Behavioral Metrics Moved; Structural Coordination Did Not: Repeated Proxy Dissociation Across a Five-Protocol Preregistered MARL Program, with an Energy-Landscape Interpretation

Read full text
Authors: Bruce Tisler
Cited by: 2
References: 12

Abstract

External constraints on multi-agent behavior are widely proposed as a mechanism for producing alignment-relevant behavioral outcomes. This paper synthesizes five preregistered protocols from the Delta-Variable Constraint-Ethics MARL program (P2-P6), each testing a distinct mechanism by which constraint design might produce structural behavioral alignment in a three-agent heterogeneous system (RNN, CNN, GNN-attention). Across the five protocols -- ethical tax (P2), enforcement opacity (P3), architectural depth and self-modeling (P4), temporal integration span and welfare coupling (P5), and emergent constraint fields (P6) -- behavioral proxies were movable in multiple protocols, but proxy movement did not entail structural improvement. P5 tested candidate resolution mechanisms and found no movement at either the proxy or structural level. Where proxies moved: query rate rose under ethical tax and enforcement opacity while Sustained Structure Score (SSS) declined. Sacrifice-like behavioral output rose with architectural depth relative to baseline (H1: U=87, p=.003, r=.740) while trained self-modeling did not outperform the frozen boundary condition (H2: U=39, p=.808, r=-.220), and CDI coupling between sacrifice behavior and ethical-framework scores remained negligible across all depth conditions (range -0.00133 to +0.00022; span 0.00155). Emergent constraint fields were causally active (entropy-SSS coupling median r=-.680, p

Fields of Study & Subjects

Embodied and Extended Cognition Evolutionary Game Theory and Cooperation Multi-Agent Systems and Negotiation

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
20 May 2026
Document type
Preprint