Standard LLM-agent evaluation reports how well a system behaves, but it says nothing about the identity of the runtime stack actually deployed. That gap creates a blind spot: a tampered deployment that keeps the evaluated outputs intact will pass every output-only evaluator, because the evaluator never checks what is running underneath.
To address this, the authors introduce DIB-Bench, a benchmark designed to expose deployment-integrity blindness. The work highlights that behavioral scores alone are insufficient for trustworthy evaluation, and that the runtime stack itself needs to be audited.
The abstract does not provide details on how DIB-Bench works or its results, but the core claim is clear: stronger judges do not see more if they only look at outputs. The paper is available on arXiv under the identifier 2610.21135.