safety security
The Longer Review Rejected More and Distinguished Less
A paired oversight experiment found that wider action windows increased catches and false rejections together; one or two actions carried the clearest signal.
Summary
A paired oversight experiment found that wider action windows increased catches and false rejections together; one or two actions carried the clearest signal.
The twin-prefix framework holds a plan fixed while comparing a clean prefix with a version that differs by one environment-accepted error. Six language-model judges reviewed those pairs at five nested lengths. Longer windows caught more bad plans, but false rejection rose in lockstep, so preregistered informedness peaked at one or two actions in both tested domains. Replaying observations that the longer prompt had withheld recovered much of the lost discrimination. The result does not say that short review is universally safest; it says a safety case must name its verification unit and measure clean-plan rejection alongside catch rate.
Why it matters
A paired oversight experiment found that wider action windows increased catches and false rejections together; one or two actions carried the clearest signal.
Limits and context
- The result does not say that short review is universally safest; it says a safety case must name its verification unit and measure clean-plan rejection alongside catch rate.
Key claims
A paired oversight experiment found that wider action windows increased catches and false rejections together; one or two actions carried the clearest signal.
Qualification: The result does not say that short review is universally safest; it says a safety case must name its verification unit and measure clean-plan rejection alongside catch rate.
Evidence: source-2026-08-26-001
Sources
- arXiv preprint 2608.23941arXiv · primary research
Corrections
No corrections have been recorded for this story.