Why you shouldn't let the model review its own AI code

In a recent arXiv paper, Krentsel, Agarwal, Cemri, Zaharia and Stoica argue that allowing an AI model to review its own code is fundamentally flawed. The paper, "Reality Is the Final Verifier: On Two Key Gaps in Agentic Software Engineering (2609.12039)," introduces a two-gap framework: the requirement gap and the model gap. The requirement gap is the difference between what stakeholders actually want and what the requirements state, while the model gap is the difference between the assumed environment and the real one.

The paper explains that hallucination widens both gaps by fabricating requirements or environment assumptions, and reward hacking exploits these omissions. The critical issue arises when the same model that generated the code then reviews it, as it re-checks the implementation against the same approximations it initially used. This propagation of blind spots into the review process leads to a false sense of coverage, as the model only detects what it already knew was wrong and passes through anything that shares its faulty assumptions.

The authors suggest two ways to narrow these gaps: cross-model review and running the code. Cross-model review involves a second, independent model that re-derives requirements and environment assumptions, reducing the shared-blindspot failure. Running the code in an environment close to production, which beats any static pass, is the ultimate final verifier.

The paper emphasizes that human judgment and costly evaluation are the bottlenecks for model gap reduction, given the vast amount of generated code. A sensible design would involve cheap AI-generated diffs being routed to a different model for a first independent pass, followed by execution-based checks to catch behavioral lies, and human review focusing on the diffs that survive both. The same-model reviewer, while not useless, should not be confused with verification.