The ten AEGIS cases are authored synthetic training cases, and nothing in this paper says how any real office counts or weighs its reports. What the paper shows is how one designed record set frames the step from a pile of reports to a judgment about its worth. The set counts sources exactly and reads the count in two ways; it recommends the original source as the unit and rules out the errors on either side of it; and it leaves the judgment of sameness, on which every count depends, to an author's label and to inputs it asks for and does not hold.
Discussion
The first finding is about meaning. The same arithmetic, four reports on three sources in six of the ten cases, is read as waste in five cases, as a doubt about independence in three and as a question of which source is the standard in two. These readings are not in conflict, and one case holds two of them in a single sentence. But they lead to different handling: waste is cleared by merging reports into one strand a source, and doubt is answered by marking the source beside each report. What is absent is the third reading the literature offers. Martin Landau, Redundancy, Rationality, and the Problem of Duplication and Overlap (1969), opens on the reliability of a redundant aircraft, and A. Avizienis, The N-Version Approach to Fault-Tolerant Software (1985), reviews an approach that runs several versions so that the faults of one can be tolerated. No case in the set treats a report beyond its root as a second channel worth keeping. Where a second record is kept, it is kept as a check that may disagree, which is the use a weigher of testimony would make of it.
The second finding is about symmetry. The set guards against error in both directions. Three cases rule out binding different sources into one report, because the source behind a figure could no longer be recovered; seven rule out leaving copies of one source to pass as many. The recommended option in every case sits between the two, and the consultation case shows that the unit is the root rather than any fixed number of documents: there, with one root, a single text is recommended. In this study's reading, that is a coherent position, and it takes something from each literature: from the epistemic side, the warning of Krishna K. Ladha, The Condorcet Jury Theorem, Free Speech, and Correlated Votes (1992), that the advantage of a majority assumes independent votes, and of Alvin I. Goldman, Experts: Which Ones Should You Trust? (2002), that going with greater numbers depends on problematic assumptions; and from the side of record-keeping, the demand that a figure stay traceable to where it came from.
The third finding is about what a template can hold. The calculation counts labels, and the labels are the author's. Nine cases ask, as something still needed, for the register, definition, grounds or criteria that would make the labels a finding rather than an assumption. Two cases go further and read the ratio of sources to reports as a measure of independence, one in its run as independence fallen by half, one in its condition for reopening as a threshold that its own figures already pass. In this study's reading, the literature that has studied dependence most closely, from D. E. Eckhardt and L. D. Lee, A Theoretical Basis for the Analysis of Multiversion Software Subject to Coincident Errors (1985), to John C. Knight and Nancy G. Leveson, An Experimental Evaluation of the Assumption of Independence in Multiversion Programming (1986), and Luc Bovens and Stephan Hartmann, Bayesian Epistemology (2004), treats it as a degree to be modelled or measured. A binary root cannot carry a degree. The cases ask for what would ground their labels; as written, they do not hold it.
Martha S. Feldman and James G. March, Information in Organizations as Signal and Symbol (1981), suggest a reason why such piles exist that the cases never consider: that information is gathered beyond use because its use is embedded in social norms. A case that read the pile that way would ask what the duplicate reports were for before deciding whether to clear them. None of the ten asks.
Contributions
This paper makes three contributions to the debates on redundancy and on agreement as evidence. First, it shows that a designed set of decision records can absorb the administrative complaint about duplication and the epistemic worry about dependence while leaving out the engineering case for redundancy: in the ten cases, a report beyond its root is waste or a false witness, and never a safeguard. Second, it shows that the set holds a symmetrical position, guarding against the loss of sources as firmly as against their false multiplication, with the original source as the unit of every recommendation. Third, it locates the limit of that position in the gap between a count that can be run and replayed and a sameness that has to be judged: the independence the cases count is binary and assigned by the author, where the literatures treat it as a degree, and the records ask for what would ground it without holding it.
Limits
The cases are authored synthetic settings. Their reports, offices, ledgers and figures are the author's assumptions, and the study draws no conclusion about real reporting, real records or real decisions. The set is small: ten cases from one programme, chosen by one rule, out of 97 cases across AEGIS that run the same calculation. The classifications rest on keyword rules applied to the Korean records; such rules can miss a reading or force one, and the rules are stated beside every count so that a reader can check them. The English is this study's translation, made once and checked against the Korean, but a translation is a reading. Some fields are frame, repeated across AEGIS's records, and were not read as findings; two of the recommended reasons are the same sentence. And the readings through the works are this study's interpretation, drawn only from what their abstracts or opening passages say.
Further research
Three lines of work follow. The other 87 cases that run the same calculation, spread across 22 further programmes, could be read with the same rules to see whether any of them reads a repeated report as a safeguard, and whether the two opposite errors are ruled out as evenly. The programme's other rules that touch dependence and overlap could be read beside this one, to see whether the set treats shared resources and shared sources alike. And a version of these cases that held the register or the criteria they ask for would allow a test this study could not make: whether the author's labels of sameness survive a comparison with how the reports were actually written.