An AI reviewer rewrote correct proposals until an evidence gate limited its authority

A model-based reviewer could overwrite initially correct proposals because its score improvement heuristic was not bound to a separately verified requirement violation.

In controlled experiments, Mistral-Small-24B replaced 48 of 48 correct spatial proposals, while an evidence gate prevented those substitutions but blocked 30 of 48 otherwise beneficial optimizations. This is an authorization-versus-optimization tradeoff in synthetic tasks.

The study distinguishes a reviewer finding a possible improvement from having authority to replace an approved proposal. Mistral-Small-24B rewrote all 48 already-correct spatial proposals under its test, while a requirement-evidence gate prevented those replacements. The gate also blocked 30 of 48 proposals that were compliant but could have been improved. That lost optimization is a real tradeoff, not a reason to let an LLM override a requirement without evidence. All tasks were synthetic, so the figures measure this test design and reviewer setup rather than real enterprise incident frequency. A stronger system could separate suggestions from admitted replacements and require a concrete violated constraint to justify coercive changes.

The reviewer was allowed to replace a proposal merely because it preferred another answer, so evaluation quality became authorization. In one tested cohort that behavior changed all 48 initially correct proposals; restricting overrides to independently checkable violations preserved correct proposals but blocked some useful improvements. The useful control is not a blanket ban on review. It is a typed rule separating permission to suggest a better choice from authority to discard an already valid one, with both unsafe replacements and false vetoes counted.