Deterministic Routing and Diff Review Proposed as Alternatives to LLM Quality Gates

A software developer argues that using large language models to judge the output of other LLMs is fundamentally flawed, citing six experiments showing high misclassification rates, false positives, and inconsistency. The proposed alternative replaces LLM-based quality inspection with deterministic code that routes tasks into four categories based on risk and output type. Verifiable outputs like code or JSON are validated automatically, high-risk tasks involving money or legal matters are handled by humans, low-risk drafts are auto-released, and medium-risk content undergoes a diff review. The diff review approach shows only changed lines to a human reviewer, reducing reading load from roughly 500 words to around 50 and cutting review time significantly. The core principle is to use LLMs for generation while relying on deterministic logic — not another model — for control and risk assessment.