Chapter 14. The Hypotheses Tested

14.1How the verdicts were reached

Every book in this series writes its findings into a ledger another book can read; this study's ledger holds the six axes of Chapter 2, coded for 53 records by two models of different families, each giving a value, a basis — stated, implied or none — and a reason for it. The eight hypotheses were fixed before any chapter was drafted, and the rule for calling a verdict was fixed with them: a hypothesis is called supported where 60% or more of the cases it applies to supports it, called against where 40% or fewer of those cases support it, and called unclear where the share lies between those two marks or where the unclear cases outnumber the decided ones; where a hypothesis applies to no case in the corpus, the verdict is not applicable. Every verdict is then computed twice. The strict reading counts only what both coders found stated in a record's own words. The lenient reading also counts what a setting implies — the kind of evidence that comes from reading a scene and asking who would be standing where. A verdict is therefore a description of how 53 synthetic records were coded, not a statement about the worth of any option in them; this chapter reports what the records support, what they go against and what they cannot reach. Where a hypothesis's test depends on a field the record form does not have, or on a cell where the coders rarely agreed, that is reported as part of the result and is not repaired by inference. Chapter 2 sets out the axes and cells, the coders and their per-cell agreement, the strict and lenient rule in full, the frame rule under which a condition whose wording is shared across a studio's records is not counted as cross-domain evidence, and the second rule that what the record form guarantees is not a finding — every record carries a condition for revisiting because the form asks for one. Two of the counts used below come from code rather than from coders: that all 53 records carry such a condition, and that 5 of those conditions name something a reader could observe. The rest are verdicts and descriptive counts: the verdicts rest on the coder agreements the ledger reports, and a few also rest on properties the code counts, such as whether a condition can be observed.

14.2What the records support

Two hypotheses are called supported on both readings, and in both cases the strict and lenient counts are identical, which means the verdicts do not depend on inference from the settings.

H2. The hypothesis holds that few decisions write down in advance what would reverse them, and that those which do write it in words no one could observe; the coders' question was whether there is a condition and can it be observed. Strict and lenient both give the same counts: 48 cases support, 5 go against, none unclear, none not applicable, a share of 91%. The support rests on what the records' conditions say, not on what they are. All 53 records carry a condition for revisiting, but only 5 conditions — 9% — are stated with a number, a threshold or a time, which is what the ledger's code counts as observable (an event or a role it does not count); 2 conditions word it in phrases the studio's other records share to the measured frame threshold of 30% or more, which Chapter 2's rule excludes as evidence across domains. Had the hypothesis been scored on presence alone, the verdict would describe the form, since every record has an Invalidation line because the form asks for one. What makes the support more than a restatement of the form is the observability row, where the form is silent and the reading comes from the conditions themselves. Under the strict rule, the counts take only what both coders found stated in a record's own words; the 5 cases going against are those whose conditions name something checkable. Chapter 13 quotes both groups, and it is there that the split can be read case by case: what a reader would have to watch for, and what each condition leaves to judgment. This study reads H2's support as a statement about how these records are written at the moment of decision — a studio treating the revisiting of a risk decision as something that must be named, and naming it in words that a reader could usually check against nothing. The records are not organisational files kept over time; what they show is the writing, not the memory.

H8. The hypothesis holds that the options which cut the immediate loss most often give up evolvability and resilience among the four higher goals. Strict and lenient again agree: 24 cases support, 13 go against, none unclear, 16 not applicable, a share of 65%. Under the verdict rule the share is taken over the cases to which the hypothesis applies, so the 16 cases the ledger marks not applicable drop out; the aggregate numbers give the count but not the reason for each exclusion. The option-level figures give the shape of the evidence: the ledger holds 159 options across the 53 records; both coders agree that 44 of those options reduce the immediate loss; and what those options give up, as both coders read it, is resilience 18 times, long-term-stability 17 times, ethical-coherence 13 times and evolvability 7 times. The hypothesis names evolvability and resilience, and the tallies put resilience first and evolvability last of the four. This study reads the per-case verdict and the option-level tallies as two counts answering different questions — the first counts records, the second counts options across the corpus — and it does not reconcile them, because the ledger sets out no rule for turning option-level tallies into a per-case test. What the support rests on in the end is a pattern the coders could see in most of the records where the question applies: a restriction, a re-sizing or a halt that cuts an immediate loss and, in the same record's option list, gives up one of the goals that keeps a system able to change. Thirteen records went the other way, and each is marked against H8 in the case ledger, though the aggregate numbers do not name them. One of those records asks, in its own terms, how a widened stop should change the size of a trade:

The ledger records what the option list gives up at the point of decision; this study reads that moment as touching both prospect theory, read in Chapter 7, on outcomes judged as gains and losses from a reference point, and tail risk, read in Chapter 10, on rare heavy losses; neither work describes an option list giving up a goal, so the connection to H8 is this study's own reading. As evidence it bears on how these records are written, with the trade and its loss named in the same breath, rather than on behaviour in markets.

14.3What the records go against

One hypothesis is called against on both readings, and it is the only one in this study called against on both.

H5. The hypothesis holds that a restriction outlives the memory of its reason; the coders' question was whether the reason for a restriction is recorded. The strict counts are 0 supporting, 40 going against, 13 unclear, none not applicable, a share of 0%. The lenient counts are 2 supporting, 45 going against, 4 unclear, 2 not applicable, a share of 4%. The verdict is against in both readings, and the lenient reading moves only a little: the count going against rises from 40 to 45, two cases drop out as not applicable, and the direction is unchanged. The ledger's row on recorded reasons gives the same picture from the other side: 40 cases both coders found a reason stated, 5 more they found implied, 2 no reason, 2 not applicable, 4 unsettled — 49 of the 53 on one reading or another.

Why recorded reasons cut against the hypothesis in this domain comes down to what these records are. They are not files that an organisation has kept and an outsider has audited. They are settings written at the moment a question is asked, by a studio whose form asks for a question, a set of options and an example judgment. A reason travels with a restriction in these records because the form asks for an example judgment, and the coders found a reason recorded in 49 of the 53 cases on one reading or another. Where a record is silent about the reason for a limit, the form gives the reason no separate field, and this corpus cannot show what caused the silence: a loss limit that refills at midnight and a leverage limit that was raised each time it was exceeded both carry their reason in the option that questions them. This study therefore reads the verdict against as a finding about the corpus and only about the corpus: in a record written at the moment of decision, memory has not yet had time to fail, so the score on H5 does not settle the question the hypothesis asks. The hypothesis is about what happens to a restriction after the record has been written — through days, retrainings, personnel changes — and the corpus contains no later record of the same restriction to compare against. The Korean-language risk series places settings in the past, in the present and in the future, but each version is written as its own setting, so the versions do not show whether a reason survives.

The organisational accounts would ask for evidence this corpus does not hold. Vaughan's normalisation of deviance in The Challenger Launch Decision (1996), read in Chapter 5, describes an organisation coming to accept as normal a departure it once treated as a warning — this study reads that process as a reason failing to travel, though an organisation can keep its reasons and the memory of its warnings and still change how it reads a departure, and a single-moment record cannot show the link either, because the departure and the memory of the warning would have to be visible together; a synthetic setting can describe departures repeated over time, but it holds no longitudinal evidence from a real organisation. Chapter 15 sets the ledger's rows beside The Risk Management of Everything (2004), where Power describes documentation written to defend an organisation displacing judgment, and this study alone relates that displacement of judgment to the silence the ledger shows about a restriction's reason; it is a question for that chapter to answer, not this one. The verdict stands as reported: in most of these 53 records the coders found a reason recorded, and H5's claim that a restriction outlives the memory of its reason is not tested by a corpus written at the moment of decision.

14.4What the records cannot judge

Five hypotheses are not answered by this corpus. Three are unclear or not applicable on fields the record form does not carry, one is left to other books by its pre-registration, and one succeeds only where the evidence is thin.

H1. The hypothesis holds that the cost of waiting falls not on whoever decides but on those nearest the point where things stop; the coders' question was whether the bearer is someone other than the decider. The strict counts are 0 supporting, 0 going against, 53 unclear — the whole corpus undecided. The lenient counts are 5 supporting, 12 going against, 36 unclear, none not applicable, a share of 29%; the verdict is unclear, because under the rule unclear also holds where the unclear cases outnumber the decided ones, and the 36 unclear cases outnumber those decided in either direction. The bearer row explains it: 29 cases unsettled, 12 the decider implied, 7 unknown, 4 those nearest the stoppage implied, 1 third parties implied. Coder agreement on the bearer is 24 of 53, or 45%. The records carry no field naming who waits, and the decider row was agreed in 33 of 53 cases at 62%, its values almost all implied, with no record holding a decider field at all — so the comparison H1 asks for needs two cells, and both are thin. This study reads H1's unreadability as a property of the form and the coding together, not as evidence that the waiting falls on the decider: in 36 cases the coders simply could not tell. Chapter 10 reads the bearer row against the records' own settings, which Chapter 12 follows across past, present and future; Chapter 10 is where the hypothesis can be discussed at all, and where the chapter explains why agreement on the bearer was so low.

H3. The hypothesis holds that where judgments diverge, material is missing. Both readings return not applicable on all 53 cases. Two reasons, both structural. The aggregate table supplied with the ledger shows no divergence cell among the cases the two coders read: the records present one question, three options and a single example judgment, so there is no second judgment for the first to diverge from. And the missing-material cell is empty too — 0 cases list what is missing, with 0 items listed in all. A hypothesis that needs a divergence and a list of missing material can be scored on neither, and the coders could not be blamed for the gap. This study reads the not-applicable as a property of the corpus: these records are written to pose a decision, not to record an argument between judges. To test H3 would need a different form — records where two people write two judgments of the same question and where the record lists what it does not know. The nearest thing in this study is the coders' own disagreement, but those are disagreements about how to read a record, not judgments about a decision, and the ledger keeps them in the agreement rows rather than treating them as evidence about missing material.

H4. The hypothesis holds that the harder a decision is to undo, the further apart are those who decide and those who bear the cost. The strict reading returns not applicable on all 53. The lenient reading returns 1 supporting, 3 going against, 3 unclear, 46 not applicable, a share of 25%, and a verdict of against — the share counts only the cases decided in either direction, not the unclear ones, and it says very little in either direction. The reversal rows explain both readings. The constraint row was agreed in 11 of 53 cases, or 21%: easy implied in 9, costly in 1, unknown in 1, unsettled in 42. The action row was agreed in 13 of 53 cases, or 25%: costly implied in 6, unknown in 5, impossible in 1, easy in 1, unsettled in 40. No record holds a reversal field; the coders read reversibility from the options, and in 42 cases they could not agree on whether the constraint was easy or costly at all. Under the strict rule, nothing was stated, so the hypothesis could not be scored; under the lenient rule it became scoreable in a handful of cases, and the results split. This study reads H4 as the hypothesis this corpus is least able to carry, because the axis it tests is the one where the coders agreed least — 21% and 25%. Chapter 11 sets out the two kinds of undoing the axis separates, undoing a restriction against undoing what the restriction lets happen or stops, and reads the value of keeping the option to preserve, later called quasi-option value, in Arrow and Fisher (1974) and Dixit and Pindyck's option value of waiting beside the records; that is where a reader can see what a verdict on H4 would need, and why the corpus does not supply it.

H6. The hypothesis holds that the longer the horizon, the less is decided now and the more conditions are set for revisiting. It was pre-registered to be read across the books in Volume 0, so it carries no verdict here, and the ledger gives what this study can offer toward it. The horizon row was agreed in 30 of 53 cases, or 57%: days-to-months implied in 11, years implied in 1, unknown in 18, unsettled in 23. On the lenient reading, the by-horizon table holds 11 cases at days-to-months and 1 at years, with 0 at decades and 0 at centuries or more; strict, every horizon band holds 0 cases. The conditions column in that same table reads 1 for days-to-months and 1 for years, which is what one would expect, because every record carries exactly one condition — the form's Invalidation line. This study reads that constancy as a limit on what this corpus can show, and H6 is left to Volume 0 rather than scored here: the hypothesis asks whether conditions are set differently as the horizon lengthens, and a corpus in which conditions are set by the form, in every case, at one per record, cannot show that. Volume 0 will let other books of the series, in other domains and other settings, be coded on the same axes and by the same rule, so that H6 can be judged on evidence that varies across cases instead of evidence that is constant.

H7. The hypothesis holds that the more a decision is left to an AI or an automatic rule, the more the human's place — approval, exception or review — is empty or a name only. The strict reading returns not applicable on all 53. The lenient reading returns 5 supporting, none going against, none unclear, 48 not applicable, a share of 100%. The strict reading fails because no applicable case supplies the combination of stated evidence it needs: the place row reads human-decides implied 25, human-approves implied 3, ai-or-rule-decides implied 2, unknown 3, unsettled 20, with agreement on place at 33 of 53, or 62%; the human-role row reads no (implied) 38, no (stated) 1, yes (implied) 1, yes (stated) 1, not-applicable (implied) 1, unsettled 11, with agreement at 42 of 53, or 79%. Under the strict rule the hypothesis needs the automated place and the empty human role both found stated; in no case was that so. Under the lenient rule it became scoreable in 5 cases, and in all 5 the setting implied an automated or approving step with no named human role. This study reports the 100% exactly as the ledger gives it and does not extend it: five cases out of 53 is a small base, and the other 48 are counted not applicable, which is a limit on where the hypothesis reaches rather than a sign that the coders could not agree. Chapter 8 reads the human's place axis in full, including what it means that the records name a human step — approve, release, review — without a role, a condition or a time by which it must happen.

Three of the five hypotheses fail on fields the form does not carry or codings the coders could not settle; H6 fails on evidence that is constant by design; H7 fails strictly on evidence that is never stated. This study reads the pattern as one about the record form rather than about the hypotheses themselves: those needing two cells, or a cell the corpus lacks, are the ones this reading cannot answer. A different corpus — with a decider field, a deadline field, a reversal field and two judgments per question — could score them differently, and the ledger was built so that it could.

14.5What the verdicts are worth

The verdicts describe synthetic settings. No record reports a real firm, a real trade or a real market event, and no record has an outcome, because the whole point of the form is to hold the decision before anyone knows how it ends. A verdict is therefore a statement about a corpus of writings, not about risk in markets, and nothing here can be checked against what actually happened, because nothing in the corpus happened. The counts also describe a purposive selection: the 53 records were chosen by topic, as the Introduction sets out, from WSTiger's public atlas of records holding a scenario and options; they were not sampled, so "most records" throughout this study means most of these 53 and not a wider population.

The coders were two models of different families, each giving a value, a basis and a reason; the study reports their disagreement rather than resolving it. Agreement varied sharply from cell to cell — the reason-recorded cell was agreed in 49 of 53 cases, or 92%. Where the reading needed inference the agreement thinned: 45% on the bearer, 21% on the reversal constraint, 25% on the reversal action. The strict and the lenient readings exist precisely so that a reader can see how much of each verdict rests on inference: H2 and H8 stand on both, H7 changes from not applicable to supported with leniency, and H1 stays unclear on either. This study is itself written by AI, its readings of the works are interpretations marked as such, and every block quote is a record's own sentence copied exactly, so that a reader can check the evidence behind any count without trusting the analysis around it.

One studio's frame limits the verdicts in a second way. Chapter 2 measures how much of a field's wording is shared across a studio's records, and where a field's wording is 30% or more shared, it is not counted as evidence across domains; two of the 53 conditions fell into that class. What remains of H2 is a statement about what a record form chooses to write when it asks when a judgment should be revisited — and a form is a studio's decision, which makes this verdict as much about WSTiger's publishing practice as about decision records in general. That is a limit and also the subject: writing a decision down before its outcome is itself a decision about risk.

What Volume 0 will do with this ledger is read it again. The six axes, the cells, the strict and lenient rule and the verdict rule are the same across the series, so that a book in another domain — safety organisations, public services, or a different market setting — can put its counts beside these and a reader can see where two corpora agree about how decisions are written down. H6 was pre-registered to be read across the books in Volume 0; this study's corpus cannot carry the evidence it needs, and the cells for H1 and H4 were thin here as well. The comparison Volume 0 makes is between corpora, not between markets, and since the hypotheses are about records, a verdict in this study does not transfer to a real trading desk, an insurer or an agent system. What the verdicts are worth in the end falls into three parts: they are counts any reader can check against the ledger's numbers; they are a diagnosis of what the record form leaves out — no decider field, no deadline field, one judgment per setting, conditions set by the form in every case; and they are a base for the next coding, strict first and lenient second, so that the same eight hypotheses can be tested again on records that carry more of what they ask for.

Chapter 15 sets these findings beside the theories the earlier chapters read: Knight's distinction between measurable and unmeasurable uncertainty, the accident and reliability literature, the option value of waiting and the corrigibility of a decision, Luhmann's risk and danger, and Power's risk management — saying in each case where the ledger bears on the question and where it cannot.