Chapter 14 gave the ledger's verdicts on the eight hypotheses, one by one, with their counts. This chapter sets those verdicts beside the earlier theories the study introduced: Knight's distinction, the debate over accidents and reliability, the value of waiting, and the sociologies of who carries a loss. In each case the question is the same: where does the evidence in these 53 records bear on the debate, and where can it not bear at all? Nothing here settles a dispute among the works or claims anything about real markets. Every count comes from the ledger's own tables. Where this study's reading of a pattern is its own, it says so.
15.1Knight's distinction, read through time
Frank Knight's Chapter VII of Risk, Uncertainty and Profit (1921) separates measurable uncertainty from unmeasurable uncertainty, and Chapter VIII gives the two their names, in a passage of Risk, Uncertainty and Profit (1921): "To preserve the distinction which has been drawn in the last chapter between the measurable uncertainty and an unmeasurable one we may use the term 'risk' to designate the former and the term 'uncertainty' for the latter."
Insurance and grouping, on Knight's account, deal with risk; profit is tied to bearing true uncertainty, which cannot be insured. Chapter 3 set out how this distinction has been read since. Stephen F. LeRoy and Larry D. Singell Jr., in Knight on Risk and Uncertainty (1987), argue that Knight's distinction does not turn on whether people can form probabilities, and that he took subjective probabilities for granted; what separates risk from uncertainty is whether an insurance market can exist, because moral hazard and adverse selection can keep a risk from being insured. Richard N. Langlois and Metin M. Cosgel, in Frank Knight on Risk, Uncertainty, and the Firm: A New Interpretation (1993), answer them: for Knight, uncertainty arises where instances cannot be sorted into classes of like cases, so no frequency can be had, and such situations call for judgment. Daniel Ellsberg, in Risk, Ambiguity, and the Savage Axioms (1961), gives the distinction an experimental form: people prefer a bet whose odds are known to an otherwise equal bet whose odds are unknown. John Kay and Mervyn King, in Radical Uncertainty: Decision-Making Beyond the Numbers (2020), press the point further, arguing that many decisions face radical uncertainty that probabilities cannot capture, and recommending that a decision-maker ask what is going on here, reason with narratives, and choose strategies robust to what cannot be foreseen.
The ledger bears on this debate through its rows about grounds and time, and it fits the debate less through its numbers about anything else. The coders agreed on the reason cell in 49 of the 53 records: 40 where both read the reason as stated in the wording and 5 where both read it as implied, with 4 unsettled, 2 where the agreed value was that no reason appeared and 2 coded not applicable. Every record carries a condition for revisiting the judgment, because the form asks for one, but only 5 of the 53 (9%) state it with a number, a threshold or a time, which is what the ledger's code counts as observable; an event or a role it does not count. The horizon cell, which asks how long before an answer comes, was settled by both coders in 30 of the 53 cases, of which 11 imply a horizon of days to months, 1 implies years and 18 are unknown; the other 23 cases are unsettled. No record carries a deadline field at all. Hypothesis H2, which holds that few decisions write down in advance what would reverse them and that those that do write it in words no one could observe, is supported on both readings: 48 of 53 strict and 48 of 53 lenient, a share of 91%, where a share counts the supporting cases among those decided either way.
This study reads the ledger's fit as a question of time rather than a question of probability. Langlois and Cosgel locate Knight's uncertainty in the absence of classes of like cases; a class is what makes a frequency possible. Whether the records in this corpus write those classes is not something the ledger counts. What they write instead is an interval that is not measured: how long before a confidence score can be checked against outcomes, how long before live trading tests a backtest, how long before an approval's scope can be verified, how long before a limit refill becomes a habit. The difficulty is written here as an unobserved waiting period rather than as a number anyone could or could not form, and that is a reading of the writing only: the horizon and observability cells say when an answer would come and what a reader could watch for, and they cannot show where Knight's boundary falls.
Where the ledger cannot fit is equally clear. Hypothesis H3, which holds that where judgments diverge, material is missing, is coded not applicable on both readings, in all 53 cases, because each record carries one example judgment and there is no divergence among judgments to explain. The records therefore bear nothing on the question of whether missing material produces disagreement, and H3's silence decides nothing in the Knight debate above. The link between missing material and disagreement is this study's own reading, not a position in that debate, which turns on insurance markets and on classes of like cases. Nor can the records test Ellsberg's pattern of choice, which concerns how people select between known and unknown odds; a record lists three options and one example judgment, and does not record choices by anyone. Kay and King's question, what is going on here, calls for a narrative reading of the situation, which the ledger does not attempt: its coders record what a record states and what the setting implies about each axis. Keynes's A Treatise on Probability (1921) holds that probability is a logical relation between evidence and conclusion and that the weight of evidence differs from probability; the reason row tells whether grounds were written, and the form does not distinguish grounds given as numbers from grounds given in words, so that row cannot settle Keynes's point either. Chapter 3 already carried the older problem of David Hume's An Enquiry Concerning Human Understanding (1748), Section IV, Part II: every inference from experience supposes that the future will resemble the past, and this supposition cannot itself be proved by experience; that problem concerns whether inference from experience can be justified at all, which the ledger's horizon row does not measure: the 23 unsettled horizon cells record coder disagreement, not how much a record writes. Hypothesis H6, which asks whether a longer horizon leads to fewer decisions made now and more conditions for revisiting, is read across the books in Volume 0; the by-horizon counts — 11 cases at days to months and 1 at years, each with 1.00 condition per case — are too thin to judge here.
15.2Accidents and reliability
Chapter 5 set out the organisational literature against which the cases of warnings were read. Charles Perrow's Normal Accidents: Living with High-Risk Technologies (1984) names two independent dimensions, interactive complexity and tight coupling, and argues that in systems high on both, accidents are normal in the sense of being expected from the structure rather than from a single error. Todd R. La Porte and Paula M. Consolini, in Working in Practice but Not in Theory (1991), open the other side: some organisations, such as air traffic control and aircraft carrier flight operations, run hazardous work with very few failures, and existing organisation theory, which expects error and trial-and-error learning, struggles to explain them. Karl E. Weick's Managing the Unexpected (2001) lists the practices of high-reliability organisations: they attend to failure, refuse to simplify, stay close to operations, keep the capacity to recover, and defer to expertise where it is. Scott D. Sagan's The Limits of Safety (1993) tests the two theories against the command and control of United States nuclear weapons and finds the record closer to normal accidents theory, with dangerous near-accidents and organisational learning limited by secrecy and blame. Barry A. Turner's incubation period, from Man-Made Disasters (1978), and Diane Vaughan's normalisation of deviance, from The Challenger Launch Decision (1996), both read in Chapter 5, describe the drift in between: unnoticed or misread events accumulating against accepted beliefs about a hazard, and an organisation coming to accept as normal a departure it once treated as a warning.
From the ledger's rows on deciders and conditions, that debate looks different from the inside of a record. The decider cell was agreed by both coders in 33 of the 53 cases (62%): 23 where a team or organisation is implied, 8 where an individual is implied, 1 where a team is stated, 1 where an individual is stated, and 20 where the cell is unsettled. The human's place was agreed in 33 of the 53 cases as well: 25 implying a human decides, 3 implying a human approves, 2 implying an automated system or rule decides, 3 unknown and 20 unsettled. Whether a human role is specified was agreed in 42 of the 53 cases (79%): 38 implying no, 1 stating no, 1 implying yes, 1 stating yes and 1 not applicable. Conditions for revisiting are present in all 53, because the form requires them; conditions a reader could observe appear in 5 of 53.
This study reads that pattern as evidence about how decisions are written, not about how organisations behave. The accident literature is asking what an organisation does when a warning arrives, who hears it, and who may stop the work. In these records, that question lands on cells where the coders most often guessed: a team or organisation implied as decider in 23 cases and an individual implied in 8, the decider cell unsettled in 20, and no human role specified in 38 cases by implication and 1 by stated wording. A warning that is not heard by anyone is a different object from a warning that is written about without a hearer, and the records show only the second. Sagan's point that organisational learning is limited by secrecy and blame concerns what organisations do with near-misses; the ledger counts no near-misses, and these records contain no final outcomes.
Hypothesis H5, which holds that a restriction outlives the memory of its reason, is cut against on both readings. Strictly, 0 of the 53 support it and 40 go against it, with 13 unclear and a share of 0%; leniently, 2 support it, 45 go against it, 4 are unclear and 2 are not applicable, a share of 4%. The reason is visible in the reason row: the coders agreed on that cell in 49 of the 53 records, and in most of those the agreed value was a reason stated in the wording. Turner's theory describes unnoticed or misread events accumulating against accepted beliefs about a hazard, and Vaughan's describes a departure coming to be accepted as normal; neither is a theory of a restriction's reason being lost. The records here store the reason at the level of writing because their form asks for it at the moment the decision is made. This study reports the verdict in the ledger's terms and carries it no further: the presence of the reason is not a finding about record-keeping in any genre. A record that stores its reason is not thereby an organisation that keeps its reason. The distinction matters, because Turner's and Vaughan's theories turn on warnings and departures inside an organisation, not on the loss of a reason over time, and only the documentary claim is in evidence here.
The reliability debate also faces a structural limit from these cases. Perrow's dimensions concern systems: interactive complexity between components and tight coupling between them. The records in this corpus are written from the side of one account, one agent set, one set of venues, one vault. Some cases describe connections through which effects can propagate — cross-margin positions, collateral that falls with a position, several agents sharing one account, halts across venues — but none of that classifies a system on Perrow's two independent dimensions, so the records can neither support nor challenge Perrow's classification. What they do contain, in a small number of cells, is a named separation of roles — two keys, a supervisor switch, a gate over an account — which this study reads as documented role separations rather than an answer to La Porte and Consolini's puzzle, since the records give no performance history of hazardous work run with very few failures.
15.3Waiting, irreversibility and muddling through
Kenneth J. Arrow and Anthony C. Fisher, in Environmental Preservation, Uncertainty, and Irreversibility (1974), argue that when a development cannot be reversed and its future benefits are uncertain, the information that will arrive later gives value to keeping the option to preserve; ignoring that value biases decisions towards irreversible commitment. Avinash K. Dixit and Robert S. Pindyck, in Investment under Uncertainty (1994), make the same structure explicit for investment: a project that cannot be undone and can be delayed carries an option value of waiting, which the net present value rule ignores, so it can be rational to wait for information even when investing now looks profitable. David Collingridge's The Social Control of Technology (1980) supplies the matching warning from the other direction: while a technology is young its effects cannot be known, and once they are known it is entrenched and hard to change. Collingridge's remedy is corrigibility — keeping decisions easy to monitor and correct, and keeping options open. Charles E. Lindblom's The Science of "Muddling Through" (1959) takes a different view of how decisions are actually made: administrators decide by successive limited comparisons of a few options that differ little from present policy, rather than by comprehensive rational analysis, and Lindblom defends this incremental method as the one that works.
The records write about waiting in three places, and each is thin. The horizon cell was settled in 30 of the 53 cases, with 11 implying days to months, 1 implying years and 18 unknown, while 23 cases are unsettled; the horizon table of the ledger shows 11 cases at days to months and 1 at years, with no case at decades or longer. No record has a deadline field. The two reversal cells are thinner still: the reversal constraint was agreed in 11 of the 53 cases (21%), with 9 implying easy, 1 implying costly, 1 unknown and 42 unsettled; the reversal action was agreed in 13 of the 53 (25%), with 6 implying costly, 5 unknown, 1 implying impossible, 1 implying easy and 40 unsettled. Hypothesis H4, which holds that the harder a decision is to undo, the further apart are those who decide and those who bear the cost, is not applicable on the strict reading, in all 53 cases, and on the lenient reading it is coded against in 3 cases and supported in 1, with 3 unclear and 46 not applicable, and a share of 25%. The point here is about the shape of the two reversal rows across the whole corpus.
This study reads the record form as posing one question at one moment with three options. A small option set alone establishes neither successive comparison nor options that differ little from present policy, and some planned alternatives suspend exposure or revoke access rather than move a step at a time. Whatever fit with Lindblom's method remains is in the form rather than in any finding about the studio. The fit ends there, because Lindblom's method presupposes a decision-maker who compares and revises, and the ledger does not record who compares: the decider cell was settled in 33 of 53 cases and remains unsettled in 20. Nor does it record whether the comparison is later revised in the field rather than on paper, and the reversal rows show that neither coder could often say what revising would cost.
Against that, the records write a revisit condition in documentary form. Every one of the 53 has one. This study reads the written condition as a documentary analogy to Collingridge's corrigibility, which asks for more than a written condition: decisions easy to monitor and to correct, with options kept open. Only 5 of the 53 conditions name something a reader could watch for. This study reads the gap between those two numbers as a gap in the writing: a condition that is written but not observable leaves the later decision to whatever judgment is made at the time. The option value of waiting, as Arrow and Fisher and Dixit and Pindyck set it out, comes from information that will arrive later and from a commitment that cannot be undone, and the observability of a trigger does not determine it. The records cannot show that waiting paid, because they have no outcomes; they can show what is written about the wait, and what is written is mostly an absence: no deadline, an unsettled reversal, and a horizon that is often unknown rather than stated.
Lindblom's method and the option value of waiting describe different mechanisms, and this evidence tests neither; a record can be sparse enough to fail to evidence either without the two becoming descriptions of one gap in the writing. One says that in practice administrators decide by successive limited comparisons of a few options rather than by comprehensive rational analysis; the other says that information arriving later is worth something because a commitment is hard to reverse. The records sit in the middle: they compare a small set of options and they name a condition for revising, but they do not say who waits, for how long, at what cost, or what would tell them the wait was wrong. Chapter 10 read the waiting cases against Jensen and Meckling's agency costs and Rajan's tail-risk pay structure; Chapter 14 reported that the ledger cannot judge H4. The contribution of this section is narrower: in a corpus where the reversal constraint is agreed in 21% of the cells and the reversal action in 25%, neither the option value of waiting nor Lindblom's incrementalism can be tested, and the difference between a condition written and a condition checkable is a difference in the writing rather than in the value of the wait.
15.4Risk, danger and who carries it
Chapter 10 read the waiting cases against Michael C. Jensen and William H. Meckling's agency costs (1976) and Raghuram G. Rajan's account of peer-measured pay and tail risk (2005). Two further positions belong beside the same row. Niklas Luhmann's Risk: A Sociological Theory (1993) separates risk from danger: risk is a possible loss attributed to one's own decision, danger a possible loss attributed to something outside one's control, and the same loss is a risk to the one who decides and a danger to those who are affected but did not decide. Michael Power's two books make different points about how organisations handle this. Organized Uncertainty: Designing a World of Risk Management (2007) describes risk management as the way organisations turn uncertainty into objects they can manage and account for. The Risk Management of Everything: Rethinking the Politics of Uncertainty (2004) describes risk management spreading into all organisational activity and turning towards secondary risks — the reputational and legal risks to the organisation and its managers — so that defensive documentation can take the place of judgment. Ulrich Beck's Risk Society: Towards a New Modernity (1992), registered here only as its reading note gives it, holds that modernisation produces risks of its own making, distributed across people who did not choose them.
The bearer cell is where these positions meet the records, and agreement on it is low. Both coders agreed in 24 of the 53 cases (45%). Of the settled cells, 12 imply that the bearer is the decider, 7 are unknown, 4 imply that the bearer is whoever is nearest the point where things stop, and 1 implies third parties or the vulnerable; 29 cells are unsettled. Hypothesis H1, which holds that the cost of waiting falls not on whoever decides but on those nearest the point where things stop, is unclear on the strict reading, in all 53 cases, with no cell that could support or go against it; on the lenient reading it remains unclear, with 5 supporting, 12 going against and 36 unclear, a share of 29%.
Against Luhmann, this study reads the row as a distributional finding rather than a theoretical one. In the settled cells, the bearer is the decider in 12 cases and someone else in 5. Where the record names anyone other than the person who decides, it does so in a small and specific group: 4 cases where the people nearest the stoppage wait, and 1 case where the reward is taken and the loss falls on depositors. On the strict reading, Luhmann's distinction cannot be tested, because the coders could not agree on the bearer in most of the corpus. On the lenient reading, the only settled cells naming a bearer other than the decider are the 4 that name those nearest the stoppage and the 1 that names third parties or the vulnerable, while 29 are unsettled and 7 are unknown. Either way, the ledger codes who bears a cost, not whether a possible loss is attributed to one's own decision or to something outside one's control, which is Luhmann's distinction. That is consistent with Luhmann's asymmetry being written out of these records rather than decided about: the form asks who decides less than it asks what is decided, and it does not ask at all who is affected. The records bear on Luhmann in this narrow sense only, and they cannot tell how any loss was attributed, whether to one's own decision or to something outside one's control, because no record codes that attribution.
Power's two books fit the corpus from the other side, and the fit is documentary by design. Every record in the corpus is an object with fields: a scenario, a question, options, an example judgment, a condition for revisiting. The ledger's findings about those fields — agreement on the reason cell in 49 of 53, revisit conditions in 53 of 53, no decider field in any, no deadline field in any, no list of what is missing in any — describe exactly the shape that Power's first book says organisations produce: uncertainty turned into objects that can be managed and accounted for. Power's second book warns about a further step, where documentation serves the organisation's legal and reputational exposure rather than its decisions, and where defensive paperwork takes the place of judgment. Chapter 2's frame rule forbids reading the first shape as a finding, because the form guarantees it; a reader can see the warning applied in a weaker form here. The records record reasons and revisit conditions abundantly and record almost nothing about authority, timing or impact. Whether that shape serves the studio's judgment or serves the studio's account of itself cannot be settled by these 53 records, because no record describes what the documentation was for. This study registers the question and leaves it.
Beck's position is the one the corpus touches least. A risk distributed across people who did not choose it requires a record of distribution, and the ledger contains 5 settled cells in which anyone other than the decider appears as bearer. The automation settings in the future cases of the corpus — agents that propose, agents that halt, agents whose scope widens — are where Beck's question would be asked directly. What this chapter adds is the negative result: with the decider cell settled in 33 of 53 cases and the bearer cell in 24, the different agreement rates do not show that authority and exposure are documented apart, since both are read from the same records and some cases settle both cells, and the records do not let a reader see who is exposed.
15.5What the study adds and what it does not
The Introduction stated three contributions, each narrower than the work it builds on. This section sets each beside the earlier literature and says what it does not do. The counts are the ledger's, and they describe these 53 records only.
The first contribution is a reading of Knight's distinction, and of the debate over it, through the time between a decision and its resolution, as decision records write that time. Beside the earlier work this is a small addition. Chapters VII and VIII of Knight's Risk, Uncertainty and Profit (1921) set the measured against the unmeasurable; LeRoy and Singell, in Knight on Risk and Uncertainty (1987), locate the difference in whether an insurance market can exist; Langlois and Cosgel, in Frank Knight on Risk, Uncertainty, and the Firm: A New Interpretation (1993), locate it in whether instances can be sorted into classes of like cases; Ellsberg, in Risk, Ambiguity, and the Savage Axioms (1961), gives it an experimental form in a pattern of choice between known and unknown odds; Kay and King, in Radical Uncertainty: Decision-Making Beyond the Numbers (2020), recommend that a decision-maker ask what is going on here. The 53 records add none of those positions. What they add is a shape: the difficulty is written as an interval. In 48 of the 53 records the condition for revisiting is written in words no one could observe; the horizon cell is unsettled in 23 of 53 and unknown in 18; in 42 of 53 the reversal constraint is unsettled; and in no record is there a field for a deadline. This study reads that shape as the ledger's contribution to the Knight debate: the difficulty is written as an interval rather than as a number, and the horizon and observability cells cannot show where Knight's boundary falls. Stated that narrowly, the contribution is a description of how a genre of decision record writes the problem that Knight named. It does not resolve the dispute between LeRoy and Singell and Langlois and Cosgel, because a record setting is neither an insurance market nor a frequency series; and it cannot test Ellsberg's choice pattern or Kay and King's recommendation, because the records contain no choices by agents and no outcomes.
The second contribution is the decision ledger itself: 53 records coded on six axes, with three further cells for divergence, the human's place and what each option gains or gives up; two coders of different families, each giving a value, a basis of stated or implied and a reason; every disagreement reported; strict and lenient readings of each hypothesis; and a frame rule that keeps shared studio wording out of the findings. Beside the earlier work, the nearest comparison is Philip E. Tetlock's study of expert forecasts in Expert Political Judgment (2005): a long series of predictions scored against what happened, from which Tetlock concluded that many experts were poorly calibrated and that those who drew on many ideas and revised readily did better than those committed to one big idea. The ledger differs in one decisive way. Tetlock's study has outcomes; these records have none. The ledger therefore scores how a decision is written before the answer, not how good the decision was. Beside Michael Power's documentation critique the ledger occupies an awkward middle: it is a second document about documents, and it exists precisely because documentation can look like evidence. Its defence is that every cell carries a reason, that disagreement is reported — 62% agreement on the decider cell, 45% on the bearer, 21% on the reversal constraint, 25% on the reversal action, 92% on whether a reason was recorded — and that the counts a reader can check come from the cells themselves. Stated narrowly, the contribution is a method for reading a set of decision records on the same axes, with its own unreliability measured and published. It does not produce a population estimate, because these 53 records are the whole of the evidence and no sampling frame is recorded, and it does not resolve the cells where the coders disagreed most, which is why H1 remains unclear and H4 remains not applicable on the strict reading.
The third contribution is what those counts show about the records. Reasons and revisit conditions appear in most records; deciders, deadlines, bearers of waiting and observable triggers are rarely written. Beside the earlier work, this sits against four positions at once. Against Lindblom, the records resemble successive limited comparison in shape — one question, three options — but they do not record who compares, so the method cannot be observed in them. Against Luhmann, the records write the decider's side and leave the affected side to be inferred; the settled bearer cells name the decider 12 times and someone else 5 times. Against Turner and Vaughan the comparison is weaker than it looks: Turner's Man-Made Disasters (1978) describes unnoticed or misread events accumulating against accepted beliefs about a hazard, and Vaughan's The Challenger Launch Decision (1996) describes a departure coming to be accepted as normal, so neither is a theory of a reason being forgotten. The records store the reason for the arrangement at the moment it is made because the form asks for it, and that is no test of either theory: this study reads Turner's incubation period as not predicting absent documentation when an arrangement is made, since recorded reasons can sit beside later unnoticed or misread warning events, and Vaughan's normalisation of deviance is not an incubation theory at all. H5 is cut against on the strict reading, with 0 of 53 supporting it and 40 going against it, and cut against again on the lenient reading, with a share of 4%. Against Beck, the distribution of risk across people who did not choose it is registered in the 5 settled cells where someone other than the decider bears the cost, only 1 of which names third parties or the vulnerable, and cannot be followed further. One further count belongs here. Hypothesis H8, which holds that the options that cut the immediate loss most often give up evolvability and resilience among the four higher goals, is supported on both readings: 24 cases support it, 13 go against it, 16 are not applicable, a share of 65%. What those options give up, as both coders read the 44 options they agreed reduce immediate loss, is resilience in 18, long-term stability in 17, ethical coherence in 13 and evolvability in 7. This study reads that as the ledger's contribution to the theories above rather than to any of them individually: these records are written at the moment when the immediate and the later have to be weighed, and they record the price of the immediate reduction in a vocabulary that no single theory provides.
What the study does not add is the other half of the sentence. It adds no new theory of uncertainty, and it resolves none of the disputes among Knight's readers. It adds no evidence about real markets, real firms or real trades; every record is a synthetic setting published for study, with no outcome and no real people. Each record's example judgment is WSTiger's own reading, dated and revisable, and this study has treated it as material rather than as a conclusion. It adds no demonstration that any organisation succeeds or fails at reliability: some records describe warnings and limit breaches, but the corpus carries no real-world evidence and no final outcomes, and the ledger has no cell for near-misses or incidents. It adds no test of Lindblom's incrementalism, of Beck's distributional claim, or of Power's warning about documentation replacing judgment, for the reasons given in each section above: the evidence is silent, and the form of the records is not a finding about organisations.
The conclusion takes the research question and the four sub-questions one at a time, answers them from the chapters that carry the evidence, and states what would make the study's own answers change.