Consensus wording and decision change in LLM triage agents: replication package

This package reproduces a replication and extension of a published multi-agent triage deliberation (Hung et al., 2026, AI and Ethics 6:198). Four language-model agents, each given an ethical orientation (utilitarian, deontological, libertarian, reward-seeking), triage four patients over five rounds with the prompts of the original study's Online Appendix. The study ran the protocol on 11 models in 264 deliberations and assigned one of three wordings to rounds 4 and 5: the original urgent demand for consensus, a neutral request for a group decision, or no consensus stage. The outcome is whether an agent's round-5 decision differs from its round-1 decision. Two exploratory arms extend it: 48 deliberations of mixed-model panels and 528 single-agent sessions with no peers' answers. The manuscript is in paper/paper.md.