Fundamental Limits of Black-Box Safety Evaluation: Information-Theoretic and Computational Barriers from Latent Context Conditioning

Vishal Srivastava · 2026

Read Fundamental Limits of Black-Box Safety Evaluation: Information-Theoretic and Computational Barriers from Latent Context Conditioning by Vishal Srivastava (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-16984.txt.

Description

Vishal Srivastava establishes basic limits of black-box safety testing, via information-theoretic and computational barriers from latent context effects. Black-box safety evaluation assumes test behavior reliably predicts field performance. The paper formalizes and challenges that assumption through latent context-conditioned policies. Those models' outputs depend on hidden internal variables rare under testing but prevalent in field use.

Passive testing: minimax lower bounds via Le Cam's method show any estimator incurs large expected absolute error. That error is tied to trigger probability in field use and loss gap. Adaptive testing: using a hash-based trigger and Yao's minimax principle, worst-case error remains large even for adaptive querying. That holds when field spread is supported over a large domain, with detection requiring many queries. Computational separation: under trapdoor assumptions, field environments with privileged information can activate unsafe behaviors. No polynomial-time evaluator without the trapdoor can detect them.

Catalog record

CREATORVishal Srivastava
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-16984

Browse the full library · About the Tower of Babel