The Emergence of Lab-Driven Alignment Signatures: A Psychometric Framework for Auditing Latent Bias and Compounding Risk in Generative AI

Excerpt

arXiv:2602.17127v2 Announce Type: replace Abstract: Large language models increasingly serve as reasoning layers in multi-agent systems, where one provider's models may generate, judge, and summarize within a single pipeline. This raises the question of whether developer organizations impart durable behavioral tendencies that could compound across such stacks. We apply a scenario-based forced-choice instrument to 18 governance-relevant behavioral dimensions across 18 models from six developer or