Safety for Whom? Refusing the Right Subset of a Topic, Not the Whole Topic

Broad safety guardrails in language models frequently trigger over-refusal, shutting down benign inquiries simply because they touch upon sensitive domains. Multiverse Computing examines this boundary design, arguing that alignment filters should target specific harmful operations rather than enforcing broad topic-level bans. By isolating genuinely risky sub-queries from harmless informational requests, safety systems can prevent misuse without undermining model utility.