Evidence · DiggingBeagle record

Evidence-gated authorization for AI reviewer proposal overrides

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. Research/disclosure claims are specific to the documented settings; do not infer unreported victims or additional compromise.

Source role
primary disclosure

Evidence record

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. Research/disclosure claims are specific to the documented settings; do not infer unreported victims or additional compromise.

Read the original source ↗

Claim-level citations (2)