Temporal Integration Span and Welfare Coupling Did Not Resolve the Optimization-Sacrifice Dissociation: Protocol 5 Results

This record reports Protocol 5 of the Δ-Variable Constraint-Ethics MARL research program. Protocol 5 tested whether temporal integration span and welfare coupling alter sacrifice-choice behavior and ethical-framework coupling in constrained multi-agent reinforcement learning. Sixty confirmatory runs were conducted across six conditions: short-span individual reward, short-span welfare-coupled reward, long-span individual reward, long-span welfare-coupled reward, and frozen-boundary variants. The study found a complete null across five preregistered hypotheses under the tested parameterization. Temporal integration span and welfare coupling did not resolve the dissociation between sacrifice-like behavioral output and ethical-framework coupling. The record includes explicit caveats: non-frozen C1 manifest files were not generated because the condition was run before manifest-copy logic was added; condition assignment remains filename-verifiable. Long-span conditions showed approximately 30–35% null sacrifice-opportunity epochs due to communication-load gating, so cross-span comparisons should be interpreted with this limitation. The paper and supplementary files document the preregistered design, deviations, confirmatory results, and limitations. Preregistration DOI: 10.5281/zenodo.19038790. Repository: github.com/btisler-DS/constraint-ethics-necessity.