This study began from one question about war and its edges, and one question about the records that hold such decisions: what must be decided now, and what would make the decision change? The materials were fifty defence questions published by Meridian, the decision-record platform of the RISCON family, each a synthetic setting in a fictional polity, each carrying thirteen AI-generated judgments, each coded on the same axes by two models of different families and read strictly and leniently. The study read those records beside four literatures on deciding in war — how wars end and ceasefires hold, how warnings are believed or missed, who governs in an emergency and who supervises machines, and what is owed after mass violence to people who were not in the room. What follows answers the study's research question and its four sub-questions with the ledger's counts, sets out the contributions, states plainly what the evidence cannot carry, names what could be studied next, and writes the conditions under which this study's own answers would change.
C.1 The answers
The research question asked how the time between a question and its answer is written down in Meridian's defence questions and their judgments: who decides by when, who waits and pays, what can be undone, what would reopen the decision, and where the judgments part. The short answer is that the form writes that time down densely on four of those five points and thinly on one. Every question names a decision actor and a mandate. Every question names a deadline. Every question lists what is missing. Every question carries conditions for revisiting the judgment. What the records hold with much less force is the identity of the people who wait while the decision is pending, and the question of whether anyone outside the deciding room could in fact observe a reopening condition being met.
Four sub-questions carried that answer in parts.
The first asked who decides and on what grounds, and what is missing. The ledger's answer has two halves. On the side of what is written, the form guarantees a decider: cases with the decider in the record number 50, cases that list what is missing number 50 with 148 items listed in all, and the human's role is specified in 50 of 50 cases with a reason recorded in 50 of 50. On the side of what two independent readings could settle, agreement is thinner. The coders agreed on the kind of decider in 28 of 50 cases, 56 per cent, with 22 unsettled between them; the values they settled on were team-or-organisation stated in 11 cases, institution-or-state stated in 10, institution-or-state implied in 4, multilateral-or-civilisation implied in 2 and stated in 1. On the human's place the agreement runs far higher: 46 of 50 cases, 92 per cent, read as human-decides, stated in 33 and implied in 13, with 4 unsettled. What is missing is likewise counted rather than described: 148 items across the fifty questions, a number the ledger gives without characterising the items themselves. Where the thirteen judgments part, the count is 13 of 50 questions with published judgments choosing more than one option, out of 13 judgments per case. On that branch the ledger supports hypothesis H3 — where judgments diverge, material is missing — in all 13 cases the hypothesis applies to, a share of 100 per cent in both the strict and the lenient reading, with 37 cases counted as not applicable. Chapter 9 develops the first half of this answer, on the decider and the human's place; Chapter 14 develops the second, on the questions where the judgments divide.
The second sub-questions asked about the deadline, the cost of waiting and who bears it. On the deadline the record is complete by construction: cases with a deadline in the record number 50. On the bearer it is not. The coders agreed on who waits and pays in 42 of 50 cases, 84 per cent, with 8 unsettled; the settled values were those-nearest-the-stoppage stated in 22 cases and implied in 15, third-parties-or-the-vulnerable stated in 2 and implied in 2, and future-generations stated in 1. Hypothesis H1, which says the cost of waiting falls not on whoever decides but on those nearest the point where things stop, is supported in both readings: strictly, 25 cases support it against 0, with 25 unclear carrying no part of the share, which reads 100 per cent; leniently, 42 support it against 0, with 8 unclear, again a share of 100 per cent. The strict split matters. Half the collection supports H1 in the record's own words and half cannot be settled without inference; the lenient reading shows how much of that half the settings themselves will carry. Chapter 10 works through six questions on this axis and sets the count beside the distinction between a risk one decides to run and a danger one only suffers.
The third sub-question asked what can be undone, what would reopen a decision, and whether a reader could observe the reopening. The ledger answers with two reversal rows and one conditions row. On the reversal constraint the coders agreed in 40 of 50 cases, 80 per cent: easy implied in 27, costly implied in 12, costly stated in 1, with 10 unsettled. On the reversal action they agreed in 37 of 50, 74 per cent: costly implied in 30, costly stated in 1, impossible implied in 4 and stated in 2, with 13 unsettled. Hypothesis H4, that the harder a decision is to undo the further apart stand those who decide and those who bear the cost, is supported in both readings: strictly 3 cases support it and 47 are not applicable, a share of 100 per cent over the decided cases; leniently 30 support it, 7 are unclear and 13 not applicable, again 100 per cent over the decided ones. On reopening, the collection is unusually rich and unusually hard to check. Cases with a condition for revisiting the judgment number 50 of 50, and the conditions in all number 798. Of those, 650 lie in the field the judgments carry for stopping conditions, of which 602 are observable by the ledger's standard, and 148 lie in the question's own reopening field, of which 27 are observable by that standard. Conditions stated with a number, a threshold or a time number 629 of 798, which is 79 per cent. Conditions whose wording is 30 per cent or more the studio's frame number 0 of 798. Against that record, hypothesis H2 goes against in both readings: supports 0, against 50, a share of 0 per cent. The hypothesis said that few decisions write down what would reverse them and that those which do write it in words no one could observe; these fifty questions write it down in every case, and mostly in words carrying a number, a threshold or a time. Chapter 11 sets out what the reversal rows show and what corrigibility would ask of them; Chapter 13 returns to eight questions for their reopening lines alone and applies a harder test than the ledger's, asking whether a reader outside the institution could reach the measure and the evidence a line turns on.
The fourth sub-question asked whether the same series reads differently in the collection's four layers of time: 2026, the 2030s, the 2040s, and the twenty-second century and later. The horizon row is the place where the layers meet the ledger, and the coders agreed on it in 40 of 50 cases, 80 per cent: days-to-months implied in 13 and stated in 9, years implied in 8 and stated in 5, decades implied in 3 and stated in 1, centuries-or-more implied in 1, with 10 unsettled. Hypothesis H6 — the longer the horizon, the less is decided now and the more conditions are set for revisiting — is recorded in the ledger not as a verdict but as a reading to be carried out across the books of Volume 0. What this study's ledger can supply is the shape of the count: conditions per case run 15.89 for days-to-months cases and 16.00 for years cases in the strict reading, against 15.91, 16.00, 16.00 and 16.00 across the four lenient horizons, with observable conditions per case running between 11.00 and 13.40 strictly and between 11.75 and 14.00 leniently. On these fifty questions the number of conditions barely moves with the length of the horizon. Whether the same holds across the other books is for Volume 0 to say. Chapter 12 reads one series across the four layers, adds testimony from the deepest layer, and returns to three series read earlier in the study to ask what changes in who decides, how long the waiting lasts and what can be undone.
Three hypotheses stand outside the sub-questions and belong to the answer in their own right. H5, that a restriction outlives the memory of its reason, goes against in both readings: supports 0, against 50, share 0 per cent, with the reason recorded in 50 of 50 cases. H7, that the more a decision is left to an AI or an automatic rule the more the human's place is empty or a name only, is not applicable across the whole collection: 50 of 50 cases count as not applicable in both readings, which follows from the human's place row, where 46 cases are read as human-decides and none is read as a decision left to an automatic rule. H8, that the options cutting the immediate loss most often give up evolvability and resilience among the four higher goals, is unclear: supports 20, against 28, not applicable 2, a share of 42 per cent, between the marks the verdict rule sets. The underlying count is 150 options across the fifty cases, with both coders agreeing that 82 reduce the immediate loss, and with what those options give up read as long-term-stability in 23, ethical-coherence in 22, evolvability in 17 and resilience in 12. The direction the hypothesis expected does not dominate these records.
So the study's verdicts are, exactly as the ledger gives them: H1 supports, strictly 25 supports against 0 with 25 unclear and leniently 42 against 0 with 8 unclear, share 100 per cent both ways; H2 against, 0 supports against 50, share 0 per cent; H3 supports, 13 supports with 37 not applicable, share 100 per cent; H4 supports, strictly 3 supports with 47 not applicable and leniently 30 supports with 7 unclear and 13 not applicable, share 100 per cent over decided cases; H5 against, 0 supports against 50, share 0 per cent; H6 to be read across the books in Volume 0; H7 not applicable, 50 cases; H8 unclear, 20 supports against 28 with 2 not applicable, share 42 per cent. Chapter 15 sets these out with the rule that produced them, and Chapter 16 sets the findings beside the theories the earlier chapters read.
Put together, the answers say something about the form in which a decision about war can be held before its answer. A record that names its decider, its mandate, its deadline, its unknowns and its conditions for reopening does two things at once. It makes the decision revisable in principle; this study reads that as an answer to what the literature on surprise describes, warnings not told from noise or not acted on in time. And it leaves the waiting invisible, because the deadline is written into the form and the bearer of the waiting is not; when a bearer is named at all, it is named as a category — those nearest the stoppage — rather than as people with a schedule and a claim. The study's own question was how such a record holds time when some answers cannot wait and others must be kept for generations. The records hold the urgent end well and the long end in outline: a deadline for every question, and conditions for revisiting in every case, running from a distribution cycle to a voyage measured in decades.
C.2 The contributions
Three contributions are offered, each narrower than the work it stands beside.
The first is a reading of the classic texts on war and its ending as texts about the time of a decision. Chapter 3 reads the second galley in Thucydides' Book III, a decree reconsidered the next day and a second ship that arrived just as the commander was preparing to carry out the first, beside Kant's first preliminary article on a peace that hides the next war. Chapter 5 reads Hobbes on war as a tract of time and on the covenant that makes a sovereign, Antigone on an edict and the unwritten law against it, and Clausewitz on war as an instrument of policy, on friction and on uncertainty. Chapter 4 reads Cassandra, a warning that arrives and changes nothing. Chapter 8 reads the Odyssey, where return turns on what has kept going meanwhile. The reading offered is that these stories are all about the interval between an answer existing and an answer being used, and about who can be reached inside that interval. That framing is this study's construction; the reading notes give the episodes and stop there. The modern literatures on ending wars, on warning, on civilian control and on justice after war are set beside the same episodes, each at the level its note supports, and each is asked a question it was not built to answer.
The second contribution is the ledger itself. Fifty defence questions on six axes set in advance and three more supplied by the material, three kinds of cell kept apart — what the record says, what the coders read, what the frame repeats — two coders of different families recording a value, a basis and a reason for each cell, and every count read strictly and leniently so that a reader can see how much rests on the record's own words and how much on inference. Chapter 2 sets the instrument out in full and codes two questions end to end. The frame rule is part of the contribution: the studio's repeating lens is identified and kept out of the findings, and its measured share in the reopening lines is 0 of 798, which is one reason those lines can be counted at all. The question form guarantees a decider, a deadline and reopening lines in every case; their presence is a property of the form, and the ledger says so rather than claiming it as a discovery.
The third contribution is what the counts show about these records, stated as findings about Meridian's synthetic questions and about nothing else. Every number in this study comes from the ledger of the fifty cases and carries what it counts: 50 of 50 questions with a deadline, 50 with what is missing listed, 148 missing items in all, 798 conditions for revisiting of which 629 carry a number, a threshold or a time, 13 of 50 questions where the published judgments chose more than one option, agreement between the coders ranging from 28 of 50 on the decider to 50 of 50 on the recorded reason. These are descriptions of a collection assembled by a single platform and a single date. They are checkable, and they are small. A finding about how a decision form writes the time of a decision is worth having, provided it stays a finding about the form.
C.3 The limits
Four limits bound every answer above, and each is a property of the material rather than a difficulty that more care in reading could remove.
The questions are synthetic. Each setting names a fictional polity, a fictional committee or council, and a fictional situation; the premise policy states that organisations, events and figures are settings for decision-making, and the questions say of themselves that they are not reports of actual events. Nothing in this study describes a real state, a real force, a real war or a real person, and no inference about how any of those behave can be drawn from these counts. The value of such questions is that they hold a decision before its answer, with the deadline, the options, the trade-offs and the reopening conditions all in one document; the cost is that nothing happened afterwards, so the ledger cannot say whether a decision was right, whether a reopening condition was ever met, or who actually waited.
The judgments are AI-generated. Thirteen accompany each question, from eleven named perspectives, a staff synthesis and a judgment made alone; in the records examined they carry a single timestamp and revision 1, which makes them a first published pass rather than settled verdicts. Their agreement therefore measures something about generation as much as about judgment. The ledger records published dissent in 13 of 50 cases, and Chapter 2 says outright that this is a count of published dissent, not a measure of how divided a real staff would be. Where the judgments part, this study reports both sides and picks neither.
The coders are also artificial. Two models of different families read each record and recorded a value, a basis and a reason; a cell is settled when both chose the same value, and unsettled cells count as unclear in the verdicts. Agreement runs from 28 of 50 cases on the decider to 42 of 50 on who bears the waiting, 40 of 50 on the horizon and on the reversal constraint, 37 of 50 on the reversal action, 46 of 50 on the human's place, and 50 of 50 on the human role and on the recorded reason. Those numbers describe how two models read the same words. They do not describe how people read documents, and the twenty-two unsettled decider cells are the clearest sign of where interpretation, not evidence, does the work.
The collection is one platform's defence questions published by 2026-10-05: fifty cases, all marked synthetic, twelve series of four with two more in two of the series, and no selection inside that set. The counts describe that collection on that date. The settings run from 2026 to 2194, with twelve cases set in 2026 and the rest distributed across the following years and centuries; the spread is an authoring convention, not a periodisation of conflict. Finally, this study is written by AI from Meridian's records and from reading notes supplied for the works it cites. Its readings of the classic and modern texts stay inside those notes, and where the study goes past a note it marks the step as its own. A reader who wants the tradition itself should go to the works; the bibliography lists only what this study used.
C.4 Further research
Four lines of work would carry these questions further than this study can.
The first runs across the series. This study is one of forty-nine in "The Time of Questions and Answers", and each book codes its own domain on the same axes. H6 is recorded in the ledger as a reading to be made across the books of Volume 0 rather than a verdict any single book can give: whether the longer horizon brings fewer decisions now and more conditions for revisiting is a question about a distribution across domains, not about fifty defence questions. The same applies to the comparisons Chapter 15 raises — this ledger set beside the WSTiger ledger and the others in the volume, so that hypotheses about waiting, about undoing and about the observability of reopening can be checked against collections whose question forms differ in what they guarantee. Where a form guarantees a decider and a deadline, as Meridian's does, the interesting counts are the ones the form does not force: the bearer, the observability of a condition, and where the judgments part.
The second is human coding. Two models of different families produced the agreement rows used here. A study in which people read the same fifty questions and record values, bases and reasons on the same axes would show which cells are genuinely in the record and which are inferred by the reader. The decider row is the obvious candidate: 22 unsettled cases out of 50 is high enough that human agreement might settle most of them, or might reveal that the question of what kind of body a fictional council is cannot be answered from the text at all. The bearer row would test something deeper, since it asks who waits, and a person reading a record may name where a model named a category.
The third is real public records. Decision documents of the kind Meridian's form imitates — cabinet conclusions, emergency orders, procurement approvals, access regimes for archives, sunset clauses in delegated authority — exist in public archives and in official publications. Coding a set of them on the same axes would show which of this study's findings survive when the question form does not guarantee the decider, the deadline and the reopening lines. Two of the findings are ready-made tests: whether a bearer is named at all when the form does not ask for one, and whether a reopening condition is written in words a reader outside the institution could observe. On Meridian's collection, 798 conditions exist and 629 carry a number, a threshold or a time, while the observable share falls from 602 of 650 in the judgments' stopping conditions to 27 of 148 in the questions' own reopening field. That gap is a hypothesis about forms worth carrying to documents that were written to be acted on.
The fourth is the material the ledger counts but does not describe. The 148 items listed as missing are counted across the fifty cases and never characterised; the four higher goals given up by the options — long-term-stability in 23 cases, ethical-coherence in 22, evolvability in 17, resilience in 12 — are recorded as coder readings of options rather than as arguments. A study that reads what is actually missing in each question, and how each option's trade-off is argued rather than labelled, would move from the form of the record to its content. So would a study of the dissenting judgments: 13 questions where the published judgments chose more than one option are counted here and read in outline in Chapter 14, but what a lone dissent weighs against a synthesis is a matter of argument, and the ledger records only the shape of the split.
C.5 What would change this study's answers
A book that asks what would reopen a decision owes its readers its own conditions for reopening. The ones below are written so that a reader could check each of them without access to anything but the published records and the ledger.
If Meridian publishes defence questions after 2026-10-05 and the collection leaves 50 cases, every count in this conclusion changes with it. The tests are the rows themselves: 50 of 50 with a deadline, 50 with what is missing listed, 798 conditions for revisiting, 629 of 798 stated with a number, a threshold or a time, 13 of 50 with judgments choosing more than one option. If any of those counts moves when new cases are added, the answers above must be re-run rather than extended.
If human coders read the same fifty questions and settle the decider cell differently from 28 of 50, the first sub-question's answer changes shape. The specific threshold is the 22 unsettled cases: if human readings settle fewer than half of them in one direction, the claim that the form names a decider while the kind of decider stays open becomes a claim about these records' wording rather than about two models' disagreement.
If a recoding settles the 25 cases left unclear in H1's strict reading in a way that falls below 60 per cent support among decided cases, H1's strict verdict moves off supports. The same test applies to H4, where the strict verdict rests on 3 decided cases out of 50, and to H3, where it rests on 13. Each of those shares reads 100 per cent over decided cases; each is sensitive to a small denominator, and the ledger says which cases carry it.
If revision 2 of the judgments is published and the divisions differ from the 13 of 50 recorded here, the answer on where the judgments part changes, and so does H3 with it. The present judgments carry a single timestamp and revision 1. A second pass that parts in a different set of questions would show how much of the agreement is the frame's and how much is the argument's.
If the observability shares move, the answer on reopening changes. The two figures to watch are 602 of 650 observable in the judgments' stopping conditions and 27 of 148 observable in the questions' reopening field, together with the 0 of 798 whose wording is 30 per cent or more the studio's frame. A new coding, or a stricter test of what an outside reader could reach — the test Chapter 13 applied to eight lines — could reduce the observable counts sharply and reverse H2's verdict of against. H2 stands on the record saying yes to both halves of its question: a condition exists, and it can be observed. Anything that undermines the second half reopens the verdict.
If Volume 0's cross-book reading settles H6 or H8, two of this study's open questions close from outside. H6 is marked in the ledger for reading across the books of the volume; the conditions-per-case counts here — 15.89 to 16.00 across the horizons, strictly and leniently — supply one book's share of that reading. H8 is unclear at a share of 42 per cent, between the rule's marks, on 150 options of which 82 are agreed to reduce the immediate loss. A larger option set, or a domain whose options trade differently, could carry it to either side.
And if any of these records were ever shown to describe a real state, force, war or person, the study would need to be withdrawn rather than revised. Every setting here is synthetic, every judgment AI-generated and dated, every reading of the works confined to notes written for this purpose. The counts describe how a question form holds the time between a question and its answer in fifty fictional settings. That is what they were made to do, and it is all they do.