12.1The rule in brief
This chapter gives the verdicts. Every count in it comes from "The ledger's numbers": the counts of the two codings, the shares as they stand there, and the verdict rule as Chapter 2 set it out. Nothing here was computed again from the records.
Eighty-nine cases were coded, from Skeptron's medicine history (17 cases), its disaster history (47) and its industrial history (22). Each case was coded by one model run twice, producing two independent codings; each cell carries a value, a basis (stated or implied) and a reason. A cell is settled when both codings chose the same value, and a cell where they differ counts as unclear in every verdict. The strict reading counts only what both codings found stated in the record's own words; the lenient reading also counts what the setting implies.
The verdict rule was registered before any case was coded: a hypothesis is supported when 60% or more of the cases it applies to supports it, opposed when those cases go against it in 60% or more of cases, which is to say when the supports-share stands at 40% or less, and unclear in between; a hypothesis is also unclear when the unclear cases outnumber the decided ones. Two groups of hypotheses are kept apart. E1 to E5 are this study's own, registered on 6 October 2026 before any case was coded, each with a prediction written beside it at the same time. The series' hypotheses are the ledger's, drawn from the series' axis on decision records; seven are judged on this ledger and one is read across the books in Volume 0 as well as here. This chapter reports the results, including the predictions that went wrong: the rule does not change when the count runs against the expectation set beforehand.
12.2E1 to E5, strict and lenient
For each hypothesis, the prediction registered before coding is given first, then the strict result, then the lenient result, both with all four counts — supports, against, unclear and not applicable — then whether the prediction held. Five predictions stand beside E1 to E5, one more beside the test, and one beside E1's spread across the three groups of the ledger; seven in all, in the order of section 2.4. Five of the seven failed on the registered reading, two held, and this section names which below. A prediction registered in two parts is counted as failed if either part failed, and the section says which part held. The shares over E1, E3, E4 and E5 run among the cases where the two codings agree a public act is shown, thirty-six of the eighty-nine, against thirty-nine where no act is found and the hypothesis does not apply; E2 rests on thirteen cases where both codings settled the count cell, and where the strict reading supports it this section says "eight of thirteen" rather than "most".
E1. Strict: supports 8, against 3, unclear 61, not applicable 17; share 73%; verdict unclear. Lenient: supports 25, against 5, unclear 20, not applicable 39; share 83%; verdict supports. The prediction, supported leniently and unclear strictly, held on the registered reading. Not applicable stands at 17 on the strict reading and 39 on the lenient; the lenient figure of thirty-nine is the count of cases where the codings found no public act at all, and the share in each row runs over the cases where a verdict could be given. The act-before-proof field was agreed by both codings in 69 of the 89 cases, with 17 reading yes by implication, 8 yes in the record's own words, 4 not-said and 1 no; the remaining 20 cases are unsettled and no verdict can rest on them.
E2. The prediction registered for E2 had two parts: that E2 would be supported among the read cases, and that across the ledger it would be unclear, because many of the disaster and industry cases give no count. The first part held: among the twenty-three read cases (12.6) the count cell is qualified or provisional in six and settled in two, six of eight, a share of 75%, with five unsettled and ten with no count. The second part failed, since across the ledger the lenient reading supports the hypothesis at 62%, and this section reports the prediction as failed on the registered reading. On the strict reading eight of thirteen support the hypothesis and five go against it, which is the base of the 62% share. Fifty-one cases are not applicable on the strict reading, a statement about the ledger rather than about counts: most of these histories carry no settled count at all. The lenient reading moves fourteen of the strict unclear cases to not applicable, from 25 unclear and 51 not applicable to 11 unclear and 65, while the share over the decided cases stays at 62%; on the strict reading the unclear cases, 25, outnumber the decided ones, 8 supports and 5 against, so that verdict is unclear.
E3. The prediction registered beside E3 had two parts: that E3 would be supported among the read cases, which were chosen for dated gaps, and that it would be weaker across the disaster and industry cases. It is counted as failed, because one part failed: the first part held, and so did the industry half of the second, but the disaster half did not. Strict: supports 21, against 11, unclear 40, not applicable 17; share 66%; verdict unclear. Lenient: supports 21, against 13, unclear 16, not applicable 39; share 62%; verdict supports. Among the twenty-three read cases the gap is named in thirteen, not named in eight and unsettled in two (12.6), 13 of 21 decided, a share of 62%, so the first part held. The group columns carry the second part, and 12.5 gives them in full; what is needed here is their direction. On the strict reading the medicine history stands at supports 10, against 5, unclear 2, share 67%; the disaster history at supports 7, against 1, unclear 27, share 88%; the industrial history at supports 2, against 5, unclear 10, share 29%. On the lenient reading the medicine history gives supports 10, against 6, unclear 1, share 63%; the disaster history supports 7, against 1, unclear 11, share 88%; the industrial history supports 2, against 6, unclear 3, share 25%, which is a verdict of against. The industrial history is weaker than the medicine history, as registered. The disaster history is not: its share, 88% on both readings, stands above the medicine history's 67% and 63%, so the registered weakening outside medicine failed there. The hypothesis named for cases that show a public act, so the disaster history's share runs over its decided cases rather than over its 47 records. Across the ledger, a gap between the event and its record was found stated in 39 cases, not named in 34 more with 3 further cases read as not named by implication, and unsettled in 13.
E4. Strict: supports 9, against 14, unclear 49, not applicable 17; share 39%; verdict unclear. Lenient: supports 10, against 18, unclear 22, not applicable 39; share 36%; verdict against. The prediction failed on the registered reading, and this section reports it as a failure rather than smoothing it. The share on the lenient reading falls below the 40% threshold, so more cases name what would change the reading than do not. Whose text that is has already been set out in 11.2: the text in this field is the archive's later reading, written after the outcome was known, not the actors' account of the hour the act was taken. The conditions named in it are a later reader's conditions, set after the fact. This study reads the lenient verdict as a fact about what an archive can supply, not as evidence about the actors' foresight.
E5. The prediction registered before coding was that E5 would be unclear. Strict: supports 17, against 6, unclear 49, not applicable 17; share 74%; verdict unclear. Lenient: supports 21, against 9, unclear 20, not applicable 39; share 70%; verdict supports. The prediction failed on the registered reading: the lenient reading supports the hypothesis, so in nine of the cases where a verdict could be given the record names who bore the wait, and in twenty-one it does not. The strict reading stays unclear because the unclear cases, 49, outnumber the decided ones, 17 supports and 6 against. The bearer field across the ledger was agreed by both codings in 70 of the 89 cases, with 6 cases where the bearer is named in the record's own words and 3 named by implication; 18 cases do not name it, with 3 further cases read as not naming it by implication, 40 are not applicable and 19 are unsettled.
The sixth prediction, beside the test. The prediction registered beside the test in 12.4 was that fewer than one in five of the cases with an act would pass it. One of the thirty-six cases with an act passes under the two codings, and the prediction held.
The seventh prediction, beside E1's spread. The prediction registered was that the medicine-history cases would show a higher E1 share than the disaster-history and industrial-history cases do. It does not hold. On the strict reading the shares among the decided cases rank the disaster history highest, at 100% with 1 support against 0 and nothing else decided, and the medicine history second at 83%, with 5 for and 1 against; the industrial history stands at 50%, with 1 for and 1 against and 15 unclear. On the lenient reading the medicine history stands at 85%, with 11 for, 2 against and 4 unclear, against 100% for the disaster history, with 8 for, 0 against and 11 unclear, and 83% for the industrial history, with 5 for, 1 against and 5 unclear. The ordering the prediction expected is not there in either reading, and this section reports the prediction as failed. The disaster history's share rests on very small numbers of decided cases, which is why the group columns are set out in full in 12.5 rather than compared here.
Why every strict verdict is unclear. In three of the five strict shares the cases where a verdict could be given already favour the hypothesis — E1 at 73%, E3 at 66%, E5 at 74% — yet the verdict rule also sends a hypothesis to unclear when the unclear cases outnumber the decided ones, and that happens in all five. E1 has 61 unclear against 8 supports and 3 against; E2 has 25 unclear against 8 and 5; E3 has 40 unclear against 21 and 11; E4 has 49 unclear against 9 and 14; E5 has 49 unclear against 17 and 6. The cause is the same in every row: few cells were found stated by both codings. The strict reading counts only what both codings placed in the record's own words, and most of this study's fields — a gap named in 39 cases, a bearer named in 6, a revision named in 15, with 22 unsettled — were found stated in a minority of cases. The lenient reading, which also takes what the setting implies, reaches a verdict in four of the five.
Of the seven predictions judged here, five failed on the registered reading: E2, E3, E4, E5, and the seventh, on E1's spread across the three groups. Two held: E1 and the test. Two of the five that failed, E2 and E3, held in the part that concerned the read cases and are counted as failed because the other part did not: E2 across the ledger, E3 in the disaster history.
12.3The series' hypotheses on this ledger
The series' hypotheses are named here by what they say, since that is what they assert.
On the hypothesis that asks whether the decider is also the bearer, strict: supports 1, against 0, unclear 88, not applicable 0; share 100%; verdict unclear, because the unclear cases outnumber the decided ones and the share rests on a single case. Lenient: supports 14, against 0, unclear 75, not applicable 0; share 100%; verdict unclear, since unclear cases outnumber decided ones. The cells this hypothesis needs are among the weakest in the ledger. The decider was agreed by both codings in 53 of the 89 cases (60%), with 36 unsettled, 30 both readings unknown, 9 institution or state implied, 8 team or organisation implied, 5 stated in both codings, 1 implied as multilateral or civilisation. Of the 53 settled decider readings, 5 are stated in the record's own words — three of institution or state, one of team or organisation and one of multilateral or civilisation — and in the remaining settled cases the codings read a decider out of the setting, or found none. The bearer cell was agreed in 52 of the 89 cases (58%), with 37 unsettled, 38 both readings unknown, 8 cases pointing to those nearest the stoppage by implication, 5 to third parties or the vulnerable and 1 stated. Any comparison resting on these two cells is one on which the coders divided, and this chapter marks it as such.
On the hypothesis that almost no record sets a condition for revisiting a decision, strict: supports 89, against 0, unclear 0, not applicable 0; share 100%. Lenient: the same four counts. Both readings support it, and this chapter says plainly what that verdict rests on. Skeptron's history records have no field for a condition of revisiting: the ledger's table shows 0 of 89 cases with such a condition, 0 conditions in all, and 0 observable ones per case in every horizon band. The verdict follows from the shape of the record form as much as from the cases. This study reads the verdict as a fact about this archive.
The hypothesis that where judgments diverge, material is missing. Both readings: supports 0, against 0, unclear 0, not applicable 89. It could not be tested here. The codings did not pair a divergence cell with what the records list as missing. The ledger does show that all 89 cases list what is missing, with 264 items listed in all, but no cell sets that field against any divergence, and the comparison has nothing to stand on.
The hypothesis that the harder a decision is to undo, the further apart are those who decide and those who bear the cost. Strict: supports 0, against 0, unclear 0, not applicable 89. Lenient: supports 0, against 0, unclear 1, not applicable 88. Not applicable on the strict reading, unclear on the lenient. The reason lies in the cells. Reversal of the constraint was agreed in 38 of the 89 cases (43%) and reversal of the action in 36 (40%), with 51 unsettled in the constraint field and 53 in the action field; both codings read the constraint as easy by implication in 14 cases and as unknown in 24, and the action as easy by implication in 11 cases, costly by implication in 1 and as unknown in 24. Combined with 0 cases naming a decider in their own words, the two halves of the hypothesis have nothing to compare. Any comparison resting on these cells is one on which the coders divided, and horizon cells belong in the same mark: the horizon was agreed in only 32 of the 89 cases (36%), with 57 unsettled, while 88 cases carry a deadline in the record.
The hypothesis that a restriction outlives the memory of its reason. Strict: supports 0, against 80, unclear 9, not applicable 0; share 0%. Lenient: supports 4, against 82, unclear 2, not applicable 1; share 5%. Both readings go against it, and here the reason is visible in the ledger. The reason-recording cell was agreed in 87 of the 89 cases (98%), with 80 finding the reason stated in the record's own words, 2 implied, 4 implying that no reason is recorded, 1 not applicable and 2 unsettled. Skeptron's cases are histories that carry the actors' stated grounds; the hypothesis was drawn from restrictions whose reason the later record has lost. In this ledger the reason is on the page. Whether the recorded reason is the reason that moved the actors at the hour is a different question, and these cells do not settle it.
The hypothesis that the longer the horizon, the less is decided now and the more conditions are set for revisiting. The ledger returns this one to Volume 0, where it is read across the books; it is not judged on this ledger alone. What the ledger supplies is the horizon itself and the conditions beside it. On the strict reading one case had a horizon both codings settled, in the band days to months, and it carries 0.00 conditions per case. On the lenient reading the bands stand at 4 cases at days to months, 6 at years, 4 at decades and 4 at centuries or more, each carrying 0.00 conditions per case and 0.00 observable ones per case. The comparison the hypothesis needs — a horizon against a condition set for revisiting — has nothing set against anything in these records.
The hypothesis that the more a decision is left to an automatic rule, the more the human's place is empty or a name only. Strict: supports 0, against 0, unclear 0, not applicable 89. Lenient: supports 2, against 0, unclear 0, not applicable 87; share 100%, over the two cases where the setting implied it. The human role cell was agreed in 62 of the 89 cases (70%), with 58 reading no by implication and 4 no stated; the place cell in 37 (42%), with 33 implying a human decides, 2 implying a rule decides, 1 unknown and 1 human decides stated. The lenient verdict rests on two cases. It is reported because the ledger gives it, not because two cases carry a claim about institutions.
The hypothesis that the options which cut the immediate loss most often give up evolvability and resilience among the four higher goals. Both readings: supports 0, against 0, unclear 0, not applicable 89. The ledger's options field stands at 0 options, and both codings agree that no case's options reduce the immediate loss; what such options would give up was not settled in any case. The hypothesis cannot be tested on this ledger.
Of the eight, one is supported in every case for a reason that lies in the record form, one goes against for a reason visible in the codings, one is returned to Volume 0, and the rest stand unclear or not applicable on cells where the coders divided.
12.4The test
The test asks a narrower question than the hypotheses. The hypotheses ask what a record tends to hold across eighty-nine cases; the test asks of each case one thing a later reader needs before it can judge an act at the hour it was taken. Thirty-six cases had their act cell settled by both codings, thirty-three of them stated in the record's own words and three implied by the setting.
The scene below is composed for this study from the record's own fields and the reading notes on the study the record is built on. It adds no fact, date or outcome beyond them; the record's words remain the evidence, and Chapter 7 sets out those words in full.
Philadelphia in the autumn of 1918. The ledger fixes this case at 1918-09-28, the date of the parade named in the case's title. The roles in the scene are the ones a public record of this kind carries: those who read the reports, those who wrote the orders, and those who waited between one date and the next, whom the record names as bearing the wait. Hands and eyes belong to the officials holding the reports and the order, and to the households whose days fall in the interval the record measures. What is waited for is not established in the scene's own hour: nothing dated in the record tells what the closing order did, and the question of effect is left open to a later reader.
- Unknown. The record does not date the effect of the order. This study's own coding reads the act-before-proof cell for this case as yes by implication, and the series' cell for who waits and pays reads unknown.
- Notice. The later reader's notice comes through a study published in 2007 on which the record is built. The reading note on that study places Philadelphia's first civilian cases in mid-September and its interventions only in early October.
What the passing case has is completeness on all four fields at once. The ledger's per-case table shows its row as: act, a restriction of movement, work or assembly; act before proof, yes; gap named; bearer named; revision named; count, no-count. Among the twenty-three read cases the same fields run thinner — a named gap in thirteen cases, a named bearer in five, a named revision in fourteen, with the rest not named or unsettled. Most cases carry one or two and not all three. Minamata's First Notice names the gap and the revision but names no bearer; Life Esidimeni names the bearer and the revision but no gap; Smallpox Declared Eradicated names the gap and the revision and names no one who bore the wait; Lind's Twelve Sailors names the gap and the bearer, and its revision field reads not-named. The four requirements are not alike in difficulty. Bearer named is the rarest across the ledger — stated in six cases and implied in three among the eighty-nine — and revision named, where a verdict could be given, is not a minority: eighteen cases name a revision, fifteen stated and three implied, and ten do not.
The passing case is also the one whose later study compared cities by timing. A comparison of that shape needs, in each city, a dated local sequence: when the signs appeared, when the order came, how long the days between ran. The reading note on Hatchett and colleagues' study of 2007 says it set seventeen United States cities beside one another by the timing of their nonpharmaceutical interventions in the autumn of 1918, with Philadelphia — first civilian cases reported in mid-September, interventions in early October — set beside St Louis, which acted within days; the note describes the study as observational, comparing cities by timing without isolating any one measure. Markel and colleagues' study of the same period reads the same shape across forty-three cities from historical data. Skeptron's case and those studies are not asking one question: the studies ask what early and sustained measures did to death rates, while the case asks what the archive holds at the hour the order was made. They can share a record only because that record dates its own act. This study notes the difference and adds nothing about national responses, which the records do not hold.
What the test does not show is as plain as what it shows. It measures what a record lets a later reader date; it does not judge the act. A case that fails the test may still be well evidenced about something else — the Edwin Smith Papyrus's verdicts, the Dongui Bogam's compilation — and a case that passes may still leave the cause unsettled. Philadelphia's record dates the order and names who bore the wait without establishing what the order did, and the codings settled its act-before-proof cell as yes by implication. This study's field and the series' cell ask different questions of the same record, and a claim built on either is a claim about that cell alone.
What would change this reading is stated in the record's own terms: the fifth question's answer fixes the interval in days, which is what lets the two actions be told apart. That interval measures the dates of the acts, not their effects, so any later finding of what the order did would come from outside this record. The test reads the record's capacity, not the act's worth.
12.5Medicine, disaster and industry
The ledger keeps the three series apart, and the group columns show what each series' records are shaped to carry. The seven fields of this study's question, read group by group, stand as follows.
| Cell | Medicine history (17 cases) | Disaster history (47 cases) | Industrial history (22 cases) |
|---|---|---|---|
| act | yes 17 | no 28; unsettled 11; yes 8 | no 11; yes 8; unsettled 3 |
| kind of act | restriction 4; count 4; institution 3; trial 3; notice 3 | not applicable 28; unsettled 11; notice 4; count 3; institution 1 | not applicable 11; unsettled 4; other 3; institution 2; restriction 1; notice 1 |
| act before proof | yes 11; unsettled 4; not said 2 | not applicable 28; unsettled 11; yes 8 | not applicable 11; yes 5; unsettled 5; not said 1 |
| gap named | named 10; not named 6; unsettled 1 | not named 19; named 19; unsettled 9 | not named 12; named 8; unsettled 2 |
| count | no count 5; qualified 5; unsettled 4; settled 2; provisional 1 | no count 40; unsettled 4; settled 2; qualified 1 | no count 17; unsettled 3; settled 1; qualified 1 |
| bearer named | not named 9; unsettled 4; named 4 | not applicable 28; unsettled 12; not named 5; named 2 | not applicable 12; not named 5; unsettled 3; named 2 |
| revision named | named 11; not named 3; unsettled 3 | not applicable 28; unsettled 12; not named 4; named 3 | not applicable 11; unsettled 7; not named 2; named 2 |
The hypothesis verdicts by group, with all four counts where the ledger gives them, stand as follows.
| Hypothesis | Medicine, strict | Medicine, lenient | Disaster, strict | Disaster, lenient | Industry, strict | Industry, lenient |
|---|---|---|---|---|---|---|
| E1, act before proof | unclear (5 for, 1 against, 11 unclear; 83%) | supports (11 for, 2 against, 4 unclear; 85%) | unclear (1 for, 0 against, 34 unclear; 100%) | unclear (8 for, 0 against, 11 unclear; 100%) | unclear (1 for, 1 against, 15 unclear; 50%) | supports (5 for, 1 against, 5 unclear; 83%) |
| E2, counts qualified | supports (6 for, 2 against, 4 unclear; 75%) | supports (6 for, 2 against, 4 unclear; 75%) | unclear (1 for, 2 against, 13 unclear; 33%) | unclear (1 for, 2 against, 4 unclear; 33%) | unclear (1 for, 1 against, 8 unclear; 50%) | unclear (1 for, 1 against, 3 unclear; 50%) |
| E3, gaps named | supports (10 for, 5 against, 2 unclear; 67%) | supports (10 for, 6 against, 1 unclear; 63%) | unclear (7 for, 1 against, 27 unclear; 88%) | unclear (7 for, 1 against, 11 unclear; 88%) | unclear (2 for, 5 against, 10 unclear; 29%) | against (2 for, 6 against, 3 unclear; 25%) |
| E4, text does not name a revision | against (2 for, 8 against, 7 unclear; 20%) | against (3 for, 11 against, 3 unclear; 21%) | unclear (4 for, 3 against, 28 unclear; 57%) | unclear (4 for, 3 against, 12 unclear; 57%) | unclear (2 for, 1 against, 14 unclear; 67%) | unclear (2 for, 2 against, 7 unclear; 50%) |
| E5, text does not name the bearer | supports (6 for, 3 against, 8 unclear; 67%) | supports (9 for, 4 against, 4 unclear; 69%) | unclear (5 for, 1 against, 29 unclear; 83%) | unclear (5 for, 2 against, 12 unclear; 71%) | unclear (4 for, 2 against, 11 unclear; 67%) | supports (5 for, 2 against, 4 unclear; 71%) |
Three things follow from these columns. The first is that the disaster history is mostly not applicable rather than negative: 28 of its 47 cases show no public act in the codings, so its cells describe records that carry no order, declaration, count or trial at all. The second is that the medicine history is the only group in which every case shows a public act, and it is also the group whose counts qualify most often — five of its seventeen cases carry a count qualified by definition, scope or denominator, against one in each of the other groups. The third is that the group shares rest on small numbers of decided cases: the industrial history's strict reading of the act-before-proof field is 1 for against 1 against, and the disaster history's strict reading of the named-gap field is 7 for against 1 against with 27 unclear. These shares are what the codings found in each group's records; they are not measurements of what diseases or periods do.
This study reads the group columns as showing that the verdicts travel with the shape of each series' records. The named-gap hypothesis goes against in the industrial history on the lenient reading, as predicted, and is unclear in the disaster history, where its share, 88%, stands above the medicine history's 63% to 67% and so contradicts the registered prediction of a weaker spread outside medicine. Between the medicine and the industrial history the pattern is one this study reads as the difference between records written to act and records kept to describe.
12.6Where the twenty-three stand among the eighty-nine
The twenty-three cases read closely were chosen for the research question, not sampled from the ledger. Among them the codings find 22 acts and 1 unsettled; the medicine group's 17 cases, every one of which shows an act, sit among these. The other cases among the twenty-three come from the disaster and the industrial histories.
What the twenty-three represent is the dense end of the ledger. The ledger's whole shows the same fields much thinner: across all 89 cases, 36 show an act, 39 gaps are named in the record's own words, and the count field reads no count in far more cases than qualified ones — the codings settled the count cell in thirteen cases only, seven of them qualified by definition, scope or denominator. Read on the study's own fields, the twenty-three carry their own profile.
| Cell | The twenty-three |
|---|---|
| act | yes 22; unsettled 1 |
| kind of act | notice, declaration or certification 6; restriction 5; count or published report 4; institution, compilation or gift 3; trial or treatment 3; other 1; unsettled 1 |
| act before proof | yes 13; unsettled 5; not said 4; no 1 |
| gap named | named 13; not named 8; unsettled 2 |
| count | no count 10; unsettled 5; qualified 5; settled 2; provisional 1 |
| bearer named | not named 13; unsettled 5; named 5 |
| revision named | named 14; not named 5; unsettled 4 |
Denser, then, than the ledger as a whole, and still thin in places. A named bearer appears in five of the twenty-three and not named in thirteen; a named gap in thirteen; and on this study's own test, which 12.4 sets out, one case among the twenty-three satisfies all four conditions at once, against one among the thirty-six ledger-wide cases with an act.
What the twenty-three do not represent is equally clear. They are not a sample of epidemic history, and their verdicts are not computed on them alone: the shares and counts in 12.2 and 12.3 run over the whole ledger, and the twenty-three-case table is a second view of the same coding, not a second sample. The disaster history's 47 cases include 28 that show no public act, none of which appears among the twenty-three; the industrial history's 22 include 11 that show none. Their cells therefore describe records that carry no order, declaration, count or trial at all, and the twenty-three say nothing about them. Of some of the twenty-three this study reads Skeptron's record and the reading notes, and no source text stands behind them. On COVID-19 the boundary is sharper still: Skeptron's records hold one case on COVID-19, the World Health Organization's characterisation of 11 March 2020; lockdowns, vaccines and national responses are not in them, and this study does not use them as evidence.
Two further asymmetries are visible in the ledger itself. The twenty-three carry more named revisions than the ledger as a whole — fourteen named among them, fifteen stated and three implied across all eighty-nine — and that is a property of what the archive's later reading can supply about cases long closed as well as about cases recent. And no case among the eighty-nine lets its own record name a decider: the ledger's table shows 0 of 89 cases with a decider in the record, against 53 cases where the two codings settled the cell by reading a decider out of the setting or finding none. The twenty-three cannot turn the ledger's divided cells into settled ones, so the comparison of who decided and who bore the cost stays one on which the coders divided, as 12.3 says.
The verdicts of this chapter are therefore findings about records rather than about epidemics: what an act's own hour looks like in the archive, how often a record dates the gap it should date, and how rarely it names the people who waited. The earlier accounts — the histories of epidemics, the studies of numbers and of deciding before evidence — are set beside these findings in Chapter 13.
12.7A reader of another family
The verdicts in 12.2 to 12.6 rest on two codings of one model run twice. A third reading exists: one reading, made by a model of another family, of the same eighty-nine records. It was not registered before the coding began, so it does not enter the verdict rule, and it is reported here apart from the verdicts as a check on them: how far the fields carry against a third reader.
The table gives, field by field, the share of cases in which two readings gave the same value, and Cohen's kappa, which discounts the agreement chance alone would give; kappa is 1 for perfect agreement and 0 for chance. Brackets give the cases compared; all three readings cover all eighty-nine.
| Cell | Two codings | First coding and the further reading | Second coding and the further reading |
|---|---|---|---|
| act | 84%, κ 0.69 (89) | 75%, κ 0.50 (89) | 75%, κ 0.50 (89) |
| kind of act | 83%, κ 0.75 (89) | 73%, κ 0.54 (89) | 73%, κ 0.54 (89) |
| act before proof | 78%, κ 0.62 (89) | 73%, κ 0.49 (89) | 71%, κ 0.45 (89) |
| gap named | 85%, κ 0.71 (89) | 76%, κ 0.54 (89) | 82%, κ 0.64 (89) |
| count | 88%, κ 0.69 (89) | 89%, κ 0.70 (89) | 90%, κ 0.72 (89) |
| bearer named | 79%, κ 0.64 (89) | 69%, κ 0.42 (89) | 70%, κ 0.43 (89) |
| revision in case | 75%, κ 0.59 (89) | 63%, κ 0.33 (89) | 64%, κ 0.35 (89) |
Two fields divide the readers most, and both are ones this study leans on. What would change the reading drops from κ 0.59 between the two codings to 0.33 and 0.35 against the further reading; who bore the wait drops from 0.64 to 0.42 and 0.43. Every comparison resting on them is one on which the readers divided, and the mark 12.3 placed on the series' hypotheses stands here as well. The count cell is steadiest, κ 0.69 to 0.72, the one place where the third question can be answered without a divided field. Yet two of the three readings settled a value in 89 of 89 cases for act, gap and count, and 84 of 89 for revision: most cases have a value; the trouble is how strongly each reader holds it.
Where at least two of the three readings gave the same value, the verdicts were read again without registration, and the ledger reports them apart:
| Hypothesis | Strict, two of three | Lenient, two of three |
|---|---|---|
| acts before proof are common | unclear (7 for, 5 against, 67 unclear, 10 not applicable; 58%) | supports (27, 7, 2, 53; 79%) |
| counts are qualified | unclear (10, 9, 21, 49; 53%) | unclear (10, 9, 0, 70; 53%) |
| gaps are named | unclear (21, 11, 47, 10; 66%) | supports (22, 14, 0, 53; 61%) |
| the text does not name a revision | unclear (11, 14, 54, 10; 44%) | against (12, 19, 5, 53; 39%) |
| the text does not name the bearer | unclear (19, 6, 54, 10; 76%) | supports (24, 10, 2, 53; 71%) |
Four keep their direction: E1, E3, E4 and E5 carry the same lenient verdict under the two codings and under the two-of-three reading. One turns. E2, which 12.2 reported as supported on the lenient reading, becomes unclear, on a sample of nineteen decided cases split ten to nine. The share over the decided cases is 53%, inside the band the rule sends to unclear, and on the strict reading the unclear cases, 21, outnumber the decided ones, 19. This study reads the turn as a statement about how little is settled: a cell where ten support and nine oppose leaves no direction for a further reading to confirm, and the turn comes from the ledger's thin decided base, not from a new reading of any record.
The test of 12.4 turns the same way. Under the two codings, one case met all four requirements; where two of the three readings agree, three of the thirty-six cases with an act meet them, and three among the twenty-three read closely. The ledger gives these as counts and no list of cases, so this study reports them without names. Two figures should therefore be carried into the debate: 1 of 36 under the two codings, 3 of 36 under the two-of-three reading. Neither is chosen here; each stands as what its reading found, and every claim in 12.4 about how rare a complete record is was made on the smaller number.
12.8Readers of four families, and a reference reading
The verdicts in 12.2 to 12.6 rest on two codings of one model run twice. Three further readings of the same eighty-nine records were made by models of other families; a reference coder of a family other than the readers' read twenty-five cells where the readers divided, blind to every reader's answer. None of the four families' readings, and no part of the reference reading, was registered before the coding began, so none enters the verdict rule. They are reported apart, and nothing here revises a count of 12.2 to 12.6. This section adds no verdict. It touches two of the study's axes almost entirely: what would make a decision change, and who waits and pays meanwhile. The readers divide most on those two fields, where the further readings carry their value and their disagreement limits any claim built on them.
How often each pair gives the same value. Brackets give the cells compared.
| Pair | Same value |
|---|---|
| the chapter model and the second family | 81% (412 of 509) |
| the chapter model and the third family | 88% (446 of 509) |
| the chapter model and the fourth family | 81% (236 of 290) |
| the second family and the third family | 89% (554 of 623) |
| the second family and the fourth family | 93% (325 of 350) |
| the third family and the fourth family | 93% (326 of 350) |
How often each family agrees with the majority of the others.
| Family | Same value as the majority, where they have one |
|---|---|
| the chapter model | 88% (414 of 472) |
| the second family | 92% (498 of 544) |
| the third family | 97% (502 of 517) |
| the fourth family | 94% (324 of 344) |
The chapter model — the registered reader, whose two codings the verdict rule uses — agrees with the majority least often of the four. The three further families agree with one another more often than any of them agrees with the chapter model, and the third and fourth families stand above 93% against each other. Agreement is not accuracy: families can agree in error.
The cells on which every family agrees run to 399 of the 623 compared. On those cells the fields of this study's question hold their value across four readings.
The verdicts where every family agrees. All five are unclear on the strict reading, with all four counts:
| Hypothesis | Every family agrees (strict) | A majority of families (lenient, reported apart) |
|---|---|---|
| E1, acts before proof | unclear (11 for, 2 against, 37 unclear, 39 not applicable; 85%) | supports (19, 5, 6, 59; 79%) |
| E2, counts qualified | unclear (6, 3, 19, 61; 67%) | supports (9, 6, 4, 70; 60%) |
| E3, gaps named | unclear (10, 5, 35, 39; 67%) | unclear (15, 13, 2, 59; 54%) |
| E4, text does not name a revision | unclear (2, 3, 45, 39; 40%) | supports (16, 8, 6, 59; 67%) |
| E5, text does not name the bearer | unclear (8, 4, 38, 39; 67%) | supports (14, 7, 9, 59; 67%) |
Under a majority of the families, E1, E2, E4 and E5 turn to supported and E3 stays unclear. E4's turn is the one to carry: under the two codings it read against on the lenient reading, as 12.2 reports, and under a majority of the four families it reads supported. Its direction depends on which readers are asked.
The test, both ways. Where every family agrees on the cells, 0 of 21 cases pass the test, and, reported apart because it was not registered, 2 of 28 where a majority of the families agrees. Neither is the registered result of 12.4, which is one of 36 under the two codings.
The reference reading. The twenty-five cells were drawn by field; the coder marked 14 of the 25 as borderline. How many of the cells it read each reader gives as the reference does:
| Reader | All cells | Not borderline | revision in case | bearer named | act before proof | gap named | count |
|---|---|---|---|---|---|---|---|
| the chapter model, first coding | 9 of 25 | 3 of 11 | 4 of 8 | 1 of 4 | 1 of 4 | 2 of 5 | 1 of 4 |
| the chapter model, second coding | 8 of 25 | 3 of 11 | 3 of 8 | 0 of 4 | 1 of 4 | 4 of 5 | 0 of 4 |
| the chapter model, where its two codings agree | 4 of 15 | 1 of 6 | 3 of 6 | 0 of 3 | 0 of 2 | 1 of 1 | 0 of 3 |
| the second family | 17 of 25 | 7 of 11 | 6 of 8 | 3 of 4 | 2 of 4 | 3 of 5 | 3 of 4 |
| the third family | 16 of 25 | 7 of 11 | 6 of 8 | 1 of 4 | 2 of 4 | 4 of 5 | 3 of 4 |
| the fourth family | 12 of 17 | 6 of 9 | 2 of 5 | 2 of 3 | 2 of 3 | 3 of 3 | 3 of 3 |
The chapter model agreed least often with the reference coder on these cells — 9 of 25 against 17 for the second family and 16 for the third. On the field E4 rests on, what would change the reading, the same holds: 4 of 8 in its first coding, 3 of 8 in its second, 3 of 6 where its two codings agree, against 6 of 8 for each of the second and third families. On who bore the wait the chapter model matched least often still — 1 of 4, 0 of 4, 0 of 3 — while the second family matched 3 of 4.
Three things follow. First, the registered verdicts stand as registered, from the two codings, and are not changed after the fact; the further readings and the reference reading are reported apart for that reason. Second, E4 is counted as a prediction that failed: 12.2 reports it as failed on the registered reading, where the lenient verdict is against; its direction turns with the reader, and on the field it rests on the reference coder's blind reading favours the other reading, but the registered verdict stands. A prediction whose direction moves when the reader moves is a prediction on divided ground, and this study reports the division beside the count without changing it. Third, the limits. The sample is only the 25 cells where the readers divided, chosen because they divided, so the table measures no reader's accuracy over the ledger; the reference is one coder, not a reading of a family's size; and 14 of 25 cells were marked borderline, yet on the not-borderline column the second and third families still stand at 7 of 11 while the chapter model stands at 3 of 11.
What this section bears on the study's questions is small and precise. On the axes the series states, the further readings say most about the last two: what would change the reading, and who bore the wait. Those are the two fields 12.3 already marked as divided between the two codings, and the two on which 12.4's test rests. Where four families agree, 399 of 623 cells carry; where they divide, this study names the division rather than resolving it. Every comparison built on a divided field is one on which the readers divided.