The New Turing Test: A Methods Demonstration in Irreversible History and Permissible Futures

Abstract

This record presents an exploratory, reproducible computational methods demonstration, motivated by the question of whether artificial agents behave differently when irreversible history changes their attainable futures. The Turing framing is a motivating application, not a claim that the experiments detect or establish consciousness. This paper formally introduces and defines the six-component candidate governing-class architecture G = (ℒ, Φ, Θ, κ, 𝒱, U_adm), comprising constrained admissible laws, state reconstruction, admissible parameter variation, local-law selection, viability construction, and interventions. It builds on, but is distinct from, the earlier ACT Hypothesis and Theory article (accepted in Frontiers in Immunology) and the separate Permissible Phenotype Conjecture working paper. Neither earlier proposal is validated by this synthetic study. A biological instantiation of G remains future work; any instantiation of G must prespecify its components before observing outcome data. Five sequential experiments with researcher-specified synthetic agents tested actual versus represented vulnerability, policy-selected versus physically accessible futures, ordinary reward optimization, adversarial phantom-law comparisons, and neural-policy replication. At a 36% risk level, a frozen ordinary planner predicted a 32.14-percentage-point increase in preservation after an experimentally engineered loss; 150 simulated trajectories per comparison condition yielded a 25.34-point sample contrast. The response followed from an authored state-dependent terminal-value rule, not a discovered existential mechanism. An apparent additional neural-policy residual weakened from 6.69 to 1.70 points in a subsequent cross-architecture challenge, without excluding smaller effects. The contribution is a transparent experimental protocol, formal introduction of a candidate governing-class architecture, and a demonstration of why causal interpretation, statistical identifiability, and negative controls matter when interpreting preservation-like behavior. The study does not establish a new law of biology or cognition, an autonomous self-model, sentience, or subjective experience. Chalmers’s hard problem remains open. All protocols, code, recorded outputs, and neural checkpoints for the five experiments are included for independent reproduction. Researchers interested in the methods or in future applications of G are invited to contact the author.