The paper asked of twelve authored AEGIS training cases what a written condition for reversal commits the decider to, and what the record says about where the condition's numbers come from. Each condition commits the deciding council or committee to a count over a small test named in advance, run before approval and recorded by a single officer, with a tolerance of one failure or none. The rules lean in one direction: in the four cases whose recommendation is the more restrictive choice, only a clean result reverses it, and in the eight others a single failure, or the failure of a single item, does. The costs of every option are written beside it, so the lean is a presumption chosen with the costs in view rather than a blindness to them. And every record says where its figures come from, whether the earlier text it corrects, the author's additions or design values, and five say that something in that text was not adopted. Because all twelve cases are authored synthetic settings, these are findings about how the AEGIS record set frames reversal, not about how any real body decides.
Discussion
On whether to bind a future decision, the records side with binding. Sunstein and Ullmann-Margalit list rules and presumptions among the second-order strategies; Elster argues that people may benefit from constraints on their options. The AEGIS conditions are rules of this kind, fixed in sample, count and consequence before any trial is run. The preregistration argument of Nosek and others gives the same structure a reason drawn from science: a condition stated before the outcome cannot be bent to fit it. The records add a choice of direction: they state beforehand whether a single failure or a clean sample is decisive.
On caution, the records occupy a middle position that neither side of the debate quite describes. Sunstein, in Laws of Fear, argues that precautionary steps create dangers of their own and asks for close attention to costs and benefits. The records meet the first half of that demand: each of their thirty-six choices names what it gives up, including the restrictive ones. They do not meet the second half in the form of a balance. Instead of weighing the named costs against one another when the test result arrives, they decide beforehand that a restrictive recommendation stays until the sample is clean. On this paper's reading, that is a presumption in Sunstein and Ullmann-Margalit's sense, placed on the side of caution, and it is the most distinctive thing the set does.
On small samples, the records show awareness of the problem they create. Hanley and Lippman-Hand ask what a study with no adverse events can show, and note how differently zero and one are treated. The clean-result rules rest on exactly that difference, over samples as small as three. Yet one record says that its three trials do not adequately represent the condition they test, and nine of the twelve conditions close with a sentence refusing a generous reading, such as treating the absence of anomalies as a finding of safety. The records do not solve the zero-numerator problem; they decline, in writing, to let an unconfirmed quiet stand in for a confirmed clean result.
On the standing of numbers, the records do two things that the literature tends to keep apart. Porter describes quantification as a strategy of impersonality where trust is short, and the reversal rules are impersonal in that sense: once written, they need no one's judgment to apply. Van der Sluijs and others describe pedigree analysis, a qualitative review of how strong each number is. The records give their numbers a pedigree in plain words, inherited, added or declined, and so tell the later reader that the impersonal figure is an exercise value. Dewar's assumption-based planning asks that vulnerable assumptions be brought into the open and monitored; the records do the first and leave the second to the officer who records each test. What they leave unwritten is how an exercise value would become a tested tolerance, and every record says its predictive performance was not measured.
Contributions
The paper makes three contributions. It gives a measured description of how one authored record set writes reversal rules in advance: counts over named samples of three to twenty, a tolerance of one or none, tests run and recorded before approval by an officer other than the deciding body, and a lean that keeps restrictive recommendations until a clean result while letting conditional ones fall at the first failure. It shows that the same records mark the standing of their numbers, naming inherited figures, added exercise values and design values, and in five records inherited claims they declined, a practice found in about 1% of AEGIS's authored cases. And it places the set within the debate on deciding how to decide, showing that the records side with precommitment and preregistration on when to state a reversal, choose a cautious presumption over a balance of named costs on how to apply it, and supply in their own pedigree statements the caution about precision that the literatures on numbers and on zero numerators ask for.
Limits
The limits are those of the material. The cases are authored synthetic training cases; the study can say how they frame reversal, not how any real decision is reversed. The selection is one programme of twelve cases, chosen because its scope announces the features the paper reads, so the paper cannot say how common those features are in AEGIS except by the counts it gives against the whole authored library. The classification of rules as single-failure, item-by-item or clean-result, and of recommendations as more or less restrictive, was made by stated rules on the Korean wording; the four clean-result cases are few, and a different reading of what counts as restrictive would move the line in at least one of them. The English is this study's own translation, made once and labelled; a different translator might have chosen other words, though every quote can be traced to the Korean by the hash each English seed carries. The frame measure undercounts in Korean, which is why it was repeated in English among the twelve; a measure over more English records would be firmer. Finally, the records describe themselves as corrections of an earlier list that this study did not read, so it cannot say what the corrections changed beyond what the records themselves state.
Further research
Three lines of work follow. The first is to read the other AEGIS programmes and sets whose records state the sources of their numbers, which the counts list, with the same questions, to see whether the lean toward caution and the plain-word pedigree hold across them or belong to this programme. The second is to compare these rules with reversal conditions elsewhere in the RISCON family that name no numbers, using the same frame measure, to see what a record gains and loses when it turns a condition for revisiting into a count. The third is to ask of any reversal rule the question these records leave open: by what evidence would an exercise value be replaced by a tested tolerance, and who would decide that it had been. A record that wrote that path beside its threshold would show not only what would reverse the decision, but what would change the rule that reverses it.