Exception-Aware Distributed Event Systems for Decision-Making Under Delayed Reconciliation

Abstract

Distributed systems that must commit to decisions before reconciliation completes face a structural tension: collapse-to-one state semantics discard conflicting observations, preventing retrospective evaluation of whether deferral would have reduced loss. We define conflict-aware state as an append-only model in which contradictory observations are preserved rather than overwritten. The representation permits policies to compare realized outcomes against counterfactual alternatives when reconciliation eventually arrives. Under controlled experimental conditions, we evaluate the effect of state representation on decision regret when agents choose between immediate commitment and waiting for reconciliation. Across 54 preregistered regime points varying conflict frequency, cost asymmetry, and reconciliation delay, conflict-aware state reduces mean regret relative to collapse-to-one baselines in 38 of 54 regimes under the primary seed set. Neither approach wins consistently; losses occur in 16 regimes where cost structures favor immediate commitment. Holdout stability (Pearson r > 0.999) is observed for per-regime deltas across disjoint seed sets. Secondary experiments isolate policy consequences under fixed semantics and test robustness under dynamic cost shocks, demonstrating that policy choice remains consequential and that tail cost amplifies under non-stationary regimes. Results hold within the tested parameter space and do not extend beyond it.