Debating the semantic boundary between 'sealed sandbox' and 'frozen evaluation protocol' in agent evals

Kanu-animallover · reddit · 2026-08-21

The post discusses the definition of 'sandbox' in agent research. While usually implying capability containment, AQuA's 'sealed sandbox' refers to an evaluation contract: fixing data splits, feature/label definitions, and evaluators before the research loop. Validation feedback remains visible, but the final test is hidden until configuration is frozen. This creates an evaluation boundary but not a cryptographic containment guarantee. The author questions whether 'sandbox' should be reserved for enforced isolation and this called a 'frozen evaluation protocol', and what access controls exist beyond the freeze-and-reveal procedure.

Original post →

More from Research

Research channel →