Chapter 2. Method: The Decision Ledger

This chapter sets out how WSTiger's records were turned into counts. It describes what a record holds, the axes on which every record was read, the three kinds of cell the ledger contains, the rule by which a hypothesis is called supported, unclear or against, the rule that keeps a studio's writing habits out of the evidence, and the coding of two records in full. The study recommends no option. For each alternative in a record it shows what the alternative gains, what it risks, who is close to the cost and when, whether it can be undone, and what the record itself says would make its example judgment be reassessed. Every count in this chapter and in the chapters that follow comes from the decision ledger computed from the published records; nothing else in this study counts cases.

2.1Records as evidence

A WSTiger record is a small, fixed bundle. It has a Scenario: the setting in which a decision has to be made. It has a Question: the decision put as a choice. It has three Options, named as alternatives rather than ranked. It has an Example judgment: WSTiger's own reading of what to do, offered as an example and not as a verdict the reader must accept. It has an Invalidation: the line saying when the judgment should be revisited. Alongside these fields the record carries its studio, its domain, a time-of-setting marker, its kind (every record used in this study is marked synthetic), a pack identifier, the locales in which it exists, a content hash and a fetch timestamp. Those last facts let a reader confirm that this study has read the record that WSTiger published, and not a later or altered version.

Every record used in this study is a synthetic setting. It describes no real firm, no real trade and no real person. WSTiger publishes such settings for study. The example judgments are WSTiger's readings of the situations the studio writes; they are dated and revisable, which means they are conclusions someone reached, not facts about how the world turned out.

What a record can therefore show is the shape of a decision written before its answer. A reader can see which alternatives the record names, what grounds it offers for choosing among them, what it leaves unnamed, and what condition it sets for coming back to the choice later. The record holds the moment at which a question is open: the options are still options, the example judgment is a reading, and the invalidation line is a promise that the reading can be taken back. The Introduction sets out why those moments are worth holding onto: histories written afterwards lose them.

What a record cannot show is what happened. There is no outcome in any of the 53, no profit or loss attached to an option, no statement that an option was chosen and no test of whether the example judgment was good. A record therefore cannot tell a reader which option is right, and this study does not say. What it can tell is how the decision was written: what the record gives as grounds, what it omits from those grounds, how it treats the time before the answer arrives, and how carefully it states a condition for revisiting.

Two features of the records matter for everything that follows. The first is that a record's form is shared across a studio's records. A field that appears in every record was asked for, so its presence tells a reader about the form rather than about the decisions. Section 2.5 sets out that rule in full; it is mentioned here because it governs how every count later in this study may be read. The second is that the Example judgment and the Invalidation come from different places. The example judgment is a reading of a situation; the invalidation is a statement of what would make that reading be questioned. The two records this chapter codes in full show how differently an invalidation can be written: one is a short clause pointing to a control deadline, the other is a change condition followed by a next check.

The reading of the examples belongs to WSTiger. The readings through the works belong to this study. The counts belong to the ledger. Keeping those three apart is what allows the chapters that follow to place a record beside an author's argument without pretending that either settles the other.

2.2The six axes and three more

Each record in the ledger carries a value for six axes, read from the record's own words, and three further values that the case chapters use alongside them. A value comes with a basis: stated means both coders found it in the record's words, implied means at least one inferred it from the setting, and none means neither found it.

A1, who decides. The decider axis records whether a record names the person or body that decides. None of the 53 records lists a decider. The axis is therefore read from the setting's language: an office, an operator, a protocol, an agent. The record's own wording may name a role, or it may leave the decision floating between a system and a person who could approve it. Chapter 8 reads this axis in full; here it is enough to say that the axis is read by inference rather than copied from the record in nearly all of its cells, which is why the coders' agreement on it (33 of 53 cells, 62%) is a measure of how similar two readings of an unnamed setting can be.

A2, grounds and what is missing. This axis asks on what a record's decision rests, and what it leaves out. A record grounds its decision in its scenario, in the evidence it names, and in its example judgment. What is missing is the harder half: a decision needs material, roles or intervals that the record does not supply. The ledger records this in two ways. It records whether a reason for the decision is present, and it records whether the record lists what is missing. In these records no case lists what is missing; the ledger's row for that count reads zero, and the count of items listed in every record is zero as well. The reasons, by contrast, are usually there: the reason-recorded row shows 49 of 53 cells (92%) on which the two coders agreed.

A3, deadline, horizon, the cost of waiting and who bears it. This axis carries four separate questions. Is a deadline written into the record? None of the 53 has one. How long is it before an answer can come? The horizon cell takes one of the values unsettled, unknown, days-to-months or years; across the 53, the coders settled on one value in 30 cells (57%), and the values they settled on were unknown in 18 cells, days-to-months in 11 and years in 1. What does waiting cost? The ledger answers this through the options, read for what they do to the account or the arrangement while the answer is not in. And who bears that cost? The bearer cell takes a value such as the decider, those nearest the stoppage, third parties or the vulnerable, or unknown; the coders settled on one value in 24 of the 53 cells (45%), which is among the weakest cells in this ledger. Chapter 10 treats the waiting; this chapter only defines the axis.

A4, whether the decision can be undone. The axis splits in two, because a restriction and the thing it stops are undone at different costs. Undoing the restriction means withdrawing the rule, limit or requirement again; the reversal-constraint cell reads easy, costly, unknown or unsettled. Undoing what the restriction lets happen, or stops, means recalling an act that has already been done; the reversal-action cell reads easy, costly, impossible, unknown or unsettled. A limit can be lowered again in minutes while the exposure it allowed cannot be un-taken at all, or the reverse: a restriction may be irreversible even though nothing has yet been done under it. The two cells are coded separately for that reason, and they are the weakest cells in the ledger: the coders settled on one value for the constraint in 11 of 53 cells (21%) and for the action in 13 of 53 cells (25%). Chapter 11 reads this axis in full.

A5, the conditions for changing course. Every record carries an Invalidation line, and the axis asks two questions about it: is there a condition, and could a reader observe it? A condition counts as observable when it is stated with a number, a threshold or a time, which is what the ledger's code can find; a condition that names only an event or a role is not counted, so the count is a floor. Across the 53 records the condition cell is present in 53 of 53 (100%), and the code-level count of conditions a reader could observe is 5 of 53 (9%).

A6, the time of the setting and the time of the judgment. Two times run through a record. The first is the time of the setting, which the ledger records as present, past or future: 35 of the 53 settings are present, 9 are past and 9 are future. The second is the time of the judgment, which is the moment at which the example judgment is written. Because no record carries an outcome, these records do not report how any of their choices ended; the decision each setting holds open is unresolved, and the date on which WSTiger wrote its reading is a separate matter that the records leave aside. The markers the records themselves carry differ in style, and the two cases coded below show both: one scenario line is marked 1995, the other 2026, in language that belongs to the studio's pack rather than to the setting. Chapter 12 returns to the settings' clocks.

A7, divergence. This axis records where judgments come apart: where two readers, or two coders, read the same setting differently. In this ledger divergence is measured against what a record lists as missing. Since no record lists what is missing, the hypothesis that ties divergence to missing material (H3) is not applicable in all 53 cells under both readings; the ledger reports that plainly and Chapter 14 repeats it. That absence of listings keeps this ledger from testing H3; it does not make other divergence uninformative, since readers can disagree about the evidence, the values or the interpretation a record does supply.

A8, the human's place. This axis records where a human appears in the decision: whether a human decides, approves, reviews, or is named without a role. The place cell reads human-decides, human-approves, ai-or-rule-decides, unknown or unsettled; the coders settled on one value in 33 of the 53 cells (62%). A separate cell records whether a human role is specified at all, and the human-role row shows that most cells read "no": 38 on an implied basis and 1 on a stated basis, against 1 implied and 1 stated that read "yes". Chapter 8 and hypothesis H7 are where this axis carries weight.

A9, the four higher goals. Every option is read for what it gains and what it gives up among four goals named in the ledger: resilience, long-term-stability, ethical-coherence and evolvability. Across the 53 records the options number 159, and both coders agree that 44 of them reduce the immediate loss. What those options give up, as both coders read it, is resilience for 18 options, long-term-stability for 17 options, ethical-coherence for 13 options and evolvability for 7 options. These four goals are the study's own vocabulary for what an option abandons in order to cut a loss now; they are not terms taken from any of the works cited later.

2.3Three kinds of cell

A ledger cell is not one kind of thing. Some cells are copied, some are counted, and some are read. Keeping them apart is what allows a reader to challenge a count without challenging the record.

The first kind is copied by code. These are the record's fields, taken character for character: the Scenario, the Question, the three Options, the Example judgment, the Invalidation, and the record's own metadata. No judgment enters into them, and any reader can re-derive them from the published record. When this study block-quotes a record, the words come from this kind of cell.

The second kind is counted by code. Two counts of this kind appear in the ledger, and both are mechanical. The first asks whether a condition for revisiting is stated with a number, a threshold or a time, which is what the ledger's code counts as observable. The ledger's answer is 5 of 53 conditions (9%). The second asks how much of a field's wording is shared with the rest of a studio's records, which is the frame measure taken up in Section 2.5. Its answer is that 2 of the 53 conditions carry wording that is 30% or more the studio's frame. Both counts are reproducible, and both are about the writing of records rather than about the decisions inside them.

The third kind is read, and these are the cells that need judgment. A1, A2, A3, A4, A5 as a case value, A6, A7, A8 and A9 all fall here, along with the cells for each hypothesis. Each of these was coded independently by two models of different families, and each coder gave, for every cell, a value, a basis (stated, implied or none) and a reason. Where both coders chose the same value the cell is settled; where they did not, the cell is unsettled, and unsettled cells count as unclear in the verdicts. The reasons are kept with the index so that a reader can see why two readings came apart.

Agreement varies sharply by cell, and the ledger reports it. In the table below, "settled" means both coders chose the same value; the value column lists what the cells held once coded, with each value's basis.

Where agreement is low it must be said plainly rather than smoothed over. The two reversal cells are the weakest: 42 of 53 constraint cells and 40 of 53 action cells were unsettled. The bearer cell is next, with 29 of 53 unsettled. By contrast, whether a record contains a reason is nearly always agreed, and the decider, place and human-role cells are settled in most records. The pattern itself is informative: coders agree most easily about what a record says on its face, and least easily about what a record implies about who pays for waiting or how hard a choice would be to withdraw. That is a fact about these cells, and Section 2.7 returns to what it does to the hypotheses.

2.4Strict and lenient, and the hypotheses

The ledger tests eight hypotheses. They were written down before the books in this series were drafted, so that the counts would test a statement rather than describe one. They appear here in the ledger's own words.

• H1: The cost of waiting falls not on whoever decides but on those nearest the point where things stop. (Is the bearer someone other than the decider?)
• H2: Few decisions write down in advance what would reverse them, and those that do write it in words no one could observe. (Is there a condition, and can it be observed?)
• H3: Where judgments diverge, material is missing. (Divergence against what is listed as missing.)
• H4: The harder a decision is to undo, the further apart are those who decide and those who bear the cost.
• H5: A restriction outlives the memory of its reason. (Is the reason for a restriction recorded?)
• H6: The longer the horizon, the less is decided now and the more conditions are set for revisiting.
• H7: The more a decision is left to AI or an automatic rule, the more the human's place (approval, exception, review) is empty or a name only.
• H8: The options that cut the immediate loss most often give up evolvability and resilience among the four higher goals.

A hypothesis is called supported when 60% or more of the cases it applies to supports it, against when 40% or less of them supports it, and unclear in between. A hypothesis is also unclear when the unclear cases outnumber the decided ones. A hypothesis that does not apply at the level of the ledger's cells is reported as not applicable rather than forced into a verdict; H6 is not counted in this ledger at all, whose verdict for it is that the question is read across the books in Volume 0.

Every verdict is read twice. The strict reading counts only what both coders found stated in a record: a value that at least one coder inferred does not enter the count. The lenient reading also counts what the setting implies. The two readings answer different questions. Strict asks what these records say outright; lenient asks what their settings suggest once a reader grants the inference. A verdict that appears only on the lenient reading rests on inference that a reader may or may not grant. Where the two readings give the same verdict from the same stated counts, the result rests on the records' own words; where they give the same label from different counts, as H1 does, the shared label is not by itself evidence that stated grounds carried it. A verdict that appears only on the strict reading would be strange, and no verdict in this ledger behaves that way. The verdicts are reported in Chapter 14, which takes each one apart: what it rests on, what it cannot carry, and why a verdict can be both computed and uninformative at once. This chapter only sets the rule, so that the later chapters can be checked against it.

2.5Frame and finding

A record is written by a studio, and a studio writes with habits. Some fields in a record carry wording that belongs to the studio rather than to the situation: the same phrasing, the same tags, the same shape of assumption-and-evidence. That wording is useful for reading one record and dangerous for reading many, because it makes every setting look alike in the same respect regardless of what the setting is.

The frame rule keeps that out of the evidence. If a field's wording is 30% or more shared across a studio's records, that field is not evidence across domains; it is treated as the studio's writing habit, and findings are not built on it. In WSTiger's records the code-level measure found that 2 of the 53 conditions for revisiting carry wording that is 30% or more the studio's frame. The measure was run on the invalidation wording; it does not settle how much of the other fields carries frame, and this study marks its own readings wherever a record's language looks like the studio's pack style rather than like the setting. One of the two cases coded below shows the pattern in the open, and this study reads it as such.

The second rule is that what the record form guarantees is not a finding. A form that asks for a field in every record will produce that field in every record, and counting the field cannot then say anything about the decisions the records describe. In WSTiger's records this rule bites in three places. Every one of the 53 records has a condition for revisiting the judgment, because the form asks for one; that the condition exists is a fact about the form. By contrast, none of the 53 records has a field for a deadline; that absence is a fact about the form and not evidence that decisions in this studio ignore deadlines. Missing material is different: the form has no dedicated field for it, yet a scenario, a question or an example judgment could still list what a decision lacks, and none of these 53 does, which the form alone does not guarantee. The same caution applies to any cell the form makes automatic. This study reads the high agreement on the reason-recorded cell as a statement about what the example-judgment line contained in these 53 records rather than as a fact about how decisions are made in general, and a reader comparing these counts with another studio's records should first check whether that form also asks for a judgment line. Where the ledger's counts are informative is in the cells the form does not dictate: which conditions a reader could observe, how much wording is frame, and what the two coders inferred about deciders, bearers, horizons and reversal.

The two rules together define what may be called a finding in this study. A finding is something the records show about a scene or about the set of records, read through the axes and reported with its counts. It is not the frame's habit of repeating itself across cases, and it is not the form's habit of producing a field. The chapters that follow name the differences between these three things each time one of them is used.

2.6Two cases coded

One person trades and explains

Synthetic scene. In a fictional trading office whose scenario line is marked 1995, one operator sits between two duties that the record leaves in the same hands. Orders go out from the same hands that work through reconciliation breaks: this study reads a break as a difference between what the account records and what the market shows, and clearing one is part of the operator's ordinary work rather than part of anyone else's. The office has a daily control deadline, and a break that remains unresolved at that deadline is what would bring the judgment back. No independent approval appears anywhere in the record's setting, and nothing in it says that anyone else reads the breaks. What is waited for before that decision is not stated in the record; its invalidation names a break unresolved at the daily control deadline as the trigger to come back, and what would be decided is whether the arrangement as written is kept, checked or stopped.

The three options are keeping the arrangement as it stands, requiring independent approval, or suspending new exposure. The record's own reading separates the two duties rather than choosing between speed and caution:

And the condition for coming back to that reading is:

Coded on the axes, this is how the ledger reads the record. A1, who decides: the index cell is team-or-organisation, on an implied basis. The record names an office and an operator but names no one as the decider; the coders inferred an organisational decider from the setting's language. The decider row shows both coders chose the same value in 33 of the 53 cells overall, so this is one of the cells on which inference carries the reading rather than the record's words. A2, grounds and what is missing: the reason-recorded cell agrees with the ledger's overall picture, where 49 of 53 cells (92%) are settled, 40 of them on a stated basis. The example judgment gives a principle rather than a calculation, and the record lists no material that is missing; the ledger's count for records that list what is missing is zero in every case. A3, deadline, horizon, waiting and the bearer: the record contains no deadline field, and the ledger's row for a deadline in a record is zero across the 53. The horizon cell is days-to-months, on an implied basis, which this study reads as drawn from the daily control deadline as a recurring marker rather than from any stated interval. The bearer cell is unsettled: the coders did not converge on who waits and pays, and the ledger's bearer row shows this is common, with 24 of 53 cells (45%) settled and 29 unsettled. A4, whether it can be undone: the index's undo cell is unknown for this record, and it does not split into a reversal constraint and a reversal action. The record names no cost of withdrawal for any of its three options, and the reversal rows explain why this is the weakest pair of cells in the ledger, with 42 of 53 constraint cells and 40 of 53 action cells unsettled across the whole set. A5, the condition: the invalidation cell is present, and the condition names a daily control deadline and a state of affairs, a break that remains unresolved. The condition names a deadline and a state of affairs a reader could watch for. This record's H2 cells go against on both readings, and since every record carries a condition, that places it among the 5 of 53 conditions the code-level count marks observable. A6, the two times: the setting is past, marked 1995 on the scenario line, and the example judgment is written before any outcome, as every record's judgment is. A7, divergence: the coders' values converge or not cell by cell as reported above, and since the record lists nothing that is missing, H3 is not applicable to it, as it is to all 53 under both readings. A8, the human's place: the place cell is human-approves, on an implied basis. Independent approval is one of the three options, so a human step appears, but the record specifies no role for that human, no condition and no interval. The human-role row shows that most cells across the 53 read "no" on that point, 38 implied and 1 stated. A9, the four higher goals: the index gives no case-level goal cell, so this study reads the option set in its own terms and reports the ledger's option counts beside it. Suspending new exposure cuts the immediate exposure; requiring independent approval slows the next act without touching what is already open; keeping the arrangement keeps speed and keeps the same person explaining what the same person did. The ledger's option counts are that both coders agree 44 of the 159 options across the 53 records reduce the immediate loss, and that what options give up is resilience for 18 options, long-term-stability for 17 options, ethical-coherence for 13 options and evolvability for 7 options.

The hypothesis cells for this record are worth setting out because they show how a case carries or refuses a hypothesis. On the strict reading the record goes against H2, is unclear for H1 and H5, and supports H8. On the lenient reading it goes against H2, is unclear for H1 and H5, and supports H7 and H8. H7 is instructive here: this study reads the lenient count as turning on the unnamed approving human, a step named without a role, which is exactly what that hypothesis describes. This study reads H2's going against on the strict reading as worth pausing over, because the record does write a condition; the hypothesis's second half, that conditions are written in words no one could observe, is what the strict count does not confirm from this record's own words.

Deposits versus performance

Synthetic scene. In a prop system whose scenario line is marked 2026, someone looks at an account and has to say, if its balance has risen, what the rise was. During operation, deposits and withdrawals occurred, and transfers between accounts were mixed into the balance changes; the same number on the screen therefore carries money that arrived, money that left, and money that moved between accounts inside the reporting scope. The evidence named in the record is exact: timestamps and amounts of deposits and withdrawals, net asset valuation at the same timestamps, and transfers between accounts within the reporting scope. What is waited for is the reconciliation, and what is being asked is how much of the rise may be recorded as investment income at all.

The three options are recording the balance increase as income, calculating performance with cash-flow separation, and presenting returns side by side by purpose. The record's example judgment separates external cash movements from investment gains or losses and asks that strategy comparison use performance that distinguishes the impact of cash flows alongside performance that reflects the timing of capital deployment, as seen in the actual investment experience. Its invalidation is written as a change condition with a next check: performance is confirmed when the reconciliation of trading gains or losses, costs and external cash flows matches, and the return corresponding to the purpose is displayed; the next check concerns how internal transfers are treated between account-level reporting and consolidated reporting.

Coded on the axes, the record reads differently from the first. A1, who decides: the index cell is unsettled. The record's assumption line describes flows rather than people, and no role appears that the coders could settle on as decider; the decider row shows 33 of 53 cells (62%) settled overall, so this is one of the cells where two readings did not converge. A2, grounds and what is missing: the example judgment gives a principle, and the scenario line lists the evidence a reconciliation would have to produce. This study reads the required-evidence line of the scenario as the record's own partial listing of what its judgment rests on; the record does not list what is missing from those grounds beyond the evidence named, and the ledger's count for records that list what is missing is zero in every case. A3, deadline, horizon, waiting and the bearer: the horizon cell is unknown and the bearer cell is unknown as well. The record names no interval before an answer can come and puts no party close to the cost; the horizon row across the 53 shows 23 cells unsettled and 18 unknown, so unknown is a reading the coders reached often enough to be ordinary in this ledger. A4, whether it can be undone: both reversal cells are unsettled. The record's question is about measurement rather than about a commitment, and the coders found nothing on which to hang a value. A5, the condition: the invalidation cell is present, and the condition is a change condition plus a next check. This study reads the wording as naming evidence a reconciliation would have to produce, which looks closer to observable than the shorter clause of the first case. The ledger classifies it otherwise: this record's H2 cells support on both readings, and since every record carries a condition, that places the condition among those the code-level count does not mark observable. The counts follow the ledger's classification; this paragraph's reading of the wording departs from it and is this study's own. A6, the two times: the setting is present, marked 2026 on the scenario line, and the judgment precedes any outcome. A7, divergence: the record lists nothing that is missing, so H3 is not applicable to it as it is to all 53. A8, the human's place: the place cell is unsettled. The record names no human and no rule about who records performance; the human-role row shows that most cells across the 53 read "no" on specification of a human role, 38 implied and 1 stated. A9, the four higher goals: the index gives no case-level goal cell. The option set is about how performance is measured rather than about cutting an exposure, and this study reads the case as one in which no option visibly cuts an immediate loss in the sense the hypothesis describes; the ledger's not-applicable count for H8 is 16 of 53 under both readings, and this record's case-level cell for H8 is not-applicable on both readings.

The hypothesis cells for this record are H2 supported under both readings, H5 against under both readings, H1 unclear under both readings, H6 by-book, and H8 not-applicable under both readings. H2 supported is worth noting beside the first case, where the count went against on both readings: every record writes a condition because the form asks for one, so what separates the two records is not the length of the wording but whether the condition names a trigger a reader could observe. H5 is against under both readings, and Chapter 14 is where the verdict on H5 is taken apart.

2.7Limits of the method

Three limits should be held beside every count in this study.

The first is that the read cells were read by AI. Each was coded by two models of different families, each giving a value, a basis and a reason, and agreement was reported rather than assumed. Agreement is high on what a record says on its face and low on what it implies about burden and reversal: the bearer cell was settled in 24 of 53 cells (45%), the reversal-constraint cell in 11 of 53 (21%) and the reversal-action cell in 13 of 53 (25%). Those three low agreements reach the hypotheses. H1, which asks whether the bearer is someone other than the decider, is unclear on both readings, and its lenient counts (supports 5, against 12, unclear 36) show that most cases could not be decided either way. H4, which links difficulty of undoing to the distance between decider and bearer, is not applicable on the strict reading and against on the lenient reading, with only a handful of cases decided in either direction. A verdict computed on cells that most coders could not settle carries less than a verdict computed on cells they could.

The second limit is that the records are synthetic. No record has an outcome, so no chapter can say what any option produced, and the counts describe how decisions are written, not how markets behave. Every count in this study describes these 53 records and no wider population.

The third limit is this study's own authorship. The chapters are written by AI from the records, the readings of the works are this study's readings and stay within the reading notes that describe each work, and the quotes from records are exact copies of what the records say. The hypotheses and their verdicts come from the ledger, not from this chapter's authorship; this chapter states the rule by which the ledger calls them and reports counts only where they bear on the method's limits.

All of this is reported because a reader should be able to check every count against the published records and the index. Where a cell could not be settled, the ledger says so instead of smoothing it; where a field is guaranteed by the record form, the ledger's rules keep it out of the findings; and where this study has read something into a record, it is marked as this study's reading rather than presented as the record's words.

Hands on

The chapters that follow read the records against the ideas that have been used to think about risk before them. Chapter 3 begins with Frank Knight's two uncertainties, set out in Chapters VII and VIII of Risk, Uncertainty and Profit as its reading note gives them, and with the debate among later readers about what his distinction turned on; four records sit at the edge of the measurable beside that argument, and this chapter's axes are the ones they will be read on.