Chapter 1. What a Record Can Carry in Fifteen Authored Training Cases

This chapter shows that the fifteen cases let a record pass into a present decision only on stated conditions, that they rule out the neglect of a record more often than its copy, and that the original on which every condition rests is absent from the record set. It contributes to the debate on lesson-drawing and on the use of history a reading of where one designed record set places the risk of that passage.

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-NEWDOMAIN-v1 is a test set of 18 cases in AEGIS's eleventh part, which AEGIS describes as checking the form, calculation and reproduction of nine new domains. The selection rule is one: every case of that programme whose situation (the field context) names a public or declassified record, by the Korean words 공개 or 해제. The rule keeps 15 cases and leaves out three, whose situations start from a pilot result, the spread of wargame results and a reassessment ordered at a meeting. None of the fifteen belongs to a programme that earlier AEGIS papers read. Across all 6,540 cases, 1,176 situations use one of the two words, so the fifteen are not the only cases that touch public records; they are the only programme built around them.

The records are what AEGIS says they are. All fifteen are typed realistic synthetic (현실형 합성). All fifteen 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 of all fifteen says that every figure, question and actor is the author's assumption and refers to no public record and to no real person, unit or facility. Every run says there was no verification of real events or authority, and every prediction performance is "Not measured". None has a source link. Thirteen runs wait for input; two have executed a small bounded calculation on the author's assumed counts. No case carries a completed decision. The deciding roles are 15 different officers and analysts, none marked fictional, across nine domains.

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), the question, the three choices, the author's judgment mark (review), the needed inputs, the unknown, the condition for reconsidering (reconsider), the observation, the uncertain line, the programme scope 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 choice is counted as holding the record back when it has words such as 유보 (withhold), 보류 (hold), 참고 (reference), 보존 (preserve), 인용하지 않는다 (does not quote) or 남긴다 (leaves); otherwise as carrying it over whole when it has words such as 그대로 (as it is), 전체 (whole), 전면 (in full), 직접 (directly) or 동일 (equally). 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, choices, needed inputs, unknowns, conditions for reconsidering, observations, link reasons and author's marks are nearly free of frame among the fifteen, at 0% to 9% in Korean or English. Three fields are frame and are not read as findings: the uncertain line (58% 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%). The programme scope and source role are one sentence each, repeated in all fifteen; they are read once, as what AEGIS says the records are. That every case has three options and a condition for reopening is guaranteed by the template and is not a finding.

1.2What passes from a record, and what is ruled out

Everything in this section is read from authored synthetic cases, whose public records are the author's assumptions and not real sources. The fifteen cases do not ask whether a record is true. They ask how much of it may cross into a present decision, and they answer by separating layers. The first case, set in a war-history domain, gives the separation its plainest form. Its situation reads:

The sentence divides one source into two objects. What the case says its history reports, that supply lines were cut and rear logistics disabled, is not what the author puts in question. What is in question is the lesson, and the lesson is placed on the side of interpretation, where it can be right for one set of conditions and wrong for another. The author's recommended option then draws the line where the conditions match:

This is not a refusal of the lesson. It is a narrowing of it to the cases that share the condition under which it was learned. The other cases repeat the move in their own domains, and in 11 of the fifteen situations the turn is explicit: the record is named, and then a concession follows. In the manual case, the clause on front-line maintenance is said to rest on assumptions of its own time:

In the classic-text case, it is the age of the text:

And in the memoir case it is the kind of source:

Eight of the fifteen situations name a condition of the record's own making in this way: its assumptions, its selection, its era, its deletions, its reliance on memory. Seven say what the record cannot settle on its own. The wargame case gathers the whole argument into three sentences, and it is worth reading entire:

The first sentence reports a result. The second says what the result is made of, and lists the fixed inputs as plainly as a method section would. The third does not reject the result; it ties its reach to the match between those inputs and the world. The author's reason for ruling out the option of quoting the conclusion directly in a plan states the principle behind all three sentences:

This is where the cases sit in the debate. In this study's reading, the move is close to the one Cartwright and Hardie make in Evidence-Based Policy: A Practical Guide to Doing It Better (2012), where predicting a policy's outcome is betting that the situation has the support factors that let the policy play its causal role. The cases name the support factors in their own terms: a single route, skilled maintainers, troop numbers of a given age, fixed travel times. It is also close to the context that Michael Howard, The Use and Abuse of Military History (1993), a reprint of a lecture of 1961, puts beside width and depth as a criterion for studying military history. The recommended option in all fifteen cases is a transfer on conditions, and, by a keyword rule on the Korean words, 10 of the recommended reasons name a condition, a comparison, a scope or a premise in so many words.

What the cases rule out is less expected. Each case marks its first choice recommended, its second an alternative and its third excluded, in that order in all fifteen; the order is the template's. What fills the excluded place is not. By the rule of section 1.1, in four cases the excluded option is the literal transfer of the record: the whole lesson made into one rule, the manual adopted in full, the classic split in two by the letter, the wargame's conclusion quoted in the plan. In the other 11 the excluded option holds the record back, with some looseness in the rule: it is kept for reference, preserved only, never quoted, left to a later cycle, or withheld. The reasons given for excluding neglect are pointed. The declassified-documents case says:

The intelligence case gives the same reason in other words:

And the manufacturer's-figures case refuses to let the mere existence of public information become a reason to wait:

In ten of those 11 cases, the literal transfer, as the same rule reads it, has not disappeared. It stands as the author's alternative: quote all of the visible text, quote the old assessment as present fact, carry the meeting's wording as it is into the implementation documents, put the memoir's scenes into training without editing. The reasons given for these alternatives do something the choices themselves do not. Across the fifteen, seven alternative reasons open by granting the option an advantage, and several then add the very safeguards that the recommended option contains. The declassified-documents alternative is the choice to quote everything visible, but its reason ends:

The posture-assessment alternative is the choice to extract the conclusion alone, and its reason ends:

The meeting case follows the same course. Its alternative is to carry the adopted wording as it is into the implementation documents and leave execution to each department, and its reason grants the speed of that course before attaching a second step:

An alternative that quotes everything with a filter, extracts a conclusion with its assumptions attached, or carries a wording over with a report back to the body that wrote it, is no longer the literal transfer named in the choice. The reasons repair the option toward the recommendation. Read together, the excluded and the alternative places sort two errors unequally. Neglect of a record is excluded, with reasons that say it defeats the record's purpose. Copying a record is allowed as a second choice and then qualified in its own reasons until it is nearly conditional.

This ordering matters for the debate. Khong, Analogies at War (1992), and Levy, Learning and Foreign Policy: Sweeping a Conceptual Minefield (1994), locate the danger in inferences that outrun the evidence. The fifteen cases acknowledge that danger, in the four exclusions of literal transfer and in every recommendation's conditions, but they weigh the opposite danger more heavily. In this study's reading, their design is closer to Rose, What is Lesson-Drawing? (1991), for whom searching for fresh knowledge is not normal and is stimulated by dissatisfaction with the status quo, and to Levitt and March, Organizational Learning (1988), for whom organisations learn by encoding inferences from history into routines. One case shows the encoding plainly. Where a classic's principle of concentration seems to collide with a present policy of dispersed deployment, the author refuses to let either be dropped:

The excluded option there is to remove the classic's principle from doctrine. The case would rather keep both, each assigned to its own object, than lose the older text.

The cases also describe themselves in these terms. Each is linked to one other case of the fifteen, and the reason given for every link is that the two share a structure of judgment, not a topic. Ten of the 15 link reasons describe that structure as a separating or a dividing: the context of an original kept apart from the conditions of its present use, the public status of a document kept apart from its grade of reliability, a body's decision divided into authority, procedure and follow-up. The link from the classic-text case says it most compactly:

A record set that links its cases by a shared act of separation has made that act its subject. What it separates the record from, however, is a present that it can describe and a past that, as the next section shows, it does not hold.

1.3The original that is asked for and not held

Every recommendation in the fifteen cases depends on a comparison: the lesson with today's routes, the manual's assumptions with today's maintainers, the published figures with an independent measurement, the declassification list with the preserved original. The cases are explicit that the comparison needs the record itself. Of the 45 needed inputs across the set, 15 name it, and in each of the 14 cases that name it the record comes first. The classic-text case asks for:

The memoir case asks for:

The one case that does not name the record among its inputs is the case on committing an elite unit, whose public practice is described as a general rule rather than as a document; its needed inputs concern present conditions only.

None of these originals is in the record set. No situation gives the record a date, and none names a publisher or a place of release; the two classics are named by title alone, one title carrying a personal name. No case has a source link. And AEGIS states the matter outright in every case:

The sentence is part of AEGIS's frame, repeated in all fifteen, and it is not a finding about any one case. Its effect on the set is a finding. The cases describe a war history, a manual and a memoir; the programme says those descriptions refer to nothing public. The comparison with the original that each recommendation requires is therefore a comparison with an original the author assumed. The cases are honest about this at the level of the whole set. The uncertain line repeated in each says:

Inside the cases the absence leaves marks. The unknowns of six cases concern the making of the record itself: whether a bias entered when the lesson was drawn, how the manufacturer's measurement was calibrated, what the deleted items said, what the assessment assumed, how memory differs from the official record, how the intelligence was collected. The first case's unknown pairs the record with the present in one line:

These are the questions of source criticism, and they can only be answered from the original. Four of the conditions for reconsidering are likewise triggered by something found in the record rather than in the world: a notice of re-restriction, a conflict with the official record, a change in the assessment's conditions, an item without grounds of cross-quotation. The intelligence case's condition shows the direction of all four:

The record set also disagrees with itself about how a figure taken from a record should be held. The manufacturer's-figures case ends its recommended reason:

The life-extension case ends its recommended reason the other way:

Both sentences are the author's, written into synthetic cases of the same programme. The first speaks as if an original were present and could anchor a figure; the second speaks as the programme scope does. The two cases that run a calculation sit on the same seam. The posture case divides one count by another and states what the counts are:

The counts are assumed, and they are said to reproduce a public assessment that, by the programme's own statement, is not referred to. The arithmetic is exact; what it is arithmetic of is the author's setting.

This is the limit the cases reach. Everything above rests on assumed records, and none of it says what any real record holds. March, Sproull and Tamuz, Learning from Samples of One or Fewer (1991), describe the balance between agreeing on an interpretation of history and interpreting it correctly. The fifteen cases are designed to secure the first: a shared procedure for any record, with conditions, inputs and reopening rules. They cannot reach the second, because the history is not in them. In this study's reading, that is what a training template is for, and also the place where its teaching stops: it rehearses the comparison with an original and leaves the comparison itself to whoever holds one.