Chapter 1. What a Count of Sources Decides in Ten Authored Training Cases

This chapter shows that the ten cases read one small count in two ways, as waste and as doubt, never as a safeguard; that they recommend sorting reports by their roots and rule out the two opposite errors on either side of that; and that the sameness of source on which every count rests is the author's assumption, asked for and not held. It contributes to the debates on redundancy and on agreement as evidence a reading of where one designed record set puts the worth of a repeated report.

1.1Data and method

AEGIS publishes its authored cases as data files of one record a line, in 11 published parts. This study read all of them once, by GET, on 2026-10-06: 6,540 cases. The programme AEGIS-CALC-v1 is a set of 196 cases in AEGIS's ninth part, which AEGIS describes as cases that actually execute 12 kinds of calculation rule and keep the input hash, the output and the replay. The selection rule is one: every case of that programme whose run executes the rule source-roots-v1, which counts the distinct original sources of a list of reports. The rule keeps 10 cases. Across all 6,540 cases the same rule is run 97 times, in 23 programmes, so the ten are not the only cases that count sources; they are the only ones in the programme built around executed calculations, and none belongs to a programme that earlier AEGIS papers read.

The records are what AEGIS says they are. All ten are typed realistic synthetic (현실형 합성). All ten give the same source role: an authored question seed whose judgment mark is the author's record, not a factual basis from public records. The programme scope says every number is a value assumed by the author and has not been compared with public records, and every run says that its mapping of reports to sources is the author's assumed value. Every run has been executed and replayed with the same output; every fact check says there was no verification of real events or authority; every prediction performance is "Not measured"; none has a source link. Each case stands at "Partial calculation complete", waiting for the rest of its judgment materials. The deciding roles are 10 different officers, three of them marked as a setting, across six domains, five of them filed under disclosure and accountability.

Each record was saved as a snapshot in this study's folder with two fields removed, the source line and the group label, because some AEGIS records name a model there. The snapshots were checked for model names and none remain. The records are Korean. This study wrote the English once, sentence by sentence, from a table of translations kept with the study, and each English seed names its Korean original by hash. Every block quote below is that English.

The fields read are these: the situation (context, which AEGIS repeats as the summary), the question, the three choices, the author's judgment mark (review), the needed inputs, the unknown, the condition for reconsidering (reconsider), the observation, the run's input, output and interpretation, and the link to another case. Every classification is a stated rule on the Korean words, applied by a script, and every number comes from that script. A question is counted as asking for efficiency when it has words such as 단순화 (simplify), 간소화 (streamline), 갈래 (strands) or 묶어 (bundle); as asking about independence when it has 독립 (independent); and as asking for a standard when it has 기준 (standard) or 최종 근거 (final basis). An excluded choice is counted as collapsing the reports when it has 한 건으로 (into one), and as leaving them many and unsorted when it has 그대로 (as they are), 현행대로 (as at present), 맡긴다 (leave to), 독립 의견처럼 (as if independent opinions) or 표기 없이 (without marking). The rest is close reading.

Before any field is used as evidence about the set, it is measured for frame: the share of its five-word runs that recur in at least 5% of AEGIS's items. The situations, questions, choices, needed inputs, unknowns, conditions for reconsidering, observations, run interpretations and link reasons are free or nearly free of frame, at 0% to 7% in Korean or English, and the author's marks reach 10% only in English. Three fields are frame and are not read as findings: the uncertain line (60% in Korean), and the condition and counter-evidence attached to every link, which are the same sentences across almost all of AEGIS's links (99%, and 100% among the ten). The programme scope and source role are one sentence each, repeated in all ten; they are read once, as what AEGIS says the records are. Two recommended reasons are the same sentence word for word, and are read as one. That every case has three options and a condition for reopening is guaranteed by the template and is not a finding.

1.2One count, read as waste and as doubt

Everything in this section is read from authored synthetic cases, whose reports, offices and figures are the author's assumptions. The arithmetic is the same in all ten. A short list of reports is tagged, report by report, with the source each rests on; the run counts the reports and the distinct tags. Six of the ten come out at four reports and three sources. The first case states the whole of it in three sentences:

Nothing in the situation says what the overlap means. The meaning arrives with the question and the author's reasons, and it arrives in two forms. In this case, and in the cases of maintenance performance, fuel consumption, scattered situation reports and the alliance consultation's documents, the question asks how far the reporting can be simplified. The author's reason for the recommended option makes the overlap a warrant for removal:

The same sentence, word for word, stands behind the recommendation in the case of maintenance performance reports. The reasons given against the alternative in these cases speak the same language. The alternative of confirming the sources while keeping the present reporting is marked down because it saves nothing:

And in the consultation case, where three documents rest on a single agenda meeting, the recommendation to issue one text is justified by effort, not by evidence:

Read this way, a report beyond its root is a cost. It occupies a strand of reporting, a format, a review. Four of the cases give the burden of reporting as their reason for preferring one option over another, and the conditions for reopening in four cases are written in the same key: the matter is to be reopened if the reports double, if half of them are rewritten, or if the sources come to exceed half the reports. In this study's reading, the record set here takes up the complaint about duplication and overlap that Martin Landau, Redundancy, Rationality, and the Problem of Duplication and Overlap (1969), sets beside redundancy, and takes it without that counterweight. The opening of Landau's article, with a plane brought down safely on an impaired rudder because other features of the aircraft could stand in for it, treats a second channel as the source of reliability. No case in the ten makes that argument for a second report.

In three cases the same count carries a different meaning. The cases of press releases, training accident reports and historical records ask whether reports that share a source can be taken as independent of each other. The accident case states the doubt plainly:

The case of historical records gives the principle behind it:

Here the report beyond its root is not a cost but a false witness. It adds a voice without adding a source, and the risk is that it will be heard as confirmation. The author's reason for ruling out the accident case's last option names that risk:

This is the worry of the second literature. Krishna K. Ladha, The Condorcet Jury Theorem, Free Speech, and Correlated Votes (1992), shows what happens to the promise of a majority once votes are correlated, and finds the conditions severe for small groups; Alvin I. Goldman, Experts: Which Ones Should You Trust? (2002), holds that going with greater numbers depends on problematic assumptions. The accident case, with four reports, is a small group in exactly this sense, and its doubt is the one these works describe. The situation of the scattered situation reports puts the two readings in a single sentence, in which the independence of documents falls as far as their sources overlap; its question then asks only how far to merge them.

The two readings rest on one count and lead to different handling. Read as waste, the overlap is cleared: one strand or one representative document a source, and the rest turned into references. Read as doubt, the overlap is kept and marked: the root is written beside each record so that a reader can tell which records are copies of which. Neither reading keeps the repeated report for its own sake. Two recommended options come nearest to it. In the accident case the two reports from one office are to be cross-checked separately, and in the supply case the maintenance and inspection records are kept as supplementary checks on the ledger that is made the default source:

Both keep a second record because it may disagree with the first, not because it can stand in for the first. In this study's reading, that is the use the testimony literature makes of a second source, a check on the first, and not the use the engineering literature makes of a second channel, a substitute when the first fails. A. Avizienis, The N-Version Approach to Fault-Tolerant Software (1985), reviews an approach built on the second use; the ten cases are written entirely within the first.

Martha S. Feldman and James G. March, Information in Organizations as Signal and Symbol (1981), offer a reason why piles of reports might grow regardless: organisations gather more information than they use and keep asking for more. The ten cases begin where such a pile already exists, and their recommendations shrink or label it. They do not ask what the pile was for. In this study's reading, that silence is part of how the set frames the count: a report beyond its root is to be explained away as waste or as dependence, and no case allows that it might have been gathered for another reason.

1.3What the count cannot hold

Everything in this section, too, is read from authored synthetic cases. The calculation in each case is a count of labels. Its input is a short list in which every report already carries the name of its source, and each run says where those names come from. The fuel case says so most fully:

All 41 labels in the ten inputs are of this kind. The calculation can be replayed and gives the same result each time, but what it replays is the author's decision about which reports share a root. The judgment that a source-counting rule exists to support, that two reports are or are not the same source, is made before the rule runs, and made outside the record.

The cases know this. Nine of the ten ask, among the inputs still needed or the things still unknown, for something that would ground the labels: a register of original sources in three cases, a document defining the sources, grounds for fixing them, a record of confirming them, a comparison table, criteria for judging them. The case of historical records asks for exactly the judgment the count presupposes:

The press release case asks for the evidence behind the labels:

And the scattered reports case lists the same criteria among its unknowns:

Six cases also ask who wrote the reports or how they were written. That matters to the reading of independence more than to the reading of waste. Two reports tagged with one office may have been copied from one file, or written by two people in that office who saw different things; two reports tagged with different offices may both rest on a figure that one of them passed to the other. The supply case gives an instance of the second kind of trouble, where different sources yield one figure by different routes:

The literature on multiversion software is about exactly this gap between a label and a dependence. D. E. Eckhardt and L. D. Lee, A Theoretical Basis for the Analysis of Multiversion Software Subject to Coincident Errors (1985), model the propensity of separately written versions to fail together, and John C. Knight and Nancy G. Leveson, An Experimental Evaluation of the Assumption of Independence in Multiversion Programming (1986), find that versions prepared independently from one specification failed together in substantially more tests than expected. In this study's reading, the lesson for the ten cases is that a distinct tag does not make a distinct source, and that the dependence that matters is a degree to be estimated. Luc Bovens and Stephan Hartmann, Bayesian Epistemology (2004), treat the reliability of information sources as something to be modelled, and Michael Weisberg, Robustness Analysis (2006), describes the search for predictions common to several independent models and defends the degree of confirmation it can give. The ten cases have no place for such a degree. A report shares a root or it does not, and the count follows.

Two of the cases go further and treat the count itself as a measure of independence. The case of historical records reads its result as a fraction:

And the scattered reports case sets a share as its condition for reopening:

By the case's own figures, three sources for five reports already exceed half, so the condition as written is met at the moment the case is set; the case gives no sign of having noticed. In this study's reading, which goes beyond what Ladha's article says, a ratio of sources to reports says little about how far the reports are correlated, which is what that result turns on. Across the ten, 27 sources stand behind 41 reports, a ratio of 0.66, and the ratio is the same whether the reports beyond their roots are copies or separate observations.

What the cases rule out is clearer than what they recommend. In three cases, all of them read as waste, the excluded option is to bind the reports into one, and the reason is that the basis can no longer be recovered:

In the other seven the excluded option leaves the reports many and unsorted, quoted as they are, scattered, unmarked or distributed as if independent. The consultation case names the error those options share:

The two errors are each other's opposite. One loses the sources by merging them; the other multiplies a source by leaving its copies unmarked. Every recommended option names the original source, a representative document or a default as its unit, which places it between the two. The consultation case shows that the unit is the root and not a number of documents: with three documents on a single root, the recommended option is the very collapse into one that three other cases rule out, because here one is the number of roots. In this study's reading, the set holds a consistent position: neither the pile nor the single report is the unit, the root is. What it cannot do inside its own records is say what makes a root.

This section adds to the debate on agreement as evidence a reading of how a template handles dependence. The cases see the danger that Ladha and Goldman describe and guard against it; they also see the opposite danger, of losing track of which source said what, and guard against that too. They do both with a binary label set by the author, where, in this study's reading, the literature from Eckhardt and Lee to Ladha treats dependence as a degree that has to be estimated.