Sufficient constraints to prevent spoofed declarations in partial re-execution

Identify and validate a set of constraints for partial workload re-execution protocols (including proof-of-learning style checks) that are sufficient to rule out spoofed training or inference declarations for large-scale AI workloads.

Background

Partial re-execution aims to verify that declared workloads were actually run by re-running sampled segments and checking reproducibility. To prevent adversaries from fabricating plausible but false declarations, the report proposes adding constraints (e.g., randomness, timing, and trend-consistency checks).

The authors note that while many constraints are plausible, establishing which sets are actually sufficient to preclude spoofs remains unresolved.