Chapter 2. Method: The Decision Ledger and the Thirteen Judgments

This chapter sets out how the study reads Meridian's fifty defence questions. It describes the fields of a question and of a judgment, the nine axes on which every question is coded, the three kinds of cell and the two coders, the strict and lenient readings, the eight hypotheses and the rule that turns counts into verdicts, the frame that is kept apart from the findings, and two questions coded in full. This study recommends no option. It reports what the records contain, what two coders could agree on, and where their readings part.

2.1Questions as evidence

A Meridian question is a document, not an event. Each is stored under a code, with a title, a pack, a domain, a decision year and an assumption level. Every question in this study's collection is marked synthetic: its setting names a fictional polity, its premise policy states that organisations, events and figures are settings for decision-making, and its own words say that the issue posed is not a report of an actual event. The study quotes the records only from that setting. Nothing in them reports a real state, force, war or person.

The fields fall into groups. Identity and premise fields say what kind of matter this is: title, pack identifier, domain, family, decision year, horizon, assumption level, setting, premise policy. The decision fields say who must act: a decision actor with a role, a mandate that limits what that actor may decide, a deadline, a decision type and a decision question. A decision brief carries four more fields: why now, stakes, boundaries, and reopen-when lines. Three premises (numbered P1, P2, P3) state the synthetic conditions the decision must work under, and each carries the label "synthetic setting".

Then come the fields that make the question a decision rather than a complaint. Options: in these records three per question, each with a label and a tradeoff that names a gain and a loss. Unknowns: each marked "unverified", each phrased as a question the setting cannot answer. Watch questions: three, asking who decides and who can object, who bears the losses, and what independent confirmation would reopen the decision. Materials: background documents whose scope states what they do not prove, a symbolic or interpretive frame chosen by the author, and a value or norm that the decision-maker must argue. Source references come with their own scope statements. A response request lists six forms of answer the record asks for. And two updates, labelled independent branches, describe synthetic follow-ups after the initial input, such as a recovered register or a change of authority. They are later developments in the setting, not outcomes of the decision.

The thirteen judgments come after. Each has a slot — the chief of staff, the source officer, the domain analyst, the independent challenger, the authority counsel, the operations planner, the synthesis secretary, the decision-quality auditor, the principal's proxy, the outcome auditor, the improvement architect, then a staff synthesis and a judgment made alone — a selected option, a confidence level, a revision number and a creation timestamp. Eleven of the slots are perspectives, and the remaining two are a synthesis and a single-pass judgment. Each judgment answers with eight claims: decision, scope, accepted loss, conditions, stop conditions, milestones, uncertainty, counterposition. Each claim carries a kind (a value judgment, an inference, an unproven hypothesis, an unverified item) and a basis that says which premise or unknown it rests on. The four kinds are not decoration. A value judgment marks a claim that rests on a choice of what to protect; an inference marks one that follows from the setting by reasoning the basis makes visible; an unproven hypothesis marks a claim that is assumed and named as assumed; an unverified item marks one the material does not settle. A reader can therefore sort a judgment into what it chose, what it reasoned, what it assumed and what it left open, and check the sorting against the basis each claim carries. Every judgment here is AI-generated, dated, numbered by revision, and revisable: a later revision could replace it, and the record does not promise that it will not.

What such records can show is how a decision is posed and judged before its answer. They show the fields the posing required, the losses each option names and to whom, the unknowns the setting admits it cannot fill, and the conditions under which a published judgment says it should be reconsidered. What they cannot show is what happened. There is no outcome field. No question reports that a roster was finalised, that an alert was corrected, or that a family was paid. The study therefore treats each record as a decision frozen at the moment before its answer, held open by its deadline, and treats the thirteen judgments as a dated, revisable record of how that moment was judged.

2.2The six axes and three more

An axis is one question the ledger asks of every record, in the same words, so that fifty different settings can be compared without any of them being turned into a case study of war. Six axes were set before coding; three more follow from what the judgments and options themselves supply.

A1, who decides. The record names a decision actor and a mandate in every question. The coders then classify the level of that actor: an institution or state, a team or organisation, a multilateral body or civilisation, or unsettled. The record writes the actor out in words; the level the coders assign to that name is a code cell. The ledger reports that all 50 cases carry a decider in the record.

A2, what is missing. Each question lists its unknowns, each marked unverified. The ledger counts the cases that list what is missing and the items listed in all: 50 of 50, and 148 items. Listing the gaps in a fixed field is a requirement of Meridian's question form rather than a finding about these records; what the axis adds is what each question admits it cannot fill.

A3, the deadline, and who waits. The deadline is written out in the record: 50 of 50 cases carry one. Around it sit two coded rows. The horizon row records how far the decision's effect runs: days to months, years, decades, or centuries and more. The bearer row records who waits and pays while the decision is pending: those nearest the point where things stop, third parties or the vulnerable, or future generations. The deadline itself measures only one instant. The waiting before it and the cost after it are what the bearer row is for.

A4, what can be undone. Two rows here, and behind them two different objects of undoing. Reversal constraint asks how costly it would be to undo a restriction — to withdraw a rule or lift a limit the decision imposes. Reversal action asks what undoing what the decision permits to happen would involve, with values easy, costly or impossible. The two are not the same: a rule can be withdrawn in a day, while a payment already made or a record already released cannot be taken back by withdrawing anything. The distinction matters because a decision can be easy to reverse and still have done its damage in the meantime, and because the same answer to one row does not settle the other.

A5, conditions for reopening, and whether they can be observed. Every question writes conditions under which the judgment should be reconsidered. A condition is counted as observable when it is stated with a number, a threshold or a time, and not when it names only an event or a role. This is the axis the study cares about most in its later chapters: a condition worded so that no one outside the room could check it can be confirmed, if at all, only by the reviewers the records name, and a reader cannot tell from the wording alone whether it has been met.

A6, the layer of time. Meridian's defence questions sit in four layers: 2026; the 2030s; the 2040s; and the twenty-second century and later. The horizon row supplies a second reading of time that does not depend on the setting's date. The layer is read from the setting's year, which the record gives; the horizon is read from the reach of the decision's effect, which the coders assign. The two can part: a question set in 2026 can bind people for a decade, and a question set in the twenty-second century can be decided in an afternoon. Keeping both rows is what lets the study ask later whether the length of an effect has anything to do with how much is settled now.

Three further axes come from the records' own contents. A7, divergence, reads the thirteen selected options written into each question. Divergence is the count of questions whose published judgments chose more than one option; the ledger gives it as 13 of 50 cases, and notes that there are 13 judgments per case. A8, the human's place, asks where a human sits in the matter decided: approving, granting an exception, reviewing. The ledger's human-role row reports 50 of 50 cases in which a human role is stated, and its place row classifies the placement as human-decides, with a small number unsettled. A9, the four higher goals, reads what each option gives up: long-term stability, ethical coherence, evolvability and resilience. The coders record a give-up only where both of them read it in the option's wording.

What an axis cannot do. Each axis compresses a question to one value, and the compression loses things. A1 records a level, not the mandate that limits what the actor may decide: two committees at the same level can hold different authorities, and the mandates in these records are written out in words the code does not carry. A3 records a horizon and a bearer, not the shape of the waiting. A4 records cost classes, not the acts that would undo a decision. A5 records whether a condition is worded observably, not whether the condition would be met. A7 records how many questions divided their own judgments, not why they divided. A9 records a give-up only where both coders read one in the wording, so an option that gives up evolvability in fact but not in words is not counted. The axes let fifty questions be compared without any of them being read as a case study of war; they do not replace reading the question.

Direct fields and coded cells sit inside the same axis rather than dividing the axes into two groups. The record supplies the actor's name, the deadline, the listed unknowns, the condition texts, the setting year, the three options and the thirteen selected options of the judgments; the codes assign the level of the decider, the class of the bearer, the horizon, the reversal rows and the observability of a condition, and it is on the codes that two readers can disagree.

2.3Three kinds of cell and two coders

The ledger holds three kinds of cell, and keeping them apart is what makes the counts checkable.

A record cell is copied from a field without judgment: the decision actor, the deadline, the three option labels, the thirteen selected options, the number of conditions to revise. A reader can verify any of these against the record in the same minute. A code cell is an axis value the ledger assigns to a question: the level of the decider, the class of the bearer, the horizon, the reversal rows, the observability of a condition. A model cell is not a separate kind of cell but the evidential basis of a code cell: it marks whether the coders read the value off a field or had to infer it from the setting. Decider level, bearer, horizon and the reversal rows are code cells either way. The ledger marks the difference in its own values column: a value found in the record's words is marked stated, a value inferred from the setting is marked implied, and a value the two coders did not settle is marked unsettled.

Two coders did the work. They are two models of different families, and this study never names them. Each coder returns three things for every cell: a value, a basis and a reason. The value is the axis answer. The basis says what in the record it rests on. The reason says why that answer and not another. Where the two values agree, the cell is settled and the ledger records whether both coders found it stated or implied. Where they disagree, the cell is unsettled, and unsettled counts as unclear in the verdicts.

How the coding ran. Each coder worked from the same question record and the same axis wording. The ledger does not average the two values and does not pick between them. Where they agree, the cell is recorded once, with its basis; where they disagree, the cell stays unsettled and the case is lost to the unclear column for that hypothesis. The agreement column below is therefore a count of questions on which two readers reached the same answer, out of the fifty in the collection, and not a score for either reader. A cell can be settled on a stated value or on an implied one, and that difference is what the strict reading later counts.

Why a basis and a reason travel with each value. A value alone is hard to check: two coders can assign the same class to a committee for different reasons, and a later reader could not tell which reading of the record each had in mind. The basis names the field or premise the value rests on; the reason says why that answer and not another. A disagreement can then be located — on the meaning of a title, on the reach of a mandate, on the wording of an option's tradeoff — and a reader who disagrees can go back to the same words. This is why unsettled cells are kept rather than dropped. Dropping them would make the agreement rows look better than the coding was; silently choosing one value would hide the places where the record supports more than one reading.

Agreement varies sharply by cell, and the ledger reports it row by row:

The lowest cell is the decider, at 28 of 50. This study reads that low agreement as a fact about the coding and not about the records. The record names an actor in words; the coders must decide what level that name describes, and a committee named in a title does not by itself say whether it is an institution, a team or part of a multilateral body. The next lowest cell is reversal action, at 37 of 50, where the two rows of A4 can also be read apart: a coder may treat an option as costly to reverse while the other treats the reversal as impossible. The two highest cells are the human's role and the recorded reason, both 50 of 50 — a fact this study treats as the form's doing, not the scene's.

The options are read the same way. The ledger records what the options give up, again only where both coders read it: long-term stability 23, ethical coherence 22, evolvability 17, resilience 12. The ledger does not turn the four counts into a ranking.

2.4Strict and lenient, and the hypotheses

Every count in the ledger is read twice. The strict reading counts only what both coders found stated, meaning found in the record's own words. The lenient reading also counts what at least one coder found implied, meaning inferred from the setting. Reading both matters because the two readings answer different questions: the strict one asks what the records say outright, the lenient one asks what the records would still support if inference is allowed. Where a hypothesis flips between the two, the study reports both readings and does not choose between them.

Eight hypotheses were set before coding, in the words of "The ledger's numbers", each with the question the coders asked:

The verdict rule is fixed. A hypothesis is supported where 60% or more of the cases it applies to supports it; it goes against where 40% or less does; the space between is unclear. A verdict is also unclear where the unclear cases outnumber the decided ones.

What the rule does with cases that do not apply. A hypothesis does not speak to every question. H3 reads only the questions whose published judgments chose more than one option, and the other 37 are marked not applicable rather than counted as support or as against. H4 carries 47 such cases in the strict reading and 13 in the lenient one, and the table lists them, so the applicable set is visible and a reader can see how thin it is. The thresholds do the rest of the work: 60% and 40% are set before the counts are seen, so a verdict cannot be moved by drawing a friendlier line afterwards, and the unclear band keeps the rule from forcing a result where the cases are evenly split.

H6 is not decided by this rule inside this volume; the ledger marks it for reading across the books in Volume 0, and gives its by-horizon counts instead.

H8 is the one hypothesis with an unclear verdict in both readings: 20 supports against 28 goes-against, and the share of supports is 42%. H6 is not decided inside this volume, and H7 is marked not applicable here, so neither of them is decided by these readings either. The hypotheses are reported as the ledger gives them. Where a verdict goes against a hypothesis the ledger registered, that is a result, not an error.

2.5Frame and finding

A frame is a lens that repeats across scenes: a fixed set of fields (title, assumption level, premise policy, comparison, watch question, usage, material, published, scope, response request, relative to) that a studio reads through a question before it judges it. In this study's records the frame's wording is spread across several fields rather than standing in one place: the premise policy, the watch questions, the usage and scope lines of the materials, and the response request. The materials field is one of the fields that carry it, and it also carries case-specific material that is not frame wording — a symbolic or interpretive frame chosen by the author, such as River of Memory in the Sand River question or Cassandra in the alert question, and a value or norm that the decision-maker must argue. In the Sand River question, for example, the record supplies this norm:

The frame rule is simple. The frame is the studio's way of reading. It is not a finding about any one question, and the fact that it repeats across scenes is a fact about the studio, not about organisations, wars or decisions. A finding, by contrast, is something the records show: a count, a cell, a field, a place where two coders part. Where this study describes the frame, it does so to say what must be kept apart from the counts, and never to convert a frame sentence into a result.

What the measurement covers. The count of 798 takes the conditions for reopening and stopping that the fifty questions contain, in the two places they live. It does not measure the watch questions, the premises or the response request, and it does not say how much of any one judgment is frame wording. It says one narrow thing: that no condition line crosses the 30% threshold the ledger set before the count was taken. The frame rule needs nothing wider, because the risk it guards against is the other one — a frame sentence read out as a finding about organisations or wars.

Meridian measured its own frame share directly. Of the 798 conditions for reopening and stopping that the fifty questions contain, the ledger reports that 0 are cases where the wording is 30% or more the studio's frame. That measurement covers the reopening lines in both places they live: 650 conditions sit in the judgments' stop-condition field, of which 602 are observable, and 148 sit in the question's reopen-when field, of which 27 are observable. No condition, in other words, is 30% or more the frame's wording.

The question form guarantees certain things, and this study refuses to call them findings. Every question names a decider because the form asks for a decision actor: 50 of 50 carry one. Every question names a deadline because the form asks for one: 50 of 50 carry one. Every question carries reopen-when lines because the form asks for them, and the ledger counts 148 such lines across these fifty records. Their presence is therefore a property of the form. What is not guaranteed is what lies inside them. The observability row shows that 602 of the 650 stop conditions and 27 of the 148 reopen-when lines are worded in terms a reader could check. That difference, and not the existence of the lines, is what the ledger can say.

2.6Two questions coded

Support before the name

Synthetic scene

In the fictional Sand River region the Sand River Family Recovery Committee owes the quarterly payment roster a criterion before it is finalised at 16:00 on 5 November 2026. Eight hundred households have applied; half hold official receipts and half say their records were lost. The record names six review staff, a livelihood-support person in charge and a review lead, without saying how those roles divide; what is waited for is a criterion that says who is paid first. After the deadline the roster is finalised, and until then every application is held.

The record's own question:

The staff synthesis, dated 9 September 2026, at revision 1, with confidence set to medium, chose O3, and its decision claim reads:

And one of the record's three reopen-when lines:

The cells. A1: the record names the fictional Sand River Family Recovery Committee and a mandate that sets interim eligibility only and excludes confirmed death and final compensation; the ledger's decider cell for this question is unsettled. A2: the record lists three unknowns, all marked unverified — the share of households with lost records who can reach alternative evidence, the real spread of urgent need, and the scale of duplicate and relationship disputes. A3: the deadline is 16:00 on 5 November 2026; the horizon is days to months; the bearer is those nearest the stoppage. A4: reversal is coded costly. A5: 16 conditions to revise, and the reopen-when line quoted above names a role, a period and an event. A6: the 2026 layer. A7: all thirteen judgments chose O3, so this question is not among the thirteen divided ones. A8: the place is human-decides, and the human role is stated. A9: the options name immediate losses to named parties — delay for the recordless (O1), slower payment for everyone (O2), deferred verification and equity disputes (O3). The wording names none of long-term stability, ethical coherence, evolvability or resilience, and the coders record a give-up only where both read one in the wording.

Where the coders differed. For this question the disagreement is at the level of classification, not at the level of the record. The record names the decision actor in words; the coders did not settle what class of decider that name belongs to. The decider cell is the lowest cell in the ledger, unsettled for 22 of the 50 questions. This study reads that as a limit of the coding vocabulary and not as silence in the record, which names its committee and its mandate on every page.

What this question's record can and cannot settle. It fixes the deadline, the budget ceiling and the staffing, and it fixes the shape of the choice: three eligibility criteria, each naming who waits. It does not say how many households can reach alternative evidence, how far urgent needs differ, or how large the duplicate and relationship disputes will be, and those three unknowns are what would tell the committee whether the brief review is faster than a fuller one. The thirteen judgments accordingly choose the same option and mark the same assumption as unproven. The record can show that the assumption was named; it cannot show whether it held.

Withdrawing a crisis alert

Synthetic scene

In the Virtual Nuri border-livelihood zone the deadline for the Crisis Communications Committee is the 9:00 a.m. regular broadcast on 10 September 2026, and an overnight alert has already stopped shared transport. Internal cross-checking has found double counting in the timestamped material, while a separately received witness statement remains unverified. Review staff can examine only one body of material — the full source or the summary — before the deadline. Resident representatives wait for an explanation they can use to decide whether to return; a counterpart liaison organisation waits for material it can verify; whistleblowers wait to learn whether their request to defer publication has been heard.

The record's own question:

The staff synthesis, dated 9 September 2026, at revision 1, with confidence set to medium, chose O1:

The cells. A1: the record names the Virtual Nuri Crisis Communications Committee, whose mandate covers its own alerts and the release of material, and the ledger classifies the decider as institution or state. A2: three unknowns, all marked unverified — whether the remaining witness statement is independent, whether the counterpart organisation will accept verification on de-identified material alone, and how residents will respond to the next alert. A3: the deadline is 9:00 a.m. on 10 September 2026; the horizon is days to months; the bearer is those nearest the stoppage. A4: reversal is unsettled for this question. A5: 16 conditions to revise; the record's reopen-when lines name roles, periods and events. A6: the 2026 layer. A7: all thirteen judgments chose O1, so this question is not among the thirteen divided ones. A8: human-decides, with the human role stated. A9: the options name immediate costs — doubt that may persist at the counterpart organisation (O1), delayed public explanation for residents (O2), and review staff tied up on other alerts (O3). The wording names none of the four higher goals.

Where the coders differed. Here the settled cell is the decider and the unsettled one is reversal. The ledger marks this question's undo value as unsettled, without saying which of the two reversal rows the disagreement sits in or which values the two coders chose, so the record does not show whether they differed about the cost of undoing the decision or about what undoing it would involve. The reversal-action row is unsettled for 13 of the 50 questions in the aggregate, and that aggregate does not settle what happened in this question. This study reads that as the sharpest kind of methodological limit: a cell the coders could not settle, where the record's option trade-offs name the costs of choosing and implementing each option rather than the cost of changing it afterwards, so the record does not supply the reversal cost that would settle the classification.

What this question's record can and cannot settle. It fixes the deadline and the constraint that the review staff can examine only one body of material before it, and it names the parties to the loss: residents deciding whether to return, a whistleblower asking for deferral, a counterpart organisation wanting material it can verify. It does not say whether the remaining witness statement is independent, whether the counterpart organisation will accept de-identified verification, or how residents will treat the next alert. The correction the judgments adopt rests on the assumption that a summary can explain an error while protecting its source. The record shows the assumption and its limits; it cannot show whether the summary did both.

2.7Limits of the method

The method is built on synthetic questions in fictional polities, and it therefore cannot describe what any state, force or institution did. Its judgments are AI-generated, which explains in part why the thirteen judgments so seldom part: agreement among judgments produced under a shared frame and a shared inference plan is partly the generator's agreement, and the ledger's A7 row — 13 of 50 questions with judgments choosing more than one option — is a count of published dissent, not a measure of how divided real staff would be. Its coders are also AI models, so the agreement rows are a measure of two models reading the same words, not of two people reading the same records. The collection is one platform's defence questions published by a single date, so the counts describe that collection and nothing wider. And this study is itself written by AI: its readings of the classic works stay inside the reading notes given for them, and where the study marks an interpretation, it says so.

What the method cannot do is worth setting out. It cannot say which option a committee should have chosen: the questions are built so that each option names its own losses, and the ledger records those losses instead of ranking them. It cannot say whether any judgment was right, because the settings are synthetic and no outcome field exists to be right about. It cannot compare these fifty questions with the defence questions of any other platform, any other year or any other authoring run. It cannot measure how far the thirteen judgments are independent of one another, since they share an input and an inference plan, and their agreement may belong to that sharing as much as to the questions. And it cannot settle a coding dispute by authority: where the two coders parted, the ledger records the parting and the reader is left to read the record again.

What the method does offer is narrower and checkable. Every count comes from "The ledger's numbers" and names what it counts. Every axis is applied to all fifty questions in the same words. Every cell distinguishes what the record says from what a coder inferred, and every verdict is read strictly and leniently. A reader who disagrees with a coding can find the axis, find the case and find the field.

Chapter 3 takes the classic texts on ending wars and reads them with the ceasefire questions.