The sixteen cases share one habit: they say how far a study may be carried before they carry it. This chapter shows the habit, then shows what it leaves undone. A short account of the data and the method comes first; the two sections after it read the cases themselves.
1.1Data and method
The data are sixteen public records from GAEKJU's library, read through its public interface (a list of every question and a page for each) and saved as snapshots on 2026-10-06. The library then held 496 questions: 40 record readings, 51 outside cases and 405 synthetic scenes. Each question belongs to one of seven kinds of talk (idle talk, rough talk, a group's own code, talk that looks up, holding back, missteps and holding out) and to one of three questions GAEKJU asks of every kind: whether it is the same across domains, whether it costs performance, and whether it has changed across eras. The outside cases divide 16, 20 and 15 among those three questions.
The selection rule is one sentence: every GAEKJU question of the kind "outside" on the domain axis. It gives sixteen cases, covering all seven kinds of talk. The domain question is the question of transfer itself, and the outside kind is the one in which GAEKJU sets a published study beside a second place. The rule excludes GAEKJU's record readings, which are de-identified aggregate counts from one organisation and are not read here, and its synthetic scenes, which GAEKJU writes for comparison and which are not testimony from any trade. No synthetic scene is read here, and no fact is drawn from one.
Every case has the same parts. A title and a scene state the question and the study's setting. The evidence summarises the study by author and year, and often sets it beside counts from GAEKJU's own record. Three readings, marked A, B and C, give three ways of understanding the pairing, and the case leans to one of them. A judgment follows, dated "as of 2026-09-25". Then come a condition that would show the judgment or a reading wrong, a proposal for what to change, and the references. The sixteen cases cite 19 references to 17 distinct studies; three cases cite two studies each, and one study, Scott's 1985 work on a Malaysian village, is cited by three cases. The studies cited run from 1959 to 2012.
The records are in Korean. This study put each into English once, sentence by sentence, and labelled the English as its own translation of the Korean original, which each translation names by its hash. Every block quote below is a sentence of that English, copied exactly. Some Korean words that the cases quote from speech are given in English where they have a plain equivalent; one, nep, a brisk and dutiful "yes" used in Korean offices, is kept as written.
Three cautions follow from what the records are. First, an outside case is GAEKJU's summary of someone else's study. This study does not compare the summary with the study, and it says of a study only what the case says. Second, 14 of the 16 cases set the study beside counts from GAEKJU's own record of one organisation. Those counts belong to GAEKJU's record readings, which this paper does not read; it reports where a case uses them but does not repeat them, and the appendix leaves out the sentences that carry them. Third, a shared form can produce shared sentences. Following the series' rule, the share of each part's five-word runs that recur in at least 25 of the library's 496 records was measured, and again against the 51 outside cases alone with a threshold of 3. In both, the scenes, evidence, readings, conditions and proposals of the sixteen cases recur at 0%, and the judgments at 2%, all of it runs made of the date that closes every judgment and the verb before it. The series counts a part as formula at 30%. Nothing quoted below is the studio's boilerplate. What the form guarantees, that every case has three readings and a condition that would show it wrong, is not counted as a finding.
The counts in this paper come from a script that reads only the saved records and counts only what the Korean says in plain words. They are counts of the cases, not of the studies or of any workplace.
1.2How far each finding is carried
The plainest result is that the cases rarely carry a finding whole. They carry a likeness, and they say where the likeness stops. In 10 of the 16 cases, the evidence states what the study, or the record it is set beside, did not measure, count or see. The sentence usually comes right after the summary, as if the summary were not complete without it.
The case on banana time is the clearest example. It is an outside case, GAEKJU's summary of Roy's 1959 study of a machine work group and of a 1997 study of everyday conversation. The work group, the case says, cut the monotonous day into pieces with games and teasing, and when a quarrel stopped the games the day became unbearable. The finding is vivid, and it would be easy to carry it straight to any office where people trade jokes between tasks. The case stops itself:
Output was not measured.
The two studies go as far as saying that talk that looks pointless is everywhere. They do not speak of its relation to performance.
The phrase "go as far as" recurs; three cases use it of a study or of a likeness. It names a ceiling. Above it lies the question GAEKJU asks of every kind of talk, whether it costs the work, and the case refuses to let the factory answer it. One of its three readings puts the same point as a distinction between two claims that sound alike:
Roy did not measure output. Saying that it helped them hold out and saying that it made them work well are different things. What this case supports is only the first.
The case leans to a different reading, that the factory's games and an office's laughter "do the same work" of cutting up the day, and its judgment keeps both the likeness and the ceiling:
What remains across trades is the gesture of cutting up the day.
But this case cannot speak as far as performance. (as of 2026-09-25)
This study reads this as Cartwright's distinction between "there" and "here" written into a public record. What travels is a gesture, the cutting-up of a long day; what stays behind is the claim about output, which the study never made. Of the three cases whose judgments use the word performance, two decline in this way to carry the case that far.
The second pattern is a mismatch between what a study observed and what the second place could show. The case on laughter by team, GAEKJU's summary of a 2002 study of humour in several New Zealand workplaces, finds a likeness in that laughter differs by small unit rather than by trade. It then names the gap in one sentence:
The study read the grain of laughter, and the original counted the amount of laughter marks.
"The original" is the case's word for GAEKJU's own record, which counts words. The study looked at what laughter does, teasing or soothing; the record can only say how often a laughing mark appeared. The case on one CT and two hospitals, GAEKJU's summary of Barley's 1986 study of a new scanner arriving in two radiology departments, makes the same move, and says the distance between the two is great:
What the study saw was relations, and what the original counted was words.
This study reads these sentences through Tal Yarkoni's The Generalizability Crisis (2022). Yarkoni's complaint is that verbal claims reach further than the quantities that are said to support them. The cases turn his complaint into a rule of their own: before the claim is made, say what the quantity on each side is. A study of relations is not a study of words, and the case says so before it leans.
A third pattern concerns the kind of evidence a study is. The case on moves by direction, GAEKJU's summary of a 1980 survey in which employees in the United States wrote down the ways they had used to move someone at work, marks the survey's limits before it uses it:
These are ways people wrote down themselves, and which move got the work done better is not in the gist.
The judgment carries the limit forward:
Which move got the work done better is not answered by this survey, so the difference of direction is not linked to performance. (as of 2026-09-25)
The same caution appears where the study is a single survey. The case on unfairness and slacking, GAEKJU's summary of a 2002 survey of employees in Singapore, leans to reading slacking as a response to treatment rather than as disposition, while one of its own readings warns that a single survey cannot tell which came first. Deaton and Cartwright, in Understanding and Misunderstanding Randomized Controlled Trials (2018), say that a result applies to the sample selected for it and that extending it needs its own justification. The cases do not supply that justification; what they do is mark its absence, in the record, at the point of use. Six of the sixteen cases say in so many words that the carrying is a stretch, an assumption or an overreach. Three judgments call themselves a hypothesis or an assumption. The case on surface acting and permission, GAEKJU's summary of two studies, one of service workers who smile on the outside and one of swearing under permissive leaders in British organisations, joins the two and then demotes the result:
Neither is a study that measured the amount of swearing by trade, so this judgment is kept only as a hypothesis. (as of 2026-09-25)
The laboratory raises its own version of the problem, and one case meets it directly. The case on many hands is GAEKJU's summary of a 1979 experiment in which people shouted and clapped with less effort in a group than alone, and of Karau and Williams's 1993 gathering of 78 studies that found the effect robust but smaller when shares were visible, the work meaningful or the group valued. The case calls this the strongest evidence against the hypothesis GAEKJU holds for this kind of talk, that holding back does not necessarily mean low performance. One reading tries to keep the laboratory at a distance:
Where the loss was caught large was in meaningless tasks like clapping and shouting. In a workplace with valued colleagues and meaningful work, it does not appear at the same size.
The case does not lean to that reading. It leans to the first, that the loss is real, and its judgment refuses to explain the finding away:
In this group, the hypothesis 'not necessarily low performance' is weakest here. (as of 2026-09-25)
Here the debate on laboratories matters. Anderson, Lindsay and Bushman (1999) included social loafing among the domains in which they compared laboratory and field effects, and found the correspondence considerable. Mitchell (2012) found that correspondence uneven, with social psychology's laboratory effects the most likely to change sign in the field. The case does not cite either work, and nothing in it decides between them. What it does is place the question where those works place it, in the comparison of laboratory and field, and ask for the field studies to be taken out of the meta-analysis and looked at alone:
It is looked for in the results of taking the field studies alone out of the sample of Karau & Williams 1993, and in field comparisons after that.
Here the carried finding cuts against the studio's own hypothesis, by the case's own account, and the case is careful not to let the laboratory decide too much or too little. This study reads that balance beside Mitchell's finding that the answer differs by subfield, topic and effect size: no general claim about laboratories settles a particular transfer.
The cases thus carry findings in a consistent shape. They name a likeness, state its ceiling, and say what the study could not show. It holds across the seven kinds of talk and across studies as different as a factory study and a meta-analysis of laboratory tasks. This study reads it as a written form of what the debate on external validity asks for, kept in the record at the moment of use.
1.3One study on three questions, and where the test is placed
Careful sentences could still surround a finding that is carried anyway. A harder test is what happens when the same study is carried more than once. Scott's 1985 study of a village in Malaysia is cited by three cases, each on a different kind of talk. As the cases summarise it, the study recorded how people with little power resisted without revolt: by foot-dragging, feigned ignorance, false compliance and back-talk. If the cases merely stretched a concept, the three would read alike. They do not. The three lean to three different readings: A, B and C.
The case on resistance without words asks whether the pushback of the weak picks its field. It notes, first, how much of Scott's list is not speech at all:
Of the four, only back-talk is left in words.
The case leans to the reading that pushback changes form with the place, from the body where one cannot speak to written words where one can, and its judgment carries the concept furthest of the three:
It holds that pushback does not pick its field.
The case on false compliance and short answers puts the same concept beside a different question, whether two registers of office talk are false compliance. It starts by saying what Scott's study did not do, and calls the reading its own:
This study did not count tones of speech or the frequency of words.
It leans to the reading that calling it resistance is too much, and it names the evidence that would be needed and the assumption that remains:
To speak of false compliance there would have to be pairs in which a promise made upward and talk made below diverge, but the aggregate values hold no such comparison.
The judgment adds that carrying a concept from the village to the office is the card's own assumption. One of its readings states the difference between the two settings in plain terms:
The shape is the same but the weight is different.
The case on yes said upward asks whether the brisk nep said to a superior is the weapon of the weak. It too begins by marking distance:
Reading the office's nep as false compliance is reading one field over into another.
It leans to a third reading, that the item in Scott's list nearest to an office is not dutiful agreement but back-talk, and its judgment narrows the concept to that one item:
Reading all nep as false compliance is too much. Nep was common between colleagues too. Where Scott's frame fits is the side of the rare back-talk that drops the honorific. What village and office share goes as far as that, where power is tilted, talk splits into the outward and the behind. (as of 2026-09-25)
Set side by side, the three cases do something the debate on concepts discusses. Collier and Mahon write that applying a classical framework too strictly to a category built on family resemblance can lead to abandoning it prematurely or modifying it inappropriately. This study reads Scott's four forms, as the cases give them, as a family of that kind; the cases do not say so. On one question the whole family is carried, as pushback that changes its medium. On another the family is held back, because the evidence that would show one of its members, false compliance, is not in the record. On a third only one member, back-talk, is carried and the rest left in the village. This study reads that as adaptation case by case, close to what Collier and Mahon describe; it is this study's reading, and the cases themselves do not cite the debate. The same study, carried by the same studio in the same form, does not produce the same answer three times.
This study reads another case through Gary Johns. Johns, in The Essential Impact of Context on Organizational Behavior (2006), defines context as situational opportunities and constraints that affect both the occurrence and the meaning of behaviour. The case on wards that report more is GAEKJU's summary of Edmondson's 1996 study, in which, as the case puts it, wards with good relations and leadership reported more medication errors because they spoke of them more. The case leans to the reading that a count of spoken mistakes is a temperature of speaking in both places. One of its other readings starts from what the ward reports were for:
The wards' reports were records meant to protect patients.
The judgment then sets a limit of level, not of place:
This thermometer is used only at the level of the organisation and the place, and not at the level of the person. (as of 2026-09-25)
The ward study compared wards; the case will not let it be carried down to people. That is a transfer limited by unit. This study reads it beside Henrich, Heine and Norenzayan's warning against calling a phenomenon universal from one subpopulation; their note speaks of populations, not of units of analysis.
If the cases carry findings with this much care, the question that remains is who decides whether a carrying holds. Each case answers part of that question in the sentence that says what would show its judgment wrong. Fifteen of the sixteen such sentences state what would turn the judgment or one of the readings: the judgment is wrong, the reading collapses, a reading steps back, another is right. The cases also say where the test would be made. In 13 of the 16, it is made in GAEKJU's own record of the organisation the finding was carried to; in 8, it calls for further outside studies; 5 need both. Every case places its test somewhere, and most place it here rather than there.
The case on false compliance and short answers is typical in the care with which it names the test and its conditions:
It is checked if the side that has access to the original text counts only the number of such pairs, with people's names removed.
The case on copier stories, GAEKJU's summary of a 1991 rereading of a field study of copier technicians, names what the second place would have to show for the study to fit, and admits that the record cannot yet show it:
Whether the know-how talk that flowed sideways carried learning cannot be seen in the original.
The case on who stays, GAEKJU's summary of a 2001 study of a grocery chain and a hospital in which people stayed for their ties to the workplace more than for their satisfaction, writes the failure of the transfer into its own condition:
If people sought out by many colleagues talked of quitting and still stayed long, the embeddedness reading stands, and if such people on the contrary left first, this study is hard to carry over to the original.
This study reads it as close to Pearl and Bareinboim, External Validity: From Do-Calculus to Transportability Across Populations (2014), where a transfer is decided by what is known about the differences and commonalities of both populations. It is also close to Cartwright's support factors, the conditions that must be present here for a cause to work here. The cases do not use either vocabulary. They do place the evidence where this study reads the debate as asking for it.
What the cases do not do is say who will look, or when. None of the sixteen conditions sets a date by which the check is to be made. One names the party that would make it, the side with access to the original text, and three require someone to read the original text, anonymised or with names removed, which only the holder of that text could do. The judgments are dated; the tests are not. A judgment "as of 2026-09-25" is left standing until someone with the data, at no named time, chooses to look.
The proposals point the same way. In 11 of the 16 cases, the proposal moves, divides or lifts a load or a share: turns at the customer counter, partners for training spread across several seniors, the first hold on new equipment passed between groups. Six refuse in plain words to dig out, score, listen in on or pass upward what people said or did. The case on many hands begins with such a refusal, "Do not dig into who did less", and the case on resistance without words says the same of the system that would act on it:
The automatic judgment has no business digging out false compliance.
The proposals carry the transfer's caution into action: they change how work is divided and pass no verdict on a person. In the cases, the proposed change does not wait on the test that would reopen the judgment; the check that would confirm or undo its reasoning has no owner and no date.
To the argument about external validity, then, the chapter adds a public, dated form of transfer that keeps the study's place and the target's place apart and puts the deciding evidence in the target; to the argument about concepts, one category carried three ways. What it shows missing is the part of a decision neither debate puts at its centre: the person who will look, and the time by which they will.