Conclusion

This study began from one question about risk taken, held or stopped before anyone knows how it ends: what must be decided now, and what would make the decision change? Under that sat a narrower question about the records themselves. A decision record holds a moment before the answer comes; what does it write down about that moment, and what does it leave blank? The study read 53 of WSTiger's synthetic decision records, from a marine insurance book of 1739 to automated trading desks of the 2030s, coded every one of them on the same axes in a ledger computed by code, tested eight pre-registered hypotheses against those cells, and set what the counts show beside four literatures on deciding before the answer. What follows states the answers the evidence allows, the contributions that remain after the limits are held beside them, the limits themselves, what could be studied next, and the conditions under which these answers would change. The study ranks no option and claims nothing about real markets. Its subject is what the records write and what they do not.

C.1 The answers

The research question asked how the time before an answer, the cost of waiting, and the conditions for changing course are written down in WSTiger's records of risk decisions, and what that shows beside the main accounts of deciding under risk and uncertainty. The short answer is this. The records write reasons and conditions. Of the 53, the ledger counts 49 in which the coders agreed that a reason is recorded, with 40 of those read as stated in the record's own words and 5 as implied by the setting. Every record carries a condition for revisiting its example judgment: 53 of 53 (100%), one condition each, 53 conditions in all. Alongside those, the parts of a decision that carry its time are thin. No record carries a decider in the record's own fields, no record carries a deadline, no record lists what is missing from the material the judgment rests on. Only 5 of the 53 conditions, 9%, name something the counting code can see as a number, a threshold or a time. Those five figures frame everything this study can say: the ground of each judgment is written, the timing and the agency of each judgment largely are not.

The four sub-questions can now be answered in turn.

First: who decides in the records, on what grounds, and what is left out. Chapters 7 to 9 carry this question. Chapter 7 reads five records on how a decision should be judged before and after its answer: two set the same subject in 1956 and 2026 around a reason written before or after the outcome, and three ask what a loss, a profit and a live drawdown each prove. Chapter 8 reads seven records on authority, four of them written from inside an automated desk where agents share collateral, propose trades or widen their own scope, and three written at the moment something goes wrong, asking who may stop the system. Chapter 9 reads six records in which authority travels without the person behind it: a vault key, an API key, a coded ledger, a copied approval pack, a telegram, a voice. Read together, the three chapters describe settings in which a decision is put as a choice among options and a reason is given for the example judgment, while the person or body who would choose is left for a reader to infer.

The ledger's counts support that reading. The two models of different families agreed on the decider cell in 33 of the 53 cases (62%), and almost all of the settled values were implied rather than stated: team-or-organisation implied in 23 cases, individual implied in 8, with one case each of team-or-organisation and individual stated outright. In 20 cases the cell stayed unsettled. On the extra axis for the human's place, agreement ran at 33 of 53 (62%), with human-decides implied in 25 cases and unsettled in 20; automated or rule-based deciding was read in 2 cases. The cell asking whether a human role is specified shows agreement in 42 of 53 (79%), but the agreed value is mostly the negative: no role specified, implied in 38 cases and stated in 1, a role specified in 2 cases, unsettled in 11. What the records give instead is grounds: reason recorded in 49 of 53 (92%). What they do not give is a list of what is missing from those grounds; the ledger counts 0 items across all 53. This study reads the combination as a practice of writing the argument for a judgment while leaving the seat at the table unnamed.

H7, on the human's place where a decision is left to an automatic rule or an agent, follows the same pattern and the same weakness. On the strict reading, which counts only what both coders found stated, H7 is not applicable in all 53 cases. On the lenient reading, which adds what a setting implies, it is supported in the 5 cases where it applies, a share of 100%, with 48 cases not applicable. A verdict resting on five cases is a thin one, and the ledger reports it as such rather than widening it.

Second: what the records say about deadlines, the cost of waiting, and who bears it. Chapter 10 takes this up directly, reading six records on axis A3: a marine insurance book accumulating premium on a calm route in 1739, a weekly option premium desk in the present, two records on averages and expectancy, a vault in 2034 where a reward is received and losses fall on depositors, and a 1974 setting in which reported profit sits on the books while cash is trapped. Chapter 12 adds the clocks those intervals run on: a limit that refills at midnight, a pause after consecutive losses, the hours left in a trading day, a reaction period that expires, the rest of the day after a booked gain.

The counts here are the emptiest in the ledger. No record has a deadline field. On the horizon cell the coders agreed in 30 of 53 (57%), and the settled values leave most of the corpus undated: unsettled in 23 cases, unknown in 18, days-to-months in 11, years in 1. On who waits and pays while the answer is not in, agreement fell to 24 of 53 (45%), the lowest of the readable cells: unsettled in 29 cases, the decider implied in 12, unknown in 7, those nearest the point where things stop implied in 4, third parties or the vulnerable implied in 1. H1, the hypothesis that the cost of waiting falls not on whoever decides but on those nearest the stoppage, is unclear on both readings. Strict: 53 cases unclear, none supporting or against. Lenient: 5 supports, 12 against, 36 unclear, a share of 29%. The verdicts belong to the ledger and are reported here as it gives them. The evidence does not show that the cost of waiting falls elsewhere, and does not show that it does not; it shows that the records, read by these coders, mostly do not say who stands in the interval.

Third: whether the decisions can be undone, what would make them change, and whether those conditions could be observed. Chapters 11 and 13 hold this question. Chapter 11 separates two acts a decision about risk performs: writing a rule, and letting something happen under that rule. It reads six records through that split, on sizing and ruin, on positions that wait, on the step from simulated trading to real fills, and closes beside the option value of waiting. Chapter 13 returns to eight records read earlier, for their Invalidation lines alone, and sorts those lines into conditions a reader could go and look for and conditions that name a state of the world no one could check without interpreting it first.

On undoing, the ledger is close to silent. The coders agreed on the reversal-constraint cell in 11 of 53 (21%), with the cell unsettled in 42 cases; on the reversal-action cell they agreed in 13 of 53 (25%), unsettled in 40. H4, which asks whether 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, and against on the lenient reading with 1 support, 3 against, 3 unclear and 46 not applicable, a share of 25%. The study does not read that lenient verdict as a refutation; with 46 cases not applicable and agreement at roughly one case in five on the underlying cells, it reads it as what the corpus can reach.

On conditions, the corpus is uniform in one respect and sparse in another. All 53 records carry a condition for revisiting the judgment, and Chapter 2's rule keeps that number from counting as a finding: the record form asks for such a line, so every completed record has one. What counts is what the conditions contain. Five of the 53, 9%, name something the code can see as a number, a threshold or a time; two carry wording that is 30% or more the studio's frame and so cannot serve as evidence across domains. H2, which holds that few decisions write down in advance what would reverse them, and that those who do write it in words no one could observe, is supported on both readings: 48 supports, 5 against, 0 unclear, a share of 91%. This study reads that support as a claim about how conditions are written, not about whether decision-makers in general change course: the corpus shows a form that guarantees a trigger and a practice that mostly leaves the trigger to judgment.

Fourth: whether the same setting reads differently in the past, the present and the future. Chapter 12 carries this question. It reads five records around their clocks and resets, then returns without re-quoting to the settings that earlier chapters read in two or three versions: a warning ignored at a plant dated 1984, a leverage limit raised at a desk dated 2026 and an agent scope widening under a rule dated 2034 in Chapter 5; a reason written before a voyage in 1956 and a post-mortem written after a loss in 2026 in Chapter 7; vault keys, coded ledgers and payment instructions across the nineteenth, twentieth and twenty-first centuries in Chapter 9; a calm insurance route in 1739 and a weekly option desk in the present in Chapter 10. The ledger's time-of-setting counts are 35 present, 9 past and 9 future, all 53 marked synthetic. The same shape of question appears in all three times: who may act, how fast a loss arrives, what can be called back. What changes is the clock the question runs on and the speed of the machinery, from postal checks that take days to transfers that cannot be reversed.

H6, on the horizon and the number of conditions set for revisiting, was pre-registered to be read across the books of the series rather than inside one, and the ledger leaves it there: read across the books in Volume 0. The by-horizon figures the ledger reports for this corpus are small and should be stated as they stand: under the lenient reading, 11 cases with a horizon of days-to-months carry 1.00 condition per case and 0.09 observable per case; 1 case with a horizon of years carries 1.00 condition per case and 0.00 observable per case. In these readings the number of conditions does not vary with the horizon, and observability stays near zero in both. Whether longer horizons lead to more conditions set for revisiting cannot be answered from one book's corpus, and this study leaves the question where its pre-registration put it.

Two further verdicts belong in the answer because they bear on the same material. H8, which holds that the options cutting the immediate loss most often give up evolvability and resilience among the four higher goals, is supported on both readings: 24 supports, 13 against, 16 not applicable, a share of 65%. The supporting material is a property of the options rather than of the outcomes: the ledger counts 159 options across the corpus, of which both coders agree that 44 reduce the immediate loss, and what those options give up, as both coders read it, is resilience 18, long-term-stability 17, ethical-coherence 13, and evolvability 7. H5, which holds that a restriction outlives the memory of its reason, is against on both readings: strict, 0 supports against 40, with 13 unclear and a share of 0%; lenient, 2 supports against 45, with 4 unclear and 2 not applicable, a share of 4%. Chapter 14 gives the reason the counts suggest: these records write the reason for the restriction alongside the restriction itself, so the pattern H5 describes is not what this corpus shows. H3, on divergence among judgments alongside listed missing material, is not applicable on both readings, because no record lists what is missing, leaving nothing to compare divergence against.

Read as a whole, the answers to the four sub-questions describe one shape. The records of this corpus are careful about reasons and about triggers, and thin about people, deadlines, bearers and reversals. The care is visible in counts any reader can check; the thinness is visible in agreement rates and in fields that the form does not carry. The study's answer to its research question is therefore an answer about the writing of risk decisions before their outcomes: in these 53 synthetic settings, the time between a question and its answer is written as a condition and left unwritten as an interval, and the interval is where the cost of waiting, the identity of the bearer and the possibility of undoing would have to live.

C.2 The contributions

Three contributions were promised at the start, each narrower than the work it builds on, and each can now be stated with the evidence that stands behind it.

The first is a reading of Knight's distinction and the debate over it through time. Chapter 3 sets out Knight's two terms as Chapter VIII of Risk, Uncertainty and Profit (1921) names them, Keynes on probability and the weight of evidence, Hume on inference from experience, and the two readings of Knight held by LeRoy and Singell on one side and Langlois and Cosgel on the other, with Ellsberg's ambiguity aversion and Kay and King's radical uncertainty nearby. It then reads four records at the edge of the measurable through that debate and asks of each whether the difficulty is a missing frequency, a missing market for the risk, or a case that cannot be classed with others. Chapter 15 sets the ledger's cells beside the same distinction. What the study adds is small and specific: a decision record writes a moment rather than a state of the world, and read as moments the distinction becomes a question about how long it takes before a kind of case can be counted. The horizon cells of the ledger, settled in 11 cases at days-to-months and 1 at years, with 23 unsettled and 18 unknown, are what that question looks like in this corpus. The contribution is a way of reading the debate, not a settlement of it.

The second contribution is the ledger itself. Chapter 2 sets out what a record holds, the six axes and the three further values, the three kinds of cell — copied, counted and read — the two models of different families who coded each readable cell with a value, a basis and a reason, the strict and lenient readings, the pre-registered hypotheses and the verdict rule, the frame rule that keeps a studio's writing habits out of the evidence, and the companion rule that what the record form guarantees is not a finding. It closes by coding two records in full so the method can be followed cell by cell. What supports the contribution is its reported reliability rather than any claim of precision: agreement runs from 92% on whether a reason is recorded down to 21% on the reversal-constraint cell, and every figure is published beside what it counts. A reader can disagree with a coder's choice without touching the record, and can disagree with the frame rule without losing the counts.

The third contribution is what the counts show about these records. Reasons are written in 49 of 53 cases as both coders read them; revisit conditions exist in 53 of 53 because the form requires them; conditions naming a number, a threshold or a time number 5 of 53; the decider appears in the record's fields in 0 of 53; the deadline appears in 0 of 53; what is missing from the grounds is listed in 0 of 53; the bearer of the waiting is settled in 24 of 53. This is a finding about a studio's synthetic records and about the form in which they are written. It is not a finding about markets, firms or traders, and the study states it at that width every time it appears. What makes it worth having is that it is checkable and that it points at a specific blank: the interval between a question and its answer, in which someone waits and pays, is the least written part of these records.

C.3 The limits

Four limits run through every count and every reading in this study, and they should be restated together here.

The first is the character of the evidence. Every record used is marked synthetic: a setting published for study, with no real firm, trade, plant or market event behind it. No record has an outcome, because the form holds the decision before anyone knows how it ends. That is what makes the corpus useful for this question and what forbids a whole class of claims. Nothing here can be checked against what happened, because nothing in the corpus happened. Chapters 7 and 14 make the point from two directions: the records keep the reason before the answer, which is exactly the material hindsight destroys, and they cannot say whether a judgment was good, because a judgment's quality cannot be tested against an outcome the form does not carry.

The second is the selection. The 53 records were chosen by topic rather than sampled: the risk series and its past, present and future versions of one setting, and the September 2026 atlas expansion on position sizing, leverage, margin and authority for automated agents, taken from WSTiger's public atlas of 2026-10-05, which held 424 records with a scenario and options. Purposive selection means that "most records" in this study means most of these 53, and nothing more. The counts do not describe the atlas, the studio's wider output, or any population of decisions.

The third is the coding. The coders are AI models of two families, never named in this study, and their agreement varies sharply by cell. Where agreement is high, the counts carry weight: reason recorded at 92%, human role specified at 79%. Where it is low, the counts carry little: the bearer of the waiting at 45%, the reversal-constraint cell at 21% and the reversal-action cell at 25%. That is why H1 is unclear on both readings and why H4 cannot be judged strictly at all. The study reports the disagreement rather than repairing it by inference, and treats unsettled cells as unclear in the verdicts, as the rule requires.

The fourth is the form and the authorship. A record that carries a revisit line does so because the form asks for one, so 53 of 53 is a property of the design and not a finding about practice. A record that lacks a decider field and a deadline field lacks them for the same kind of reason, so 0 of 53 counts the form as much as the writing; Chapter 2 keeps that fact separate from the finding it can support, which is about what the fields, once offered, are filled with. The frame rule removes wording shared across a studio's records from cross-domain evidence, and 2 of the 53 conditions reach the 30% threshold that disqualifies them. Finally, this study is AI's expansion of WSTiger's records: the design was made by an editorial session, the chapters drafted by a model and checked by others, and the readings of the works are the study's own, kept inside the reading notes that describe each work. Where a reading goes past a note it is marked as this study's reading. No claim about real markets is made anywhere in the study, and none can be drawn from it.

C.4 Further research

The blanks in this corpus are specific enough that further work can be aimed at them rather than at generalities.

The first line of work is already scheduled. Each book of the series writes its findings into a ledger another book can read, and Volume 0 will read across those ledgers. H6 is reserved for exactly that reading. The by-horizon figures for this corpus — 11 cases at days-to-months with 1.00 conditions per case and 0.09 observable per case, 1 case at years with 1.00 and 0.00 — are too thin to carry the hypothesis alone, and a cross-book reading is where the question of horizon and revisiting can be approached properly. Other books of the series, in other domains, will also show whether the pattern this study found is a property of risk records or of decision records as such: reasons written, triggers written, deciders and deadlines unwritten.

The second line concerns outcomes. Records of decisions taken and later resolved would let the questions this study cannot reach be asked at all. H1 needs to know who paid while an answer was pending; H4 needs to know what was later reversed and at what cost; H5 needs to know whether a restriction survived its reason; the debate of Chapter 7 on whether an outcome is evidence about a decision needs outcomes to test. A corpus of paired records — the judgment before the answer, and the answer afterwards — would let the same axes be coded twice and compared.

The third line concerns the coders. Human coders placed beside the model coders would show how much of the disagreement is the records and how much is the reading. The two cells with the lowest agreement, the bearer of the waiting and the two reversal cells, are the natural place to start, since H1 and H4 turn on them and both are currently unjudgeable. Agreement rates are already published per cell, so any new coding has a clear baseline to be compared against.

The fourth line concerns real material. Public records of actual decisions — board minutes, risk committee papers, supervisory files, incident reviews, post-mortems — carry what the synthetic form does not: a named decider, a stated deadline, an outcome, and often a record of who bore the waiting. The same axes could be applied there, and the frame rule and the form rule would still be needed, because real institutions have writing habits too. Saltzer and Schroeder's principles, Jensen and Meckling's account of who bears risk when one party acts for another, and Luhmann's separation of risk from danger were each read in this study beside records that pose the question in miniature; public records would let those readings be tested against organisations rather than settings.

The fifth line concerns the form itself. Chapter 13 sorted the 53 conditions into those a reader could check and those that name a state of the world requiring interpretation first. Work could ask what a condition needs in order to become observable: a number in a limit, a date, an event in a sequence, a named position whose action settles the matter, or an inspectable procedure such as an independent review or a separate log field. The five observable conditions in this corpus are a starting sample for that question.

C.5 What would change this study's answers

The study owes its readers the same thing every record owes its readers: a written answer to what would make the judgment be revisited, stated in words someone could check. Five such conditions follow, each attached to a count the ledger already publishes.

First, a new coding of the same 53 records in which the coders settle the bearer cell well above the 45% this ledger reports, and the two reversal cells above 21% and 25%. That would change the answers to the second and third sub-questions before any other, because H1 and H4 are unjudgeable for want of agreement rather than for want of cases. A reader can observe this by comparing a published per-cell agreement table against the one in this study.

Second, a corpus whose records carry a decider field and a deadline field. The code that counted 0 of 53 for each would then count something, and the answer to the first sub-question would move from inference to statement. The coders' reading that a human role is unspecified in 38 of 53 cases and the H7 verdict resting on 5 applicable cases would both be open to the same test.

Third, records with outcomes, or studies of real decisions where an outcome exists. This would change the study's answers about judging a decision before and after its answer, about what a loss proves, and about H5, whose verdict of against rests on the corpus writing reasons alongside restrictions. It would also allow the debate of Chapter 7 — whether an outcome is evidence about a decision, and over how many decisions — to be examined on material that happened.

Fourth, Volume 0's cross-book reading of H6. Until that reading exists, this study's answer to the fourth sub-question is limited to what one corpus shows about clocks and resets: 35 present settings, 9 past and 9 future, with the same shapes of question appearing in all three and the machinery speeding up. A cross-book reading could change that answer, or confirm it.

Fifth, more records whose conditions state a number, a threshold or a time, lifting the count of 5 observable conditions above the 9% of the corpus it now stands at. That would change the reading Chapter 13 gives of how triggers are written, and with it the support for H2.

None of these conditions is a matter of interpretation. Each names a count, a field or an artefact a reader could go and look for, in the spirit Chapter 13 asks of the records themselves. Until one of them arrives, the verdicts stand as the ledger gives them: H2 and H8 supported, H5 against, H1 unclear, H3 not applicable, H4 not applicable strictly and against leniently on a share of 25%, H6 reserved for Volume 0, H7 not applicable strictly and supported leniently on 5 cases out of 53. The study's answers stand with them, and they are answers about what is written down before an answer comes, not about which option anyone should take. The records hold the question open on purpose. This study has tried to show, of the interval they hold open, who is named inside it, who is not, and what would show the interval to have been written better.