Igor Santos-Grueiro

The paper identifies and measures a critical failure mode where LLM agents violate policies by losing or corrupting directive-bearing state during the process of assembling the decision context, and proposes a control layer, SafeContext, to mitigate this risk.

The paper proves that for resources with structural parallelizability (like divisibility and transferability), it is impossible to enforce a linear cost for concentrating influence, demonstrating that influence can be acquired at sublinear cost.