Chapter 2. Method: The Ledger, the Frame and the Source Texts

2.1What a Skeptron case is

A Skeptron case is one record, held on Skeptron's site and fetched for this study at a stated time. Every record carries the same furniture: a studio, a kind — here always public-record — a title, an event period written as a date, an as-of date, a list of fields, a list of sources, and five questions in a fixed order, each with an Answer.

The five questions run as follows. The first asks for the date of the record: which date the public source gives for the event, how unverified months, days, times, coordinates and personal information are kept out, and how the event's date is kept separate from the date at which a document was written or published. The second asks for a concrete fact: what the source says happened, with the dates, figures, persons and roles it gives. The third asks for the boundary of what can be judged: what the source does not establish, what a later reader must not calculate across, and what would have to be reconsidered if a difference appeared. The fourth is a calendar check: the weekday of the event's date computed by arithmetic, with a statement that the arithmetic and the event's reality are recorded separately. The fifth asks for the condition of a later comparison: what would have to be observed for a comparison to hold, and what would make it withdrawn.

Each question has its own Answer, and the Answers are of different kinds. Some answer from the record, giving what the source states. One is a calculation. One states a boundary and would be withdrawn if the observations behind it changed. The five questions are questions about what a source does, not about what anyone should do: none of them diagnoses a disease, and none of them judges an act right or wrong. This study reads history cases, so the record's Decision window calls for a post-hoc review, and the third question's refusal to infer is the rule the study keeps.

The record's fields are named Scenario, Answer, Unknown, Decision window and Notice. Scenario sets the scene: it names the event's period, gives the dates and figures the case turns on, and states what must be distinguished. Decision window says how the question is to be read in time — for these history cases, a post-hoc review that keeps the facts present in the original text apart from any reproduction check performed under current execution. Unknown lists what the source does not give, what must not be filled in with actual input, and what repeated reproduction does not verify. Notice says what kind of question this is: a historical verification question, in which verified records, calculations and synthetic control groups are distinguished and current extraction is not retroactively treated as a past prediction. Sources gives the public source Skeptron checked, with the publisher, the address and the date of checking.

These cases are written in Korean. This study reads them in an English translation made once for it; where the records are quoted, the words are those of the translation, and the field names are the English labels of that translation. Skeptron's Answers are Skeptron's readings of the sources it checked, and they are dated: each carries the record's as-of date, and each source carries the date it was checked. They are not statements of the historical actors, and they are not this study's judgments. Where a record's words and a public source text differ, this study gives both; Chapter 4 does this for Snow's Broad Street account, where Skeptron's reading and Snow's own text of 1855 sit side by side.

Two habits of the studio run through every record, and both are explained in section 2.3. The fourth question's calendar check says nothing about the event: this study does not quote it as evidence and does not read it as such. The fifth question's closing sentences are the studio's recurring method rather than the case's own finding.

One thing must be said once, here. This study recommends no option. It reads records of acts on disease to diagnose what each record establishes at the hour its act was taken, what the act costs and to whom, what evidence came later, and what the record leaves unsaid. It does not decide which act was right, and it gives no medical advice.

2.2The ledger

The series' cases are read through a ledger: one row per case, one cell per axis. The series' six axes run as follows, each in one sentence. Who decides: which person or body the record shows as having decided, and whether the record names them or a later reader must infer them from the setting. Grounds and what is missing: what reason the record gives for the act, and what it lists as missing — the unknowns, the unverified times, the absent raw data. Deadline, horizon and who bears the wait: by when the decision had to be taken, how long its effects run, and who waits and pays while it is taken, delayed or not taken at all. Whether the act can be undone: whether the restriction or the act can be reversed, what it allows or stops, and what reversal would cost. What would change it and whether that could be observed: whether the record states a condition under which the judgment would be revised, and whether anyone could see that condition arise. The time of the record and of the judgment: whether the record keeps the date of the event apart from the date of its document, its count or its official recognition.

Three more cells complete the sheet. Divergence: where judgments differ, is material missing, and can the missing material be named? The human's place: where a decision is left to a machine or an automatic rule, is a human's approval, exception or review named, and is that place empty or a name only? The four higher goals: correctness and safety, evolvability and resilience, clarity and accountability, long-term value — reported here only as far as the records support any comparison of options, and the ledger's options cell holds no options at all, since both codings found none that reduce the immediate loss.

Each cell holds one of three kinds of value. Some values are copied: fields of the record itself — its event period, its as-of date, its listed unknowns, its sources, its Decision window, its frame. Some values are counted by machine: the ledger's shares and counts over the eighty-nine cases, each reported with what it counts. The rest are read by coders: the coders' values for who decides, who bears the wait, the horizon, the reversal cells, the human's place and the reason recorded, together with the seven cells this study registers in section 2.4.

The codings are two independent codings of one model — one model run twice, not two models and not two model families. Each coding gives, for each cell, a value, a basis (a field of the record, or something the setting implies) and a reason. Where the two codings chose the same value the cell is settled; where they differed it is unsettled, and unsettled cells count as unclear in the verdicts.

Two readings of the same cells run side by side. Strict counts only what both coders found stated in the record's own words. Lenient also counts what at least one coder inferred from the setting. The coders read each record without the studio's frame — the recurring sentences explained in section 2.3 — and the first coding, made with the frame present, is reported in the ledger's further-reading section, where the readings of two further readers are set beside it.

The verdict rule, registered with the hypotheses, is this: sixty per cent or more of the cases a hypothesis applies to supports it; forty per cent or less goes against it; between those, or where unclear cases outnumber decided ones, the verdict is unclear. A share is always given with all four of its counts — supports, against, unclear, not applicable — and never alone.

Agreement between the codings is not even across the cells. Here is what the ledger records for the eighty-nine cases.

The low cells are low plainly. On the horizon the codings settled 32 of 89 cases, or 36 per cent. On the reversal of the action, 36 of 89, or 40 per cent. On the human's place, 37 of 89, or 42 per cent. On the reversal of the constraint, 38 of 89, or 43 per cent. Any comparison this study rests on those cells is one on which the codings divided, and Chapter 12 marks every such comparison. Reason recorded is the one cell where they settled almost everything: 87 of 89, or 98 per cent.

The ledger's own copied counts tell a second thing about the records rather than about the codings. Cases with a deadline in the record: 88 of 89. Cases that list what is missing: 89, with 264 items listed in all. Cases with a condition for revisiting the judgment written into the record: 0 of 89. Cases with a decider named in the record's own text: 0 of 89. Those zeros are facts about what the record form asks for, not findings about the historical acts.

2.3Skeptron's frame

Skeptron's history records carry, beside their own fields, a Frame field: a set of sentences that recur across the studio's history cases with only names and dates changed. This study treats those sentences as the studio's method, not as evidence about any case. They are set out here because a reader must know which sentences in a record belong to the case and which belong to the studio.

How the frame was found. The studio's machine examined the sentences that recur within a series of cases and kept those that recur across the cases of that series, keeping a sentence when most of its five-word runs stand in five per cent or more of the cases of that series. The frame is meant to catch method sentences rather than case content. The two records read in this chapter come from different versions of the studio's history record, and their frame fields are of the same kind.

What kinds of sentence the machine found. In the medicine, disaster and industrial histories alike, four kinds recur. First, sentences that keep the date of an event apart from the date of its document: that the chronology of the event and the creation or publication of the source are not combined into the same field, and that an unverified month, day, time, coordinate or personal detail is not entered as actual input. Second, sentences that keep an announcement apart from a result: that current extraction is not backdated into a past prediction, and that verified records, calculations and synthetic control groups are distinguished. Third, the calendar checks of the fourth question, with their statement that the consistency of the date formula is recorded separately from the evidence that the event actually occurred — a check whose weekday this study neither quotes nor reads as evidence about the event. Fourth, the closing sentences of the fifth question: what must be observed for a comparison to hold, and what would make it withdrawn.

One example, quoted from the Frame field of one record, is enough.

The sentence stands in the Frame field of the Broad Street record, and sentences of the same method stand in the frames of other cases of the series, though not always in identical wording: the MERS record phrases the separation of the event's chronology from the source's creation or release differently. It is the studio's way of reading. It is not something this case shows, and it is not counted here as a finding, as a pattern, or as a link between two cases.

Two consequences follow for the ledger. The first: the coders were told to read the case's own sentences only, and to leave the frame aside. The ledger's counts of conditions for revisiting the judgment are therefore counts of conditions written into the record's own text, not of the studio's recurring sentences; that is why they stand at zero in all eighty-nine cases. The ledger also carries a check on how much of a condition's wording belongs to the frame — thirty per cent or more — and that check has nothing to work on, for the same reason. A record form that carries no field for a condition of revisiting cannot be scored on whether it carries one.

The second: this study quotes a frame sentence only to explain what the frame is. Where it needs a case's own words, it takes them from Scenario, Answer, Unknown, Decision window, Notice or the questions themselves. Where a frame sentence and a case's own sentence share wording — as the sentence above also appears, with the same wording, among the Broad Street record's answers to its first question — the study cites the field it is reading and does not treat the repetition as corroboration.

2.4The study's own question, fields and hypotheses

This study asks its own question of all eighty-nine cases, and registers the fields and hypotheses it codes before any case was coded. The registration is dated 6 October 2026.

The question is stated in the Introduction, which sets out the research question and the limits of the evidence; this chapter registers the fields and hypotheses that code it.

Seven fields code each case on that question. Act: does the case's own text show a public act on disease or health — an order, resolution, restriction, notice, declaration or certification, a published report or count, a trial or treatment, or a compilation or gift of medicines? Kind of act: if there is such an act, what kind. Act before proof: on the case's own terms, was the act taken before the cause it addressed, or the act's own effect, was established — "yes" where the case's text says the cause or the effect was unknown or uncertain at the time, was settled only later, or cannot be settled from the record. Named gap in time: does the case's own text name a gap between the event and its notice, report, publication, count, official recognition or settled cause — two such moments with different dates, or an explicit delay? Bearer named: if there is an act, does the case's own text name who bore the harm or the cost while the act was not yet taken or not yet in force — the sick, the dead, a city, workers, patients, families? Revision in case: if there is an act, does the case's own text, apart from the studio's recurring sentences, name a later observation or finding that would change how the act should be read? Each cell can also hold "not applicable" or "unsettled", and each is coded by the same two codings as the series' cells, on the same strict and lenient readings.

Five hypotheses were registered against those fields, in these words. E1: "In most cases that show a public act, the act was taken before the cause it addressed, or its own effect, was established." E2: "In most cases that give a count, the case's own text qualifies it (as provisional or revised, or by a definition, a scope or a denominator) rather than giving it as settled." E3: "In most cases that show a public act, the case's own text names a gap in time between the event and its notice, count, publication or settled cause." E4: "In most cases that show a public act, the case's own text, apart from the studio's recurring sentences, does not name what later finding would change how the act should be read." E5: "In most cases that show a public act, the case's own text does not name who bore the harm or the cost while the act waited."

The test. Each hypothesis is judged by the rule set out in section 2.2: sixty per cent or more supports, forty per cent or less goes against, between those is unclear, and unclear also when unclear cases outnumber decided ones. Shares are given with all four counts, never alone. Strict and lenient readings are both reported. Where the codings divided, the cell counts as unclear in the verdict, and every comparison this study makes on such a cell is marked as one on which the codings divided.

Alongside the hypotheses, seven predictions were registered, on 6 October 2026 and before any case was coded, and they are given here word for word as predictions; the results are left to Chapter 12, which gives every count.

1. E1 supported (lenient), unclear strict: most acts on disease came before their cause or effect was settled.

2. E2 supported among the read cases; across the ledger unclear, since many disaster and industry cases give no count.

3. E3 supported among the read cases (they were chosen for dated gaps); weaker across the disaster and industry cases.

4. E4 supported: what would change a reading is mostly the studio's recurring sentence, not the case's own text.

5. E5 unclear: the read cases name the sick or the dead more often than the frame suggests; the disaster cases name victims.

6. The book's test passes in fewer than one case in five of those with an act.

7. E1's share is higher among the medicine-history cases than among the disaster-history and industrial-history cases.

Seven predictions, then, written before coding. Chapter 12 reports which of them held and which failed, and says plainly which predictions failed.

2.5The source texts

Some cases come with public-domain texts their actors wrote, and those texts are read beside the records. Which goes with which chapter, in what edition, and on what terms.

Thucydides's plague at Athens — the History of the Peloponnesian War in Richard Crawley's translation of 1874, Book II, chapters 47 to 54 — goes with Chapter 3, which reads his purpose of setting the disease down so that it may be recognised. Boccaccio's Decameron, the Introduction to the First Day in J. M. Rigg's translation, goes with Chapter 3 for Florence in 1348. The Ordinance of Labourers of 1349 appears in translation in Adams and Stephens's Select Documents of English Constitutional History (1901); the sentence on wages is quoted from the translation printed by Ernest F. Henderson in 1892. Chapter 6 reads the Ordinance with the record's own account of its exceptions.

Lind's A Treatise of the Scurvy of 1753, Part II chapter IV, and Jenner's An Inquiry into the Causes and Effects of the Variolae Vaccinae of 1798, published by Jenner himself, go with Chapter 4. Snow's On the Mode of Communication of Cholera, second edition of 1855, goes with Chapter 4 for the Broad Street outbreak. The Mortality of the British Army of 1858 — a reprint with tables and diagrams, held in Nightingale's collections and associated with her and with the statistician William Farr — goes with Chapter 5. Pasteur's report to the Académie des sciences of October 1885, printed in Comptes rendus hebdomadaires des séances de l'Académie des sciences 101 (1885), pages 765 to 772, goes with Chapter 4. Heo Jun's Tongŭi pogam, printed by the royal medical office in 1613 with wooden movable type, supplies its preface to Chapter 9. The Constitution of the World Health Organization, adopted in July 1946 and in force from 7 April 1948, supplies its preamble and three articles to Chapter 8. The Public Health Reports of autumn 1918 accompany Chapter 7's Philadelphia section. The report certifying the interruption of wild poliovirus in the Americas, by Robbins and de Quadros (1997), accompanies Chapter 8.

No text is used to say more than it says. Each is quoted only from the passages this study's rules permit; where a text is in Greek, Italian, Latin, French or classical Chinese, its sense is given in this study's own English and labelled as such; and where a modern author is drawn on, the author and the work are named, with the chapter or book where the reading note gives one, and the author is never quoted.

Some originals could not be had. The Edwin Smith Papyrus's 1930 translation by Breasted is not available to this study, so Chapter 9 reads that case from Skeptron's record and from the reading note on Breasted's edition alone. The Life Esidimeni ombud's report, the Minamata documents and the Korean government's MERS documents are also unavailable; Chapter 6 reads Life Esidimeni from the record alone, and Chapters 8 and 10 read Minamata and MERS from their records and the reading notes. The World Health Assembly's declaration of 8 May 1980 on smallpox has no source text with it, because the original is not in the public domain. The World Health Organization's statements on Ebola in 2014 and on COVID-19 in 2020 have none either, because the organisation's terms forbid reproducing them. Skeptron's records hold one case on COVID-19, the World Health Organization's statement of 11 March 2020; lockdowns, vaccines and national responses are not in them, and this study does not use them as evidence.

2.6Two cases judged

Two cases are now shown and coded, so that the ledger is concrete. Neither is judged right or wrong; what is read is what each record establishes at its own hour.

The Broad Street Pump. The record's Scenario gives the dates and figures the case turns on, and its second sentence reads:

Around that sentence the Scenario keeps the decline after the 1 September peak and the removal on 8 September in one field, notes that the dates of death and of onset did not match, states that population outflow and the decline before the removal must be considered together, records that Snow himself wrote it was difficult to determine with certainty whether the source of infection remained in the well when the water was stopped, and closes by saying this is a case in the history of epidemiology and that current directions for diagnosis and treatment are not derived from historical counts alone. The record's first answer fixes the adopted chronology as 1854-09-08, keeps unverified month, day and time out, and records the source's publication in 1855 as a separate notation. The adopted source is a page of Snow's 1855 book on the UCLA John Snow Archive, checked on the record's as-of date.

The MERS record. The Answer that gives the dates of notification and of the response centre reads:

That Answer is one of the record's five. Its third answer, on the boundary of judgment, refuses to say that all the final confirmed cases already existed on the notification date or that all 186 of them were reports made within Korea, and instructs that the initial notification, the activation of the response, the final tallies and each country's scope be distinguished, and confirmed infections be separated from estimated transmission paths. Its adopted source is the World Health Organization's page on the 2015 MERS outbreak.

Both cases are coded below on the study's seven fields, with the values the ledger's case-by-case table records for the twenty-three cases read closely.

In every one of the seven cells the two codings chose the same value for both records; no cell in this table is unsettled. The ledger's table gives values where the codings agreed and marks unsettled where they differed, and it does not record, case by case, whether a value was found stated in the record's own words or implied from the setting; this study does not supply that detail where the ledger does not give it. What follows is this study's reading of what the record's own words carry for each value.

For The Broad Street Pump. Act: the Scenario names the removal of the pump handle on a dated day, a public act on disease within the field's definition — this study reads the basis as that sentence, and the value stands as the codings chose it. Kind of act: the ledger does not record whether this value was found in the record's own words or implied from the setting, and this study reads the basis as the Scenario's own sentence; the removal stopped the use of a water supply, which the field groups with restrictions of movement, work and assembly. Act before proof: the record's own text states that the decline began before the removal and that Snow himself could not determine with certainty whether the well still carried the infection at the time the water was stopped; the effect and the cause were unsettled at the hour of the act, and the field is "yes". Gap named: the record keeps two moments apart with different dates — the peak on 1 September and the removal on 8 September — and the event's date apart from the source's publication in 1855; the field is "named". Count: the record gives twelve cases for 8 September, attributed to Snow's table, and does not mark that figure provisional, revised, or qualified by a definition, a scope or a denominator in its own words; the field is "settled", while Chapter 5 sets out the biases of counting during an outbreak. Bearer named: the record mentions population outflow but does not name who bore the cost while the act waited; the field is "not-named", and the ledger's own bearer cell for this case is "third-parties-or-the-vulnerable"; the ledger's table does not record whether the coders found that value stated or implied, and this study does not supply the detail. Revision in case: apart from the studio's recurring sentences, the record's own Scenario names what would change the reading — Snow's own doubt about the well, and the instruction to consider population outflow together with the decline before the removal; the field is "named".

For MERS in Korea. Act: the record names two dated public acts — notification of the World Health Organization on 20 May and the start of operations of the regional emergency operations centre on 28 May; the field is "yes". Kind of act: the ledger records notice, declaration or certification without recording whether the value was found in the record's own words or implied from the setting, and this study reads the basis as the Scenario's own sentence; notification and the activation of a response body are the acts the Scenario names. Act before proof: the record's own third question asks whether all the final confirmed cases already existed on the notification date, and the answer refuses to say so, instructing that notification, activation and the final tallies be kept apart; the field is "yes", and the basis is the record's own refusal. Gap named: three moments stand in the record with different dates — notification on 20 May, activation on the 28th, and the outbreak-end count — and the field is "named". Count: the final total is defined in the record's own words as the sum of 185 cases in Korea and one in China, with 38 deaths, and the record instructs that confirmed infections be separated from estimated transmission paths; a count defined by its scope and denominator is what the field describes, and the value is the cell marked qualified by a definition, a scope or a denominator. Bearer named: the record names countries and case counts but does not name who bore the cost while the acts waited; the field is "not-named", and the ledger's bearer cell for this case is "unknown". Revision in case: apart from the recurring sentences, the record's own text names what would change the reading — that the notification, the activation and the final tallies are separate, and that each country's scope is separate; the field is "named".

The two records also carry the series' own cells, and those must be read with their disagreement in mind. For The Broad Street Pump the ledger gives decider "unknown", who waits and pays "third-parties-or-the-vulnerable", horizon "unsettled", whether the act can be undone "easy", and the human's place "unsettled". For MERS in Korea it gives decider "unknown", who waits and pays "unknown", horizon "unsettled", reversal "unknown", and the human's place "unsettled". Because the codings divided on the horizon and on the human's place for both records, any statement this study makes about how long these acts' effects ran, or about where a human stood in the decision, rests on cells the codings divided, and is marked so. The ledger's per-case values for the series' hypotheses follow the same rule: the hypothesis that the cost of waiting falls on those nearest the point where things stop is supported for the Broad Street record on the lenient reading and unclear on the strict, and unclear for MERS on both; the hypothesis that few records write down in advance what would reverse them is recorded as supported for both, for the reason section 2.2 gives — the record form has no field for such a condition, and a form's silence is not a finding about the acts.

What these two cases show is how the ledger works, not what its acts were worth. The codings read the record's own words and left the studio's frame aside; the seven study cells are settled where both codings agreed; the cells where they divided are marked; and nothing in the ledger decides whether either act was right. Chapter 4 reads the Broad Street record beside Snow's own text, and Chapter 10 reads the MERS count beside the ledger's count cell and beside the record of the Ebola emergency committee of 2014.

2.7Limits of the method

The method has limits, and they are reported here because a comparison between records is only as good as the reader's knowledge of how it was made.

The two codings are one model run twice. They are not two models, they are not two model families, and they are not the independent judgment of two people. Agreement between them shows where a cell is easy to read from a record; disagreement shows where it is not, and neither settles what the cell should be. Where the ledger gives a value without its reason, this study says so and marks its own glosses as this study's readings.

The translation is a second reading. The records are written in Korean; this study reads one English translation of them, made once for it. A phrase that carries one meaning in the original may carry more or less in English, and the field names this study uses are the English labels of the translation rather than the records' own Korean labels. Where the records are quoted, the words are the translation's.

Skeptron's selection is not a census. The eighty-nine cases are Skeptron's own selection of its history questions, not a sample of public acts on disease at large. This study codes what it has and can say how often a pattern appears among these records, not how often it appears among acts on disease in general. Chapter 12 says what the twenty-three cases read closely represent and what they do not.

The evidence is uneven between cases. Some cases come with their actors' public-domain texts; others come with none, because the original is not in the public domain, because an organisation's terms forbid reproducing it, or because it could not be reached. Where a source text is absent, the case rests on Skeptron's record and on the reading notes, and the reading is thinner by that much. Section 2.5 lists which.

This study is written by AI. Its readings of the modern works stay within what the reading notes say; where this study reads further than a note, it marks the reading as its own, in the form "this study reads … as …". No modern author is quoted, and no author is given a quotation, a page number, a date or a claim that the note does not have. The study's own readings of the records and the source texts are interpretation, and the whole study is an expansion of Skeptron's records and the public-domain texts.

All of this is reported because the study's subject is records and what they keep. What was copied, what was counted by machine, what was read by coders, where the codings divided, which cases came with texts and which did not, and what no record's form can carry at all — all of it is part of what a later reader is entitled to weigh. The counts in this chapter come from the ledger and are given with what they count; the records are quoted exactly; the readings of the works are marked; and the judgments of the hypotheses are left to Chapter 12, which gives every share with all four of its numbers and reports the predictions that failed.

The first of the three times is next: the act, and what a record lets a later reader know of the hour it was taken (Chapter 3).