0.1The question
"The Time of Questions and Answers" is a series of forty-nine books, each in its own domain, each asking the same thing in its own material: what must be decided now, and what would make the decision change? The series has one theme beneath that question, and the theme is time. A question is asked; an answer arrives later; between the two there is an interval in which someone has already chosen, someone is already waiting, and nobody yet knows how the matter will end. Each book of the series reads that interval through the evidence it has, and each book asks who sits inside it, on what grounds, under what deadline, at whose cost, with what chance of taking the decision back.
This study's domain is risk. Risk here means something plain: a chance of loss that is taken on purpose or left in place while the outcome is unknown. A desk writes insurance on a shipping route that has been calm for years. A trading system sizes a position from an estimate it cannot yet test. A firm raises a limit that was breached last month. A vault pays a reward today and leaves a loss with someone else. In every one of these the decision is made before the answer, and the decision changes what happens while the answer is not in. The series question for this domain reads: in taking, holding and stopping risk, what must be decided now, and what would make the decision change?
The evidence for this study is a set of decision records kept by WSTiger, a studio in the RISCON family. WSTiger connects economic history and market signals to judge what risk to take now, and it publishes its judgments as records that other people can read. This study reads fifty-three of those records. They are set across a long span: the earliest is a marine insurance book of 1739, and the latest are automated trading desks of the 2030s. Some are set in the past, some in the present the studio writes from, some in a future it sketches. None of them reports a real firm, a real trade or a real market event; each is a synthetic setting published for study, and each carries an example judgment that is WSTiger's own reading of that setting, dated and open to revision.
The research question of this study is the following. In WSTiger's records of risk decisions, how are three things written down: the time before an answer arrives, the cost of waiting in that time, and the conditions under which the decision would change? And what does the way those records write them show when it is set beside the main accounts of deciding under risk and uncertainty that this study takes from the literature?
That question breaks into four smaller ones, and the study takes them in order.
The first asks who decides. In the records, who is named as the decider, what grounds does the decision rest on, and what is left out of the writing? This is the question of authority and evidence together: who holds the right to act, what the acting is based on, and what a reader cannot see. Chapters 7 to 9 carry it, from the way a decision is judged once its outcome is known, through the question of who sits at the centre of an automated desk, to the way permission to act travels by key, by copied document and by instruction.
The second asks about the waiting. What do the records say about deadlines, about the horizon a setting runs over, about what waiting costs, and about who bears that cost while the answer is not in? A decision record names a question and options, and the interval that follows is paid for by someone: the trader whose position is frozen, the depositor whose funds are locked, the desk whose orders cannot go out, the counterparty whose cash is promised but not delivered. Chapter 10 takes this question.
The third asks about reversal and revision. Can the decision be undone, what does undoing cost, and what would make the decision change? These are two different questions and the study keeps them apart, because a rule can be rewritten while an exposure it created may no longer be closable at the price anyone expected. Chapters 11 and 13 carry them: the first on what can be taken back, the second on the conditions the records themselves write for revisiting a judgment, and on whether anyone could actually watch for those conditions.
The fourth asks about time in the setting itself. Some records are set in the past, some in the present, some in the future, and several threads write the same problem twice or three times over: an equipment warning in 1984 and a leverage limit in 2026 and an agent's scope in 2034; a telegram in the 1870s and a voice message today; a marine route before steam and a weekly option premium now. Does the same setting read differently when it is placed in the past, the present or the future? Who decides, how fast a loss arrives, what can be undone, and what stays the same across the times? Chapter 12 takes that question.
Under those four the study has one subject of its own, and it is worth stating before anything else is described. The subject is the writing itself. A decision record is a document produced before anyone knows how the decision ends, and that is rare evidence. Histories are written afterwards; minutes are written to be filed; reports are written to explain a result. A record of this kind holds the moment when the question was open, when three options stood beside each other and none had won, and when someone wrote down the reason and the condition under which the reason would be reconsidered. Reading many such records side by side shows what this kind of writing tends to carry and what it tends to leave out. The main accounts of deciding under risk and uncertainty were built on other material: on experiments with bets and urns, on the history of crises, on the study of accidents inside complex systems, on the records of expert forecasts. Setting the two together is what this study does, and the setting is meant to be small and checkable rather than grand.
Two things follow from that subject and run through the study. The first is that the argument is about the records and the ideas, never about which option is right. Every record offers three alternatives and an example judgment; this study shows what each alternative gains, what it risks, who stands close to the cost and when, whether it can be taken back, and what the record says would make the judgment be reassessed. The second is that the study's own position is modest. It adds no theory of uncertainty and no evidence about real markets. It offers a way of reading decision records on fixed axes, counts that any reader can check against the same records, and a comparison with a tradition of work that has been asking versions of the same question for three centuries.
0.2Three contributions
The study claims three things, each narrower than the work it stands on.
The first is a reading of Frank H. Knight's distinction between risk and uncertainty through the time between a decision and its resolution. Knight's book of 1921 gave the field its pair of words, and the field has argued about what he meant ever since. This study's reading is that the distinction can be treated as a question about time: how long must pass before cases of a kind can be counted, and what happens to a decision in the interval before that counting is possible? Knight's terms and the debate over them, where Stephen F. LeRoy and Larry D. Singell Jr. read the distinction one way and Richard N. Langlois and Metin M. Cosgel answer them, are set out in Chapter 3. Chapter 15 returns to Knight with the ledger's rows beside the debate. The contribution is not a settlement of that argument. It is a way of reading decision records against it: a record is written at a moment, and the question in each record is what can be measured in the moment the record occupies.
The second contribution is the ledger itself. The study turns fifty-three records into a table with the same cells for every case, codes each cell, and reports how far the coders agreed. Each record is read on six axes: who decides; what the decision rests on and what is missing; the deadline, the horizon, the cost of waiting and who bears it; whether the decision can be undone; the conditions for changing course and whether they can be observed; and the time of the setting and of the judgment. Three further readings sit beside those six: how far the judgments in a record diverge from one another, where a human stands in an automated decision, and what each option gains or gives up among four higher goals. Each cell was coded by two models of different families, each giving a value, a basis for it — stated in the record's words, implied by the setting, or none — and a reason. Every hypothesis was written down before any chapter was drafted, and every verdict is computed twice, once on what both coders found stated and once with what the setting implies. The studio's own writing habits are kept out of the evidence by a rule about shared wording. Chapter 2 sets out the whole of this, and the point of it is modest: counts a reader can check, and disagreements a reader can see.
The third contribution is what those counts show about these records. Two features are written into almost every one of them, and four are rarely written at all. On the first side: the coders settled the cell "reason recorded" in 49 of the 53 records (92%), finding a reason stated in the record's own words in 40 of the 53 and implied by the setting in five more; and every one of the 53 records carries a condition for revisiting its judgment, 53 conditions in all. On the second side: no record names a decider (0 of 53), no record carries a deadline (0 of 53), and only 5 of the 53 conditions (9%) state a number, a threshold or a time that code could see, the rest naming events or roles or states of the world that someone would have to interpret before saying the condition had arrived. Where the question is who waits and pays while the answer is not in, the two coders settled on one value in 24 of the 53 records (45%), and disagreed or found nothing in the rest. A related count runs across the options themselves: 159 options in all, of which the two coders agreed that 44 reduce the immediate loss, and of what those options give up, as both coders read it, resilience counts 18, long-term stability 17, ethical coherence 13 and evolvability 7.
Those counts are findings about a studio's synthetic records. They say nothing about markets, about firms, or about how anyone actually decides. They say that this form of writing carries its reasons and its revisit conditions well, and carries its deciders, its deadlines, its waiting and its observable triggers poorly. Where that gap sits, and what each of the literatures has to say about it, is the rest of the study.
0.3The records
WSTiger is a studio in the RISCON family. Its work is to connect economic history and market signals in order to judge what risk to take now, and its published output for study is a set of decision records. Meridian, the decision-record platform of the same family, publishes those records; this study is one of the forty-nine of the series "The Time of Questions and Answers", and the records are its only evidence about markets and organisations.
A record is a small, fixed bundle. It holds a setting, placed in a place and a date, in which a decision has to be made. It holds a question, which is the decision put as a choice. It holds three options as a rule, named as alternatives and not ranked. It holds an example judgment: WSTiger's own reading of what to do in that setting, offered as an example, dated and revisable. And it holds an invalidation line: the sentence that says when the judgment should be reconsidered. Alongside those fields the record carries its studio, its domain, a marker for the time of the setting, its kind, a pack identifier, the locales in which it exists, a content hash and a fetch timestamp. The last of those let a reader confirm which published version of a record a book has read.
Every record used in this study is marked synthetic, 53 of 53. That word carries a plain meaning: the setting is constructed for study. There is no firm behind it, no trader, no supervisor, no counterparty, no market event. The numbers in it, where a record has numbers, belong to the setting. What the record reports is a question posed before its answer, the alternatives standing beside one another, a reading of which way to go, and a condition for revisiting that reading.
The fifty-three split by the time of setting into 35 present, 9 past and 9 future. The past settings are written as historical material: a marine insurance book on a route that has been calm, a coded ledger readable only by its founder, a wire transfer ordered by telegram, equipment that operators leave untouched while warnings sound, a vault opened by two keys, a reason for a voyage written before the ship sails. The present settings are written from the studio's own desk: a loss limit that refills at midnight, a weekly premium from selling options, a leverage limit raised each time it was breached, a withdrawal instruction in a chief executive's voice, five agents sharing one risk budget. The future settings sketch arrangements a few years off: an agent whose scope widens each month, a mandate that retrains itself, three venues halting at different seconds, a vault whose reward token and depositor losses sit in different hands.
Synthetic settings serve this study's question well, and it is worth being exact about why. The subject of the study is the interval between a question and its answer. A real case can only be read after the answer has come, and the reading is then coloured by what happened: the reason looks prescient or foolish, the option that was chosen looks obvious, the option that was refused looks reckless. A synthetic record holds the decision at the moment before that distortion. The question stands open. The three options stand equal in the writing. The example judgment is one reading among others and is marked as such. Nothing in the record has been corrected by an outcome.
The same feature is the limit of this evidence, and the limit is large. These records have no outcomes. Nothing happened, so nothing can be checked against what happened. No record reports a real person, so no question of motive, pressure, career or blame can be asked of it. The settings are detailed enough to read a decision, and they are not datasets: there are no price series, no loss tables, no org charts. What can be asked of them is what they write down and what they leave out, and how that compares with what the literatures say a decision of this kind carries. That is what the study asks.
0.4How the 53 cases were chosen
The cases were chosen by topic and not sampled. This matters for every count in the study, so it is worth setting out the selection in detail.
WSTiger's public atlas on 2026-10-05 held 424 records with a scenario and options. That figure comes from the series' own count of the atlas and is used here only to show the size of the pool the study drew from. From that pool the study takes the records that pose a decision about taking, holding or stopping risk: a choice that adds exposure, keeps exposure in place, or takes exposure off before anyone knows how the matter ends.
Three groups make up the fifty-three.
The first is the risk series, risk-1 to risk-8, eight settings written for this part of the studio's work. They cover an arrangement in which one person owns both a trade and the explanation of it; five automated agents sharing one human risk budget; a stop button that has never been rehearsed; an assistant pack that may propose but whose right to commit is unclear; the question of who holds an agent's kill switch; one anomaly that freezes three agents at once; a mandate that retrains itself; and three venues that halt at different seconds. Each is a present or future scene about authority, stopping, or the gap between pressing something and knowing what it produced.
The second is the Korean-language risk series, translated once into English for this study. This series is written in three times: a past, a present and a future version of the same setting, and the study reads the three together where the thread runs through them. Its subjects are wide: profit reported while cash is trapped; an approval pack copied forward from an earlier session; marine insurance accumulated on a calm route; weekly option premium income; a confidence score used as an order size; one risk measure adopted by all banks; a vault whose performance token is received by one party while losses fall on depositors; a vault that opens only with two keys; withdrawal authority given to an agent that needs only to place orders; a loss limit that refills at midnight; full Kelly sizing taken from a backtest win rate; a coded ledger readable only by the founder; equipment left untouched while warnings sound; a leverage limit raised every time it was exceeded; an agent's scope widening with each success; a reason for a choice written before the voyage; a post-mortem written after a loss; an urgent wire instruction arriving by telegram; and a withdrawal instruction in a chief executive's voice.
The third is the September 2026 atlas expansion, twenty-six present settings on position sizing, leverage, margin and authority for automated agents. These are desk questions written in the expansion's form: what a stop distance does to size; whether a high win rate makes a strategy good; whether a lower displayed leverage lowers actual risk; what to do when two timeframes disagree; whether a position may be held after its entry hypothesis has been invalidated; whether to resume after a run of losses; whether entry criteria may be loosened to reach a daily target; how a planned split entry differs from averaging a loss; whether the same direction spread across exchanges is diversified; how to judge a profitable trade that broke the rules and a losing trade that kept them; whether realised gains tell the whole story of an account; whether normal entry rules hold before an announcement; whether resting orders count towards risk; whether to enter at a worse price on the same judgment; whether to hold past an expected reaction time; whether one position's loss reaches the margin of others; what happens when collateral falls with the position; whether funds in another account count as margin now; whether an analytical agent's judgment justifies wider order authority; whether to move from simulation to full capital; whether a winning day licenses more risk; whether agents sharing one account are safe within their own limits; and whether a rise in balance may all be booked as investment income.
Purposive selection of this kind has a consequence the study repeats wherever it counts: the counts describe these 53 records and nothing wider. "Most records" in this study means most of these 53. It does not mean most of WSTiger's atlas, most decision records anywhere, or most decisions about risk. A record left out of the study might have a decider named in it; the selection was made on subject, not on any feature of how the records are written.
0.5The method in brief
The method is a ledger, and Chapter 2 sets it out in full. This section gives the shape of it in one page, because the chapters that follow quote its numbers and the reader should know what they are numbers of.
Each of the 53 records is read on six axes. A1 asks who decides. A2 asks what the decision rests on and what is missing from the writing. A3 asks about the deadline, the horizon of the setting, what waiting costs and who bears it. A4 asks whether the decision can be undone, and splits that question into two: undoing the restriction that was put in place, and undoing what the restriction lets happen or stops. A5 asks for the conditions under which the decision would change, and whether a reader could observe those conditions. A6 asks about time: the time of the setting and the time of the judgment. Three further readings sit beside these six. A7 records how far judgments within a record diverge from one another. A8 records the human's place in a decision that involves automation: whether a person approves, decides, reviews, or is named without a role. A9 records what each option gains or gives up among four higher goals — resilience, long-term stability, ethical coherence and evolvability.
Cells in the ledger are of three kinds. Some are copied from the record by code. Some are counted by code: whether a condition carries a number, a date, an event or a named role; how much of a condition's wording is the studio's own frame. The rest are read, by two models of different families, each giving a value, a basis and a reason. Where both coders chose the same value the cell is settled; where they did not it is reported as unsettled rather than repaired.
Eight hypotheses were written down before any chapter was drafted, so that the counts would test statements rather than describe them after the fact. They concern who bears the cost of waiting, whether conditions for reversing a decision are observable, whether missing material explains divergent judgments, whether hard-to-undo decisions separate deciders from bearers, whether restrictions outlive the memory of their reasons, whether long horizons push decisions into revisiting conditions, whether automation empties the human's place, and whether the options that cut immediate loss give up resilience and evolvability. Each is judged by a fixed rule: 60% or more of the cases it applies to supports it, 40% or fewer goes against it, and between those marks, or where unclear cases outnumber decided ones, the verdict is unclear; where no case applies, the verdict is not applicable. Each verdict is computed twice, strictly on what both coders found stated and leniently with what the setting implies.
Two rules protect the counts from the writing around them. The first keeps a studio's habits out of the evidence: a field whose wording is shared across a studio's records at 30% or more is treated as the studio's frame and not as evidence across domains. The second keeps the form out of the evidence: what a record form guarantees is not a finding about how decisions are made. Numbers in this study come only from the ledger, and nothing outside it counts cases.
The study recommends no option. For each alternative in a record it shows what the alternative gains, what it risks, who is close to the cost and when, whether it can be undone, and what the record itself says would make its example judgment be reassessed.
0.6What the evidence cannot show
Four limits sit beside every count in this study, and this section names them before the chapters begin.
The first is the one already stated: the records are synthetic and they have no outcomes. Nothing in them happened, so no verdict in this study can be checked against what happened. The whole purpose of the form is to hold a decision before anyone knows how it ends, and that purpose is also the ceiling on what the form can show. It can show how a decision is written. It cannot show whether the example judgment was right, whether the condition for revisiting arrived, or who actually waited and paid.
The second is the shape of the record form itself. Every record in the study ends with a condition for revisiting its judgment: 53 of 53 (100%), 53 conditions in all. That uniformity is not a finding about careful practice. The form asks for such a line, so a writer completing the form supplies one. In the same way, no record names a decider and no record carries a deadline, 0 of 53 in both cases, and that silence is not a finding either, because the form has no field for either. What can be counted is what the form does not force: that only 5 of the 53 conditions (9%) state a number, a threshold or a time; that 2 of the 53 conditions carry wording that is 30% or more the studio's own frame; that the coders found a reason stated in 40 of the 53 records and implied in five more. The rule is stated once here and used wherever the subject appears: what the form guarantees is not evidence.
The third is the coding. The ledger's cells were read by two models of different families, and their agreement is reported cell by cell rather than assumed. On some cells it is high: the coders settled "reason recorded" in 49 of the 53 (92%) and "human role specified" in 42 of the 53 (79%). On others it is low, and the low cells are the ones this study's most interesting questions turn on. The coders settled who decides in 33 of the 53 (62%) and where a human stands in 33 of the 53 (62%), and settled who waits and pays in only 24 of the 53 (45%). On the two rows that record undoing, the figures are lower still: the reversal constraint was settled in 11 of the 53 (21%) and the reversal action in 13 of the 53 (25%). Where the coders could not agree, the honest report is that the record does not let a reader tell, and that is what the ledger records. Two of the hypotheses — the one about who bears the waiting and the one about what can be undone — are therefore hard to test on this evidence, and the study says so rather than inferring its way past it. The ledger's coders are AI models and are never named; the plan for the series describes them as two models of different families, and the study keeps to that description.
The fourth is the study itself. It is AI's expansion of WSTiger's records. The design was made by an editorial session: the question, the selection of cases, the works and their reading notes, the plan. The chapters were drafted by a model and checked by others. Every block quote in the study is a record's own sentence, copied character for character, and nothing else in the study is a quotation from a record. The works are used at the level of their reading notes, and where a reading goes past a note the study marks it as its own reading. That makes the study's account of the literature an interpretation, and the reader is invited to hold it as one.
Behind all four is a plain statement about subject matter. Nothing in this study is a claim about real markets, real firms or real people. The records are synthetic settings published for study; the counts describe those settings; the comparisons with the literature are readings; the whole book is a construction built to ask how a decision is written before its answer arrives.
0.7The plan of the study
The study has a preface, fifteen chapters in four parts, and a conclusion. The parts move from method, to theories read beside the records, to the ledger's axes one at a time, to the verdicts and what they are worth.
The preface says who made the study and how: a book of this series, written by AI from WSTiger's published decision records, with its design made by an editorial session and its chapters drafted by a model and checked by others. It also gives the reason for reading decision records about risk at all, which is that they keep the moment before the answer, and it states what the study will not do.
Part I holds the method and the ground it stands on.
Chapter 1, "The State of the Question: Four Literatures on Deciding Before the Answer", maps the ideas the study reads the records through. It begins with the background that all of them assume or contest, the emergence of probability as a way of deciding, in Hacking, Daston and Bernstein. It then sets out four literatures in turn, position by position and at the level of their reading notes: risk and uncertainty, from Knight and Keynes through Ellsberg to Kay and King; organisations and accidents, from Turner and Perrow through Reason and Vaughan to the high-reliability school and Sagan's test of it; waiting and irreversibility, from Arrow and Fisher and Dixit and Pindyck to Collingridge, with Lindblom's incremental method beside them; and the judging of decisions and their makers, from heuristics and hindsight through calibration and forecast records, with Luhmann's account of risk and danger, the same loss a risk to the one who decides and a danger to those who did not, and Power's of risk management turning uncertainty into objects an organisation can manage and account for. It closes by naming the gap: these literatures study markets, organisations, experiments and forecasts, and none of them reads written decision records case after case on the same axes.
Chapter 2, "Method: The Decision Ledger", sets out how the records were turned into counts. It describes what a record holds, the six axes and the three further readings, the three kinds of cell, the rule by which a hypothesis is called supported, unclear or against, the rule that keeps a studio's writing habits out of the evidence, and the second rule that what the record form guarantees is not a finding. It then codes two records in full — one person owning both a trade and its explanation, and a balance rise that may or may not be investment income — and closes with the limits of the method.
Part II reads the theories beside individual records.
Chapter 3, "The Measured and the Unmeasured: Knight, Keynes and Their Readers", takes Knight's two terms and Keynes on probability and weight, with Hume on inference from experience beneath both. It reads four settings at the edge of the measurable — a confidence score used as order size, two timeframes in collision, entry rules before an announcement, an entry at a worse price — and then the debate over what Knight meant, with Ellsberg's experimental form of the distinction and Kay and King's radical uncertainty. It closes by carrying the distinction forward as a question about time.
Chapter 4, "Panic, Margin and Collateral: Bagehot and After", sets out Bagehot's rule for a panic and the two accounts of how instability enters a market, Kindleberger's sequence of crises and Minsky's instability, with Mackay's manias as a record of how such episodes have been told. It reads four settings of margin and collateral — a displayed leverage multiple, cross-margin between positions, collateral falling with a position, funds that sit in another account — and opens the question of whether shared models stabilise or destabilise, which Chapter 6 takes up.
Chapter 5, "Warnings Made Normal: Incubation, Accidents and Reliability", reads four settings where a signal has arrived more than once or a stop has never been shown to work. It sets out Turner's incubation period, Vaughan's normalisation of deviance and Reason's latent conditions, reads one thread in three times — equipment under warning, a leverage limit raised whenever exceeded, an agent's scope widening with success — and a stop button that has never been rehearsed, then sets Perrow's two dimensions against the high-reliability school and Sagan's test of the two.
Chapter 6, "Models Everyone Uses: Measures, Monoculture and Performativity", treats a shared measure as something adopted for one purpose and used for more, and sets out MacKenzie on models changing what they measure, Daníelsson on risk produced by the models that measure it, and Bookstaber on complexity in markets, with Long-Term Capital Management as Lowenstein's account gives it. It reads four records of shared measures and asks what each measure covers, what the risk does anyway, and what would show the cover failing.
Chapter 7, "Judging Before and After the Answer", reads five settings where the question is what a result proves. It sets out the heuristics of Tversky and Kahneman and the loss-weighting of prospect theory, reads a reason written before a voyage beside a post-mortem written after a loss, and three records on what a loss and a profit each show and when evidence may be counted. It then sets Fischhoff's hindsight and Baron and Hershey's outcome bias against the use of long records of forecasts as the test of a judge, and asks what an outcome can prove at all.
Part III reads the ledger's axes one at a time.
Chapter 8, "Who Decides, and the Human's Place", takes seven settings about authority on an automated desk: five agents sharing one budget, an assistant that may propose, agents whose orders collide, an analytical agent whose widening authority is the question, and three moments at which something has gone wrong and someone must stop the system. It reads them on who decides and where a human stands, reports what the ledger says about the decider and the human's place, and sets the result beside Simon's decision premises and the garbage can model of organised choice.
Chapter 9, "Keys, Copies and Instructions: Authority That Travels", reads six settings from 1874 to 2034 in which permission to act has been given in advance and travels without the person behind it: two vault keys in different centuries, a founder's coded ledger, an approval pack copied forward, a telegram ordering a wire, and a withdrawal instruction in a chief executive's voice. It borrows two of Saltzer and Schroeder's design principles for protecting information, least privilege and separation of privilege, and closes by reading the six as one shape in three materials.
Chapter 10, "The Waiting and Who Pays for It", takes the second sub-question: what the records say about deadlines, about the cost of waiting and about who bears it. It reads six settings — a calm marine route and a weekly option premium as one shape in two times, two records on averages and expectancy, and two in which one party is paid while another's loss stands unresolved — and closes with the ledger's rows on the bearer and its verdicts on the hypothesis about who waits and pays.
Chapter 11, "What Can Be Undone", separates two acts that a decision about risk performs: writing a rule, and letting an exposure exist. It reads sizing and ruin, positions that wait, and the step from simulation into live trading, then sets the Kelly fraction and Collingridge's corrigibility beside Arrow and Fisher's and Dixit and Pindyck's value of waiting, and reports the ledger's rows on reversal and the verdict on the hypothesis that ties them.
Chapter 12, "Clocks That Reset: The Time of a Setting", reads the clocks inside the records: a limit that refills at midnight, a pause after a run of losses, a trading day with a target in it, a reaction period that expires. It then returns to the settings that earlier chapters read in two or three times and asks what the time of the setting changes and what it leaves alone.
Part IV gives the verdicts and their place in the tradition.
Chapter 13, "Conditions for Revisiting: What Could Be Observed", reads the invalidation lines alone. It shows why the existence of a condition in every record is not a finding, reads the conditions that name something a reader could check against those that name a state of the world someone would have to interpret, sets both beside Clarke's fantasy documents and Neustadt and May's thinking in time, and reports the ledger's verdict on the hypothesis about revisiting conditions.
Chapter 14, "The Hypotheses Tested", reports the verdicts on all eight hypotheses, with their counts: what the records support, what they go against, and what this corpus cannot judge. It closes by saying what such verdicts are worth, given synthetic records, AI coders and one studio's output, and what the series' later volume will do with the ledger.
Chapter 15, "The Findings Beside Earlier Theories", sets those verdicts beside the theories of the earlier chapters: Knight's distinction read through time, the argument between normal accidents and high reliability, the value of waiting against Lindblom's incremental method, Luhmann's account of who carries a loss, Power's of how organisations turn uncertainty into risks they can manage and account for, and Beck's risk society at the level its note gives. It ends with what the study adds and what it does not.
The conclusion answers the research question and the four sub-questions one by one from the chapters that carry them, restates the three contributions and the limits beside them, points to further work, and writes the study's own conditions for revisiting in a form a reader could observe: a new coding with higher agreement, records with fields for deadlines and deciders, records with outcomes.