A computational framework for comparing deterministic, stochastic, and adaptive veto-regulated decision dynamics

Decision-making is commonly modeled using deterministic dynamics, stochastic processes, or evidence-accumulation mechanisms, but these approaches often provide limited accounts of how an emerging action may be inhibited before execution. This study presents a comparative computational framework for veto-regulated decision dynamics that separates deterministic accumulation, stochastic variability, fixed-threshold inhibition, mechanistic threshold adaptation, and reinforcement-learning-based threshold selection. Five candidate models were evaluated. M1 represented deterministic decision accumulation, M2 added stochastic variability, M3 introduced a fixed-threshold veto mechanism, M4 implemented an adaptive conflict-sensitive and homeostatic threshold rule, and M5 used a demonstration-pretrained Double Deep Q-Network to select among discrete veto-threshold setpoints. M1–M4 were evaluated using independently generated phenomenological reference trajectories, bounded nonlinear parameter estimation, chronological train-test separation, and 100 paired Monte Carlo realizations. M5 was evaluated separately using 20 independently trained agents across four data splits, 100 reserved primary-test trajectories, a matched no-demonstration ablation, and 100 external-generator trajectories. For global held-out trajectory prediction, M1 and M2 achieved the lowest mean testing RMSE values, 0.2457 and 0.2460, respectively. M4 achieved a mean testing RMSE of 0.3305 and substantially outperformed the fixed-threshold model M3, which achieved 0.5487. Thus, mechanistic threshold adaptation improved explicit veto regulation but did not surpass the simpler accumulation models on global RMSE. A dense sensitivity analysis comprising 117 threshold–noise combinations and 11,700 simulations showed that higher veto thresholds reduced the veto-event rate, delayed the first veto, and prolonged decision duration. Because the inverse threshold–event-rate relationship followed partly from the predefined threshold-crossing rule, it was interpreted as a quantified implementation property rather than as an independent theoretical discovery. Stochastic noise had a smaller effect on the mean veto-event rate but increased uncertainty and between-run variability. On the primary M5 benchmark, the learned controller achieved a mean RMSE of 0.257938 compared with 0.251138 for M4; the mean difference was 0.006800, with a hierarchical-bootstrap 95% confidence interval of − 0.003077 to 0.017225. Only 3 of 20 M5 agents outperformed M4. M5 also produced a lower veto-event F1 score, showed no reliable benefit from demonstration pretraining, and generalized less effectively than M4 under distribution shift. The EEG component proceeded in two stages. An initial single-participant analysis demonstrated how an empirical motor-cortex signal could be incorporated into the framework. The analysis was then extended to 30 participants from the PhysioNet EEG Motor Movement/Imagery Dataset, with participants treated as the independent observational units. Across participants, the combined 8–30 Hz sensorimotor envelope decreased by 6.60% relative to the pre-cue baseline, with a 95% confidence interval from − 9.48% to − 3.98%. Significant attenuation was also observed in the mu and beta bands. These findings demonstrate reproducible motor-related EEG modulation and support construction of an empirical cognitive-input signal, but they do not independently validate the modeled veto mechanism. Conflict-sensitive simulations showed that increased assigned contextual conflict raised modeled veto activation while reducing decision probability and reward. Because the conflict variables were imposed computationally and entered directly into the conflict function, these findings should be interpreted as a manipulation check of model responsiveness rather than as evidence that biological conflict causes inhibitory control. Overall, the results show that conflict-sensitive mechanistic adaptation improves fixed-threshold veto regulation, whereas greater reinforcement-learning complexity does not necessarily improve accuracy, inhibitory-event fidelity, reliability, or generalization. The framework should therefore be interpreted as a reduced phenomenological and proof-of-concept model of veto-like inhibitory control rather than as a complete theory of consciousness, decision-making, voluntary agency, or free will.