AI Benchmark Half-Life in Recursive Corpora: A Theory of Validity Decay under Semantic Leakage and Regeneration

Abstract

This preprint develops a theory of AI benchmark half-life in recursive corpora under semantic leakage and regeneration, yielding validity-decay bounds and monitoring rules for evaluation systems whose items and solution traces re-enter public data. It models benchmark validity through discriminative power and construct validity, and derives jump-aware lifetime bounds, partial-identification results, portfolio design criteria, and safe sequential control under ambiguity and partial observability.

Keywords

• AI benchmark half-life
• recursive corpora
• semantic leakage
• validity decay
• benchmark contamination
• construct validity
• discriminative power
• dynamic benchmarks
• partial identification
• sequential monitoring
• lineage observability
• model metrology

Identifiers and source records

• DOI
• Works entry
• ORCID