Abstract

This paper studies decision records that say, before anyone decides, what result would overturn their own recommendation. It reads the twelve cases of one AEGIS programme of corrective review. All twelve are authored synthetic training cases whose figures and thresholds AEGIS calls training assumptions, so the paper draws no fact about the world from them. It asks what a written condition for reversal commits the decider to, and what the record says about where the condition's numbers come from. Every case counts its reversal over a small, named test of three to twenty items, and tolerates one failure or none. The four cases whose recommendation is the more restrictive choice can be reversed only by a clean result; the other eight are reversed by a single failure or the failure of a single item. Every option names what it gives up or accepts. And every record says where its numbers come from: ten name something as the original's, eleven name a figure as added by the author, and three call a figure a design or setting value. The paper reads these findings beside work on second-order decisions, precommitment, preregistration, the pedigree of numbers, zero numerators and precaution. Its contribution is a measured account of how one authored record set writes reversal rules in advance, which way those rules lean, and how openly it marks the standing of their numbers.