Chapter 15. The Hypotheses Tested

15.1How the verdicts were reached

Chapter 2 sets out the instrument that produces every number used in this chapter: the six axes on which each question is coded, the extra cells for where the judgments part, for the human's place and for the four higher goals, the two coders of different families, and the strict and lenient readings that run through the whole ledger. It also sets out the rule that converts the counts into a verdict. The rule is arithmetic and was fixed before the counts were read. A hypothesis is supported when 60 per cent or more of the decided cases it applies to support it, goes against when 40 per cent or less of those cases support it, and is unclear when the support falls between those marks; a hypothesis is also unclear when the cases whose values are unclear outnumber the decided ones. The share is taken over the decided cases alone. Cases in which the hypothesis does not reach the setting are counted as not applicable, and cases whose values the coders could not settle are left unclear; both kinds drop out of the denominator. That is why H1's strict share can read 100 per cent with 25 cases supporting and 25 unclear: the unclear 25 carry no part of it. The strict reading counts only what both coders found stated in the record's own words; the lenient reading also counts what the setting implies. Every verdict below is the ledger's, and this study keeps each verdict exactly as the ledger gives it.

Each hypothesis carries with it a test, written as a question over the coded cells: H1, whether the bearer of the cost is someone other than the decider; H2, whether there is a reopening condition and whether it could be observed; H3, whether the cases where the judgments diverge list missing material; H4, whether the further a decision is from being undone, the further apart sit the decider and the bearer; H5, whether the reason for a restriction is recorded; H6, whether the longer the horizon the less is decided now; H7, whether decisions left to an automatic rule leave the human's place empty; H8, whether the options that cut the immediate loss most often give up evolvability and resilience.

Pre-registration matters most on H2 and H5, whose verdicts run against the hypotheses. A rule written after the counts would have been a rule that could be moved; the verdicts are worth keeping precisely because they could not be.

15.2What the records support

Three hypotheses receive a verdict of supports in both readings: H1, H3 and H4.

H1, on who waits. The hypothesis reads that the cost of waiting falls not on whoever decides but on those nearest the point where things stop, and the test asks whether the bearer is someone other than the decider. Strictly, 25 of the 50 cases support it, none go against it, 25 are unclear and none is not applicable; the ledger's share is 100 per cent. Leniently, 42 support, none go against, 8 are unclear, none not applicable; again the share is 100 per cent. Two coded cells carry the verdict: the decider, where the coders agreed in 28 of the 50 cases and left 22 unsettled, and the bearer, where they agreed in 42 of 50 and left 8 unsettled. Chapter 2 sets out the cells and the agreement, Chapter 9 reads the decider, and Chapter 10 the bearer. What the two cells together show in these records is a distance: the decision actor is a committee, council or authority named in the record, while the cost sits with the people at the point where services stop. The strict count is smaller than the lenient one because the strict reading counts only values both coders found stated and the lenient reading also admits what the setting implies; cases with unsettled values stay unclear in both readings. Which cells supply the difference would take a case-level account that this chapter does not have. The lenient reading finds the pattern in 42 cases. No case in this collection was coded against the hypothesis; the 25 strict cases and 8 lenient ones left unclear show neither direction.

H3, on where the judgments part. The hypothesis reads that where judgments diverge, material is missing. Thirteen of the 50 cases had judgments choosing more than one option, and in all thirteen both coders found the missing material stated; none goes against it, and the other 37 cases are not applicable because the judgments did not part. Strict and lenient readings give the same counts: supports 13, against 0, not applicable 37, share 100 per cent. The A2 cell sits behind this: cases that list what is missing, 50 of 50; items listed in all, 148. Two things should travel with the verdict. First, every case in the collection lists what is missing, because the question form asks for it, so the hypothesis is supported in the only cases where it can be tested and the test cannot separate records that all carry the same field. Second, the divergence cell itself is a record fact, not a coder inference: 13 of 50 cases, with 13 judgments per case as the ledger counts them. Which cases divided and what each dissent turns on is Chapter 14's work; here the count stands on its own.

H4, on what cannot be undone. The hypothesis reads that the harder a decision is to undo, the further apart sit those who decide and those who bear the cost. Strictly, 3 cases support it, none go against it and 47 are not applicable — the smallness is a fact about the cells rather than about the settings. The two reversal cells show how thinly this is stated: the coders agreed in 40 of the 50 cases on the reversal constraint and in 37 of 50 on the reversal action, and almost every value they agreed on is implied rather than stated; Chapter 11 reads those cells in full. That thinness is why the strict reading leaves three cases to test on, and all three support the hypothesis. The lenient reading, which admits what the setting implies, returns supports in 30 cases, unclear in 7, not applicable in 13, with no case going against it; the ledger's share is 100 per cent in both readings. What these counts show is narrower than the hypothesis's comparative claim: among the cases where both cells can be read, none is coded against H4, while the ledger holds no comparison across degrees of difficulty in being undone and degrees of distance between decider and bearer. The verdict stays as the ledger gives it. Where corrigibility and what can be undone are read in full, Chapter 11 holds them.

15.3What the records go against

Two hypotheses go against, in every case, in both readings: H2 and H5. These are the only verdicts in the ledger that run against the hypothesis everywhere, and they deserve to be read for what they are against rather than simply recorded.

H2, on reopening conditions. The hypothesis reads that few decisions write down in advance what would reverse them, and that those that do write it in words no one could observe. The test asks whether there is a condition and whether it could be observed. Strictly and leniently alike, 0 cases support, 50 go against, none is unclear and none is not applicable; the ledger's share is 0 per cent. The counts that produce the verdict: cases with a condition for revisiting, 50 of 50; conditions in all, 798, of which 650 sit in the field judgments[].claims.stopConditions (602 observable) and 148 in the field decisionBrief.reopenWhen (27 observable); conditions a reader could observe, 629 of 798, that is 79 per cent; conditions whose wording is 30 per cent or more the studio's frame, 0 of 798. Both clauses of the hypothesis fail at once. Every one of the 50 questions writes down a condition for reopening the decision, and most of the conditions name something a reader could watch — a number, a date, an event or a role. A reopening condition sends the decision back for review rather than promising to reverse it: the review may end in keeping the decision, changing it or reversing it, and the record does not say which. Part of this is the question form, which asks for reopening lines, and Chapter 2 sets out what the form guarantees and what is left as finding. What survives that critique is the observability count: 629 of 798 conditions carry observable wording, and no condition's wording leans 30 per cent or more on the studio's frame. The two fields also differ from each other — the 650 stop conditions carry 602 observable instances, while the 148 reopening conditions carry 27 — so a reader taking only the second field would see a much weaker record than the full 798 describes.

H5, on the reason for a restriction. The hypothesis reads that a restriction outlives the memory of its reason. The test asks whether the reason for a restriction is recorded. Strictly and leniently alike, 0 cases support, 50 go against it, share 0 per cent. The cell behind it: reason recorded, 50 of 50 cases, yes (stated). The against verdict here is a statement about the shape of a Meridian record rather than about time passing. A question of this form asks for the reasons attached to a restriction at the moment the restriction is posed, so the field records the reason the setting gives at that moment. Nothing in the form stops a setting from carrying an older restriction whose original reason has been forgotten, and the collection cannot show whether any of them does. None of the 50 cases follows a restriction far enough to test whether the memory of its reason survives contact with the next decision. The verdict stands as the ledger gives it, and this study keeps it; the reader should see that what runs against H5 is a collection built to carry reasons, not evidence that restrictions in the world keep theirs. Chapter 13 reads the same records for the observable wording of their reopening conditions, and H2's against verdict sits beside that reading rather than in conflict with it.

15.4What the records cannot judge

Three hypotheses cannot be settled inside this study: H6, H7 and H8. Each fails for a different reason, and each is worth naming separately.

H6, on horizon and revisiting. The hypothesis reads that the longer the horizon, the less is decided now and the more conditions are set for revisiting. The ledger gives no in-book verdict for it; its verdict, in both readings, is the ledger's own words — read across the books in Volume 0. Three things keep the hypothesis from being judged here. First, the horizon cell is unsettled in 10 of the 50 cases and the ledger records the rest as a mix of stated and implied values; Chapter 2 sets out the cell and Chapter 12 reads the horizon distribution. At the strict level only stated horizons count. Second, the question form asks every case for reopening lines, which explains that every case carries some but sets no limit on how many, and the total of 798 conditions joins the questions' reopening lines to the judgments' stop conditions, two fields of different kinds that a count per case does not separate. Third, the hypothesis turns on the decision horizon, and this collection holds horizons from days to months to centuries, so one collection read once is not barred from bearing on it; the ledger's own words defer the verdict, and these records give no further reason for the deferral. The four layers of time are setting dates rather than decision horizons; Chapter 12 reads them, and the cross-collection reading belongs to Volume 0. The time spread of the collection is itself worth carrying: 12 cases set in 2026, 4 in 2031, 3 each in 2033, 2041, 2044 and 2047, 2 each in 2030, 2032, 2034 and 2192, and one case each in 2035, 2042, 2043, 2048, 2112, 2114, 2118, 2126, 2148, 2164, 2168, 2172, 2186 and 2194.

H7, on the human's place. The hypothesis reads that 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. The verdict in both readings is not applicable: 0 supports, 0 against, 0 unclear, 50 not applicable, share null. The reason is narrower: the hypothesis needs cases in which the coded decision has been left to AI or to an automatic rule, and the cells coded for these 50 focal decisions record none, so the test does not reach the settings. The cells that were coded are these: the human's role is specified in 50 of 50 cases, yes (stated); on the place cell the coders agreed in 46 of 50 cases and coded human decides (stated) in 33, human decides (implied) in 13, with 4 unsettled. Every record here carries a decision actor with a mandate and asks what that actor will decide. A specified human role does not settle the distinction H7 draws between a role that decides and a role that is a name only; the codes do not say which of the two these settings hold. The not-applicable verdict is therefore a fact about the collection rather than a finding about machines: it says nothing either way about autonomous weapons, automatic rules or what a human role looks like when a setting leaves the decision elsewhere. Where the human's place and the decision actor are read in full, Chapter 9 holds them on the axes for the decider and for the human's place.

H8, on what options give up. The hypothesis reads that the options that cut the immediate loss most often give up evolvability and resilience among the four higher goals. The verdict, strict and lenient alike, is unclear: supports 20, against 28, not applicable 2, unclear 0, share 42 per cent. The cells behind it: 150 options in all; both coders agree that 82 reduce the immediate loss; and what those options give up, as both coders read it, is long-term stability 23, ethical coherence 22, evolvability 17, resilience 12. The verdict is unclear for three reasons. The support falls between the two marks of the rule — 28 cases go against and 20 support, with 2 not applicable — so the arithmetic returns unclear rather than either direction. What an option gives up is a reading of a trade-off the record states, not a measurement, so both ends of the comparison rest on coder judgment about stated wording. The counts by goal do not run in the direction the hypothesis names: stability and coherence are given up more often than evolvability and resilience. The strict and lenient readings report the same totals, though equal totals alone do not show that the same cases stand behind them; the rule turns on the support share, which is 42 per cent in both. This study keeps the unclear verdict and leaves both directions open; Chapter 2 sets out the four higher goals and the verdict rule.

15.5What the verdicts are worth

Every number in this chapter passes through four artificial stages, and each stage sets a ceiling on what a verdict can carry. The questions are synthetic settings in fictional polities, written to hold a decision before its answer rather than to report an event. The thirteen judgments published with each question are AI-generated, dated and revisable, which explains part of why they seldom part. The two coders are AI models of different families, agreeing per cell from 28 of 50 cases on the decider to 46 of 50 on the human's place and 50 of 50 on the human's role and on the recorded reason. And this study is itself written by AI from Meridian's records. A verdict produced by that chain describes how these records are built and judged; it does not describe wars, states, forces or people. The scope is fixed by the collection itself: the 50 cases in this study's ledger, every one of them synthetic, with no selection inside them. The counts describe that collection on that date, and nothing larger.

Within that scope the verdicts are worth keeping, and they are worth keeping in the form the ledger gives them. H1 supports in 25 strict cases and 42 lenient ones; H3 supports in all 13 cases where it applies; H4 supports in 3 strict cases and 30 lenient ones, with no case against it; H2 and H5 go against in all 50 cases in both readings; H7 is not applicable in all 50; H8 is unclear with 20 supports and 28 against. One count in this chapter survives the form critique and is checkable by any reader: 629 of 798 conditions carry observable wording, and none of the 798 leans 30 per cent or more on the studio's frame. That statement covers the whole collection and cannot be tested by opening a single record: the totals rest on all 798 conditions and on the coding criteria, and observable wording does not mean a reader can obtain the evidence needed to tell whether a condition has occurred. The two against verdicts are worth as much as the three supports: they say that this form guarantees the thing H2 denies, and that it cannot show the thing H5 describes.

What Volume 0 will do with this ledger is written in the ledger's own verdict for H6 — read across the books in Volume 0. The series spine treats the hypotheses as questions for more than one collection, and H6's shape needs exactly that: horizons compared across books rather than inside one. Volume 0 will set this study's ledger beside the ledgers of the other books in the series, each a different domain's collection of questions coded on the same axes, so that a hypothesis which no single book can settle can be read across the collections there, with no promise that the reading will settle it. Until then, this study keeps its verdicts where they stand, unchanged and in both readings, and does not promote one hypothesis over another on the strength of these 50 records.

15.6Hands on: the findings beside the theories

Chapter 16 sets the findings above beside the theories read earlier in this study: ending wars and keeping ceasefires, warning and belief, civilian control and machines, generations and justice, with Kaldor's new wars and Walzer's two sets of rules as that chapter's own comparisons. The verdicts travel with the findings; the theories do not change them, and the counts do not confirm or refute a theory. What the comparison can do is show where a coded count and a written argument talk about the same thing and where they talk past it — and that is the next chapter's work.