Before a Marker Becomes a Mechanism Adversarial Synthetic Falsification of Dynamical Signatures in Consciousness Research, with Implications for Causal Stress Testing

Abstract

Before a Marker Becomes a Mechanism develops an adversarial synthetic falsification framework for evaluating dynamical markers in consciousness research before they are granted strong mechanistic interpretation. The study emerged from the progressive failure of its own candidate mechanism. An initial three-step causal-closure metric failed to capture longer directed loops. A generalized multiscale closure measure corrected this formal defect but was strongly maximized by a deliberately simple directed ring. State-dependent modulation of causal relations did not restore specificity. Subsequent experiments showed that functional similarity can coexist with radically different causal organization, that an acyclic system with adversarial hidden forcing can exactly reproduce the passive trajectory of a recurrent system while responding differently to controlled perturbation, that nearly identical scalar causal scores can conceal substantially different causal geometries, and that behavioral output can be removed without altering internal dynamics. A sensitivity analysis across 288 paired parameter–seed configurations confirmed that the trivial-maximizer failure was robust within the explored synthetic model family. These results do not identify a mechanism or marker of consciousness, nor do they establish whether the synthetic systems are conscious or non-conscious. Instead, they demonstrate several distinct ways in which a mathematically well-defined and potentially predictive marker can fail to justify the mechanistic interpretation assigned to it. The resulting framework is organized as an F0–F5 falsification ladder: definition and trivial-maximizer testing, passive-equivalence adversaries, scalar-collision testing, responsiveness dissociation, perturbational discrimination, and finally biological experience testing. The central methodological proposal is that candidate markers should survive inexpensive synthetic adversarial challenges before advancing to stronger biological and mechanistic claims. The paper also outlines a prospective extension of the same logic to causal stress testing of adaptive AI systems, while explicitly avoiding the use of causal metrics as optimization targets. Code, run-level data, summary tables, figures, and sensitivity-analysis materials are included for reproducibility.