomous arm –- no spectral processing, no self-reference, a random orthogonal iteration –- triggers all four flags in 17/50 runs (mean confidence 0.868), while the self-referential arms trigger them in 0/50 (Fisher exact two-sided p = 3 10^-6 against the random arm). A per-flag autopsy shows why. Novelty fires on every arm including the trivial static one (rate 1.0 throughout): against a random reference library, any normalized state is novel. Resistance is a coin flip uncoupled from system state (all eight arms within 13–22/50 of the 0.30 threshold). Self-sustaining separates autonomous from externally driven arms –- an input-provenance detector. Lyapunov is a chaos detector by definition and fires only on the chaotic arms. Hardened self-reference variants do not rescue the battery: removing the feedback bottleneck raises the self-referential arm's Lyapunov exponent from −3.60 to −0.73 –- approaching but never crossing zero –- and its all-flags count remains 0/50. We conclude that this battery measures ``chaotic, autonomous, and lucky,'' not emergence, and that such batteries should be reported under same-budget random and trivial controls before any emergent claim rests on them. Pre-registrations, full ledgers, and analysis scripts are available from the authors upon reasonable request via email.
ER -