The rubric behind the numbers

The flagship map on this site extracts the argument of the book If Anyone Builds It, Everyone Dies (Yudkowsky and Soares, 2025). Every claim on that map carries a number, and the book states almost none. So where do the numbers come from?

They come from the authors' own confidence language, read through a fixed table: "repeated categorical assertion" is 0.93, a flat unhedged assertion is 0.90, "by default" and "our best guess" are 0.85, "we expect" is 0.75, "could well" is 0.70. The table was written down before any number was assigned, and the pass that first applied it, on 2026-07-19, moved no number after seeing the solve. Later passes have changed the map and numbers on it, each recorded with its date: between 2026-09-25 and 2026-10-13 seven of them moved the title claim, five down and two up (section 4 lists them). This page is that table, the protocol that applied it, the record of what moved afterwards, and the honest account of what it can and cannot support.

Why it matters: the map's overlay marks places where the solved value and the authors' own stated confidence disagree. If the mapmaker had picked the numbers, such a mark would only say "the mapmaker disagrees with the book". Because the numbers are the authors' words run through a table fixed in advance, apart from the exceptions this page names (the objection defaults of section 2.4, the 63 flat 0.90 links of section 2.5 that carry their number by class alone, and five steps that hold by the meaning of their terms, section 4), the mark says something stronger: the book's own confidences do not add up to the book's own conclusion there. That claim is worth only as much as the pre-registration behind it, which is why the pre-registration is on this page.

For what happens to the numbers after they are assigned (the probability model, the solve, what has been proved about it), see the mathematics behind ArgMap. For writing your own maps, see the authoring tutorial.

1. The protocol

The map-wide pass ran on 2026-07-19 as a pre-registered experiment. Its rules, as recorded that day:

1. The rubric first, blind. Each extension to the table (section 2) was derived by an independent agent from the relevant source pages, with no access to any solve.
2. Five independent elicitors, one per region. Each was given only the fixed rubric and their region's text with every number stripped to ?. No elicitor saw another's region.
3. Every number registered before the region was solved once. One honest solve per region, after the numbers were fixed.
4. Nothing moved afterwards. Where the solve disagreed with the authors' stated confidence, the disagreement was recorded as a finding, not repaired by editing a number.
5. Provenance on every value. A rubric-derived number carries the ? marker in the map source. A bare number is reserved for figures the authors state as figures, and outside quoted insider odds the book states none.
6. The driving quote travels with the number. Each re-elicited node carries the sentence it was read from, either as an attributed quote line under the node or in the trailing comment that names its rubric class.

Rule 4 is the one that does the work. A rubric with a repair step is a rubric that fits the conclusion; this one has none, and the map still carries the gaps that produced. The rule bound the 2026-07-19 pass. The changes since are dated in section 4, each with the re-read, ruling or structural repair behind it, and three of them brought the title claim and its stated range closer together: the re-read of its check on 2026-09-25, the ruling of 2026-10-06 that raised the last step of its chain, and the link of 2026-10-13 that ties the growing of AIs to the build.

2. The tables

Class names are the ones used in the map's own trailing comments, so a number on the map can be traced back to the row that produced it. A trailer reading # R-COARSE C2: ... names this table.

2.1 Statement registers

The base table, derived from the chapter 5 supplement pages before the pilot region was authored (2026-07-17) and unchanged since:

The classes are symmetric: a class states how firmly the authors hold a proposition, whichever way the proposition points.

2.2 R-STEP: the language of a reasoning step

A reasoning link carries its own number, and it is read from the language attached to the step, not to the conclusion. The lookup registered before the campaign:

Two tables rather than one, kept separate on purpose even where the values coincide: how firmly an author holds a claim and how strongly they present a step are different measurements, and merging the tables would hide the difference.

A strength is a lower bound. The step holds in at least that share of the cases where its premises hold; the other cases are not argued by the line, and the solve counts them at even odds, so a flat 0.90 link realises 0.95 where nothing opposes it (0.9, plus half of the open 0.1; the mathematics page, section 3.2). An author who means an exact conditional writes the link as a pair, a floor for the conclusion and a floor against it, the form section 2.6 gives a statement; the flagship carries no such link. Reading every flat link as an exact conditional instead is the largest lever on the headline: the title claim would read 0.478 where it reads 0.624 (every link at or above 0.5 read that way), or 0.413 on the supporting links alone (measured 2026-10-13; 0.471 and 0.406 against 0.609 on 2026-10-07).

2.3 R-NORM: deontic language

The policy half of the book asserts norms ("no one may build this", "it must be shut down") rather than predictions, so the deontic classes were registered separately:

2.4 R-OPP: the opponent's material

An objection on the map is a claim the authors do not hold, so its number is the authors' credence in it, read from how they treat it:

P-CONTEST is the mirror rule: a premise the authors deny flatly (C7, 0.90) enters at 0.10, one they deny with a hedge enters higher. It never enters at 0, because "we do not think so" is weaker than "impossible".

An objection carries two numbers, and only the first is the authors'. The table above prices the objection's premise, the claim the objector asserts. The reasoning link from that premise to the claim it attacks carries a second number: how often the objection would defeat the claim if its premise were true and nothing answered it. The book almost never says, because the authors rarely grant their own answers failing. So on the flagship most objection links keep a role default under R-FALLBACK (section 2.5): 0.15, or 0.20 for an objection the book engages at length or that concedes most of its case, as the 2026-07-19 pass recorded them ("grounds .10, defeats .15"). Counted on 2026-10-06 (the objection-target census of that day): of the map's 113 objection links, 81 sit at 0.15 or 0.20; on 2026-09-23, 75 of 109 did, and five took their number from the authors' own confidence language. Those defaults are the mapmaker's, and this page's claim that the numbers come from the authors' words covers the premises and the supports, less the flat links section 2.5 counts; for most objection links it does not apply.

The default matters because of what the number does in the solve. Since 2026-09-23 (the mathematics page, section 3.8), an objection's number caps what the supports can claim in the cases where the objection stands unanswered: where its premise holds and the book's answer to it fails, the objection keeps its share of those cases, and any number of supports count there as much as the strongest of them. An objection link at 0.15 beside a support at 0.85 leaves the claim at 0.85 in those cases. Beside supports at 0.9, whose number and the objection's cannot both hold, the claim reads 0.895 there, and the overlap counts as a contradiction that makes those cases rarer; under the rule used before, five supports at 0.9 beside the same objection read 1.000. Where the answer holds, the objection is out of play. The policy leans on these defaults more than it did: raising all fifteen answered objections to the policy conclusion @shutdown from their defaults to 0.9 moves it from 0.929 to 0.838, just under the authors' own 0.85 to 0.95, and leaves the title claim at 0.624 (measured 2026-10-13 on the map with that day's grown line, solve_map.py at its default with --override setting each of the fifteen to 0.9: $obj-enf, $obj-defect, $obj-defect-doubt, $obj-monitor, $obj-ben, $obj-chips, $obj-resban, $obj-govt, $obj-goodguys, $obj-coop, $obj-waitshot, $obj-panic, $obj-messaging, $obj-hopeless, $obj-whom, every line that argues against @shutdown with a $resp-* answer; 0.928 to 0.836 with the title claim at 0.609 on the map of 2026-10-07, re-read on 2026-10-12; 0.928 to 0.834 on 2026-09-27, the evening, the title claim then moving by less than 0.001). The same test moved it from 0.952 to 0.915 that morning, with fourteen objections, and from 0.957 to 0.938 on the map of 2026-09-23; the difference is a flat support of the policy that read the authors' "aligned to whom?" page backwards and is now the objection that page answers. A pass that asks the book for these numbers directly has not run.

2.5 The procedural rules

Three registered with the tables, plus the standing authoring convention the pass worked under:

• R-FALLBACK, silence. Where the book and its supplement pages give no confidence language, the node keeps a role-based default, keeps its ?, and is listed in the pass log as residual mapmaker voice. Never invent an author number. One inheritance exception is licensed: a step that restates a sentence the authors do assert takes that sentence's class, because the assertion is the step's own language. How much of the map this rule carries, counted 2026-10-07 on the flat 0.90 links: of the 137 the solve reads (140, less three summary links whose finer steps it reads instead), 65 carried the number by its class alone, with no author sentence recorded behind it. Two of those, $goals-ev and $ext-core, the spine of the case that a misaligned superintelligence is lethal, were re-elicited from the book's own words on 2026-10-12 and stand at S2 0.90 with their sentences recorded, so 63 remain (the same 137 and 63 on 2026-10-13). Moving the 63 by 0.05 moves the title claim from 0.624 to 0.613 or 0.636 (1.1 to 1.2 points); the 65, those two among them, move it to 0.588 or 0.664 (3.6 to 4.0 points); moving all 137 moves it to 0.540 or 0.719 (measured 2026-10-13; on 2026-10-07, with the title claim at 0.609, the 65 moved it to 0.576 or 0.647 and the 137 to 0.529 or 0.700). The other row of that size is the pair a ground the book grants as fact carries, 0.85/0.05 (section 2.6), on 123 pins: at 0.80/0.10 the title claim reads 0.576, at 0.90/0.01 0.671, about five points either way (measured 2026-10-13; on 2026-10-07, on 124 pins, 0.557 and 0.663).
• R-COARSE, summary links. A coarse summary link takes its number from the book's own one-sentence formulation of the step. The book's title sentence is at least C7, and C2 where the authors' own disclaimer is attached to the summary sentence itself. A mismatch between a summary and the finer steps under it after solving is a readout, not an inconsistency to tune.
• R-COMPOSE, stacked hedges. For "we mostly expect that probably X", take the outermost hedge's class. Where two readings are genuinely defensible, record both in the log and author the weaker.
• The residual rule. A statement's own number should carry evidence that is not already somewhere else on the map, because counting the same consideration twice is exactly what the accumulation machinery assumes away.

2.6 The pair extension (2026-08-23)

The book states neither points nor intervals: it states registers. Since 2026-08-23 a statement's authored credence is a pair, a floor for the claim and a floor against it, with the gap between them the mass the authors leave open. The register table gained a pair column, registered before any node was touched, and each two-sided pair's midpoint is the class point above. The two one-sided rows leave nothing against, so their midpoint sits higher than their point: 0.95 for the deontic row, 0.975 for the categorical one.

The width does work of its own. Besides recording the mass the authors leave open, since 2026-09-08 it sets how firmly the solve holds the number when a reader drags it: a narrow pair takes many contrary flips to move, a wide one gives easily. On this map the middle pin is worth about 18 flips of evidence (the median of its 236 pins, 2026-10-13), which is less than a typical mechanism step: under a reader's what-if the authors' assertions give before their mechanisms do.

3. Where a number can land

Three slots, and they are not interchangeable:

• A pin on a claim nothing on the map argues for. These are the map's interface to everything outside it, and they constrain the solve.
• A strength on a reasoning link, from R-STEP, read as a floor (section 2.2). Also constrains the solve.
• A check on a claim the map does argue for, written # check: 0.75..0.95. This is the authors' own all-things-considered confidence, displayed against the computed value and never constraining the solve. It is how the map can say "the argument delivers 0.60, and the authors state 0.75 to 0.95" without either silently overwriting the other.

The separation is what makes an audit possible. A claim the book argues for gets its number computed from the argument, and the authors' own stated confidence in it stands beside that number as an independent reading.

4. The badge: how a disagreement is measured

The solved value of a claim is compared with its check interval. The badge is the signed distance from the computed value to that interval, zero anywhere inside it. Past 0.10 the map marks the claim, and that mark is the whole point of the exercise: it says the authors' own numbers, run through the authors' own steps, land more than ten points away from the authors' own stated confidence.

Two facts have to hold before a badge means that, and both are checkable:

1. The numbers are the authors', assigned in advance. Sections 1 and 2.
2. The badge is not an artifact of how the map is solved. This one was false for a while, and the project found it and said so. Until 2026-09-08 the deployed solve used a uniform counting reference, under which asserting a conditional pulled its antecedent down. On the map as it stood on 2026-09-06, four of the five badges past 0.10 were partly that pull rather than the book, which was measured, recorded, and fixed by changing the reference (D159, D161), with the map text untouched. The mathematics page keeps the retired reference as a dated section so the old numbers stay reproducible.

Where the flagship stands today. Measured 2026-10-13 on the map as it stands since that day's grown line (below), with the solve's reference implementation (solve_map.py at its default, which the in-browser kernel matches to about 1e-7), under the rule that decides how a support and an objection combine (one draw per case, the mathematics page, section 3.8): one claim badges past 0.10, and it is the book's title claim. @everyone-dies computes at 0.624 against a stated 0.75 to 0.95, short by 0.126, the largest gap on the map. It is what the authors' own premises deliver through their own step: 0.835 that someone builds it, 0.949 that what is built is misaligned, 0.797 that a misaligned superintelligence kills us, and the step joining them at 0.99, a step that holds by the meaning of its terms (since 2026-10-06, below). The next largest gap is @fragile, that human values are a tiny, fragile target in goal-space, computed at 0.796 against a stated 0.85 to 0.95, short by 0.054, and then that third premise, @mis-ext, 0.797 against the same interval, short by 0.053 (on 2026-10-07, with the title claim at 0.609 and short by 0.141, @mis-ext came second at 0.779, short by 0.071). The policy conclusion @shutdown computes at 0.929, inside its stated 0.85 to 0.95. Sixty-seven claims on the map carry a check for the comparison to run against, and 18 of them sit outside their interval by more than 0.01: four below it and fourteen above (on 2026-10-07, 18 of 66: five below and thirteen above; on 2026-09-28, the evening, 17: four below and thirteen above).

Several changes brought the map here, and they differ in kind. On 2026-09-23 the rule that combines the numbers changed and no number on the map moved: the title claim read 0.796 where the rule before read 0.827, and 15 claims sat outside their interval where 19 had. On 2026-09-25 the map itself changed, and the title claim with it, from 0.796 to 0.710 over that day's passes. A reader's demo had shown two claims naming one event, so the build claim was restructured; a pass found eight arguments entered twice and counted each once; and an audit read every check against the sentence it cites. Six checks turned out to price a different quantity from the claim they sit on, and were re-read. The title claim's was one: its old 0.88 to 0.98 was read from the book's "If any company or group ... builds an artificial superintelligence ... then everyone, everywhere on Earth, will die", which prices the conditional the map carries as its own step, so the check now reads the book's unconditional "The default outcome is lethal", 0.75 to 0.95. The title claim's gap is what prompted that audit, which is the risk rule 4 exists to name, and the note beside the check in the map says so. On 2026-09-26 two steps that hold by the meaning of their terms, that nobody builds a superintelligence that cannot be built and that one which cannot win the fight cannot kill us, went from 0.95 to 0.99 by a ruling of that day, and the title claim from 0.710 to 0.708.

On 2026-09-27 the map changed again, and this time the change took the title claim down: from 0.708 to 0.617. About nine points of the 0.708 had come from the solve and not from the book. Where no reasoning link on the map says anything about a claim, the network gives it even odds, and two kinds of case had no link into the title claim: a superintelligence that is built and aimed at what we want, and one that is built misaligned and does not kill us. The book does speak to both. Its chapter 12 and a chapter 11 supplement page say an aligned superintelligence leaves humanity alive, and the second case is ruled out by what the two claims mean. Two reasoning links now say so, with their strengths fixed before the solve was run (0.90 from the table for the first, the book's flat "would"; 0.99 for the second, which holds by the meaning of its terms), and the title claim reads what the authors' premises deliver and nothing else.

On 2026-09-28 two objections were rewired, and the title claim moved again, from 0.617 to 0.603. A reader who won the objection "if building it killed the builders too, they would stop" outright, with both of the book's answers rejected, saw the title claim stay near even odds (0.49 in his session of 2026-08-27, 0.37 on the map as it stood on the morning of 2026-09-28), because the objection was wired as doubt about the step from the three premises to the title: where it held, the step said nothing and nothing else on the map spoke to the claim. That objection, and the one that nothing so far has ended humanity, deny a premise: that a misaligned superintelligence causes human extinction. They are now reasoning links against that premise, and the book's answers to them stay what they were, doubts about the objection's inference. The record the second objection starts from, that no technology so far has caused anything like human extinction, is a fact the book grants, and it is now written as one (0.85/0.05, where it had carried the objection's own conclusion). That re-reading is what moved the number; the reader's won objection took the title claim to 0.104 that evening (0.106 on 2026-10-07, with both of the book's answers rejected). On the title's step, the one objection left wired as doubt is "experts disagree", which is doubt; the map's other objections aimed at a step are still to be read the same way.

On 2026-09-29 the number on the second of those objections was read afresh, and the title claim moved from 0.603 to 0.591. As a doubt about the step it had carried 0.3, and it kept that figure when it became a line against @mis-ext, where the same figure says something else: given the record, three cases in ten where a misaligned superintelligence is not lethal. Read under the table, the book gives that line no confidence language of its own, so it takes the role default of R-FALLBACK, 0.5 (section 2.5); the book's ridicule of the induction prices its answer, and the answer keeps that number.

On 2026-10-06 the step that joins the three premises to the title claim, $link-fine, went from 0.93 to 0.99 by a ruling of that day, and the title claim from 0.591 to 0.609. The 0.93 had been copied from the coarse line above it, which carries the title sentence's register and with it the authors' confidence in the whole chain. On the one step that holds once its three premises do (a superintelligence that is built, misaligned and lethal kills everyone by the meaning of the words), that figure priced the chain a second time. The step now stands where the map's other steps that hold by meaning stand, at 0.99, and the title sentence's register stays on the coarse line. Of the six changes since 2026-09-25 that moved the title claim, this is the one that raised it.

On 2026-10-13 a reasoning link tied the claim that AIs are grown by gradient descent, @grown, to the build step, and the title claim moved from 0.609 to 0.624. The title sentence fixes the technique ("using anything remotely like current techniques"), and until then the map read the building and the growing as independent. The link says that a superintelligence built with today's techniques is grown, at 0.99, a fifth step that holds by the meaning of its terms; a second link carries the authors' flat statement that today's AIs are grown whether or not one gets built, at the floor of C7's pair, 0.85 (section 2.6). @grown's pin became a check, 0.85 to 0.95, read from the same register, and it computes at 0.981, one of the fourteen claims above their interval. Of the seven changes since 2026-09-25 that moved the title claim, this is the second that raised it.

It is worth stating what this result is and what it is not. It is not the map agreeing or disagreeing that the book is right. It says the authors' stated confidences are mostly consistent under this model: the policy conclusion sits inside the authors' interval, and every gap but one is under the badge threshold. The one is the title claim: its premises, multiplied through the book's own conjunctive step, deliver 0.62 where the authors state 0.75 to 0.95, and the map keeps that badge standing. A perfectly coherent argument can still rest on false premises, and this instrument has no view on that.

5. What the audit found, and what happened to each finding

The 2026-07-19 pass ended with badges standing, and the rule was to keep them rather than repair them. Their history since is the honest test of that rule:

• Steps outrun summaries, across the whole map. Where the authors argue flat, several convergent unhedged assertions compound to near-certainty while the authors' own summary of that section stays hedged. Under the model that is a finding about the book's prose, not an error, and those gaps are still on the map: the ten largest today sit between 0.033 and 0.126, and six of the ten are claims the argument delivers more firmly than the authors state them (measured 2026-10-13; the largest, 0.126, is the title claim's shortfall, the finding below, and the second and third, 0.054 and 0.053, are @fragile and the title claim's third premise @mis-ext; on 2026-10-07 the ten sat between 0.033 and 0.141, also with six above, @mis-ext second at 0.071; on 2026-09-28, the evening, the ten sat between 0.033 and 0.147, also with six above, on 2026-09-27, the evening, between 0.033 and 0.133 with seven above, and that morning between 0.033 and 0.054). Part of the old gap was the solve's own rule letting many supports outvote the book's objections; the one-draw rule closed that part without touching a number (the summed distance to the 64 intervals of the day fell from 0.695 to 0.578 on 2026-09-23; on 2026-09-11 the ten largest sat between 0.040 and 0.053).
• The evolution analogy under-delivered. The pass computed @evo-analogy at 0.449 against a stated 0.85: the authors grant the disanalogy so thoroughly that their own concessions left the analogy standing at half their stated confidence. It stayed on the map as a badge for seven weeks. It closed on 2026-09-08, and the way it closed is the point: nobody moved a number. A later pass found that the chapter's own case for the analogy had never been mapped, mapped it, and the claim rose to 0.858 (measured 2026-09-11), inside its interval, where it still sits (0.890 against 0.75 to 0.95, measured 2026-10-13; 0.847 on 2026-10-07). A badge is a question about the map or the book, and the licensed answers are "map the passage that was missed" and "leave it standing", never "adjust the number".
• The title claim's filler. Until 2026-09-27 the title claim read 0.708, 0.042 under the authors' 0.75 to 0.95, and about nine points of that number were the network's even odds in two kinds of case no link reached (section 4). A reader asking what the map says if the built superintelligence is not misaligned saw the title claim at 0.43, mostly filler. Mapping what the book says about those cases, with strengths fixed from the table first, took the claim to 0.617 and made it the one badge past 0.10 (0.603 after the objection rewiring of 2026-09-28, 0.591 after the re-read of 2026-09-29, 0.609 after the ruling of 2026-10-06 on the last step of its chain, and 0.624 since the grown line of 2026-10-13, section 4). Rule 4 forbids editing toward agreement; this edit mapped what the pass had missed, and it moved the number away from the authors' interval, where the badge now stands.
• The headline claim's residual. The pilot region reproduced the authors' conclusion to within 0.07 and localised what was missing to one named conjunct and one deliberately excluded reasoning link, rather than leaving it a mystery. The exclusion was itself pre-registered: adding the link after seeing the solve would have broken rule 4.

6. Honest limits

• A verbal-to-probability table is a modeling decision. The table is fixed and public, which makes the numbers auditable. It does not make them correct. A different defensible table would move most of them, and the project claims only the first thing.
• Reading a register as a credence has a known cost. An author's stated confidence in a claim may already discount what they believe about its consequences, so feeding it back in as evidence can apply the same reasoning twice. There is no operational alternative, because the text never gives the undiscounted figure. The size of the effect was measured and is small at these widths, and the compensation is to show the interval rather than present a midpoint as the author's number.
• The classification is judgment, sometimes. A sentence can sit between two rows of the table, and by rule 2 each region's classes were read by a single elicitor, with no second reader over the class itself. Where a register was borderline, one person's ear settled it. The 2026-07-19 pass log records two numbers a second agent read differently, @mis-ext and @precedented, both settled by authoring the weaker of the two readings. (@precedented was re-read on 2026-09-28: it now carries the record the objection starts from, which the book grants, and the objection's own conclusion moved onto the objection's link; section 4.) The one label on this map that did get a systematic second pass is a different one, the kind of step each reasoning link makes: on 2026-09-08 two blind passes over all 348 links disagreed on 68 of them, and a third reader settled those.
• This rubric is for one book. It was derived from the confidence language of one text in one genre. A spoken debate needed its own table (the register for speech is looser than for edited prose), and maps written since carry their rubric inside the map file itself.
• The solve is not a truth oracle. It says what a set of authored numbers jointly implies, and has no view on whether they are right.

7. The record

The dated registration entries, one per pass, live in the project's authoring log, which is not yet public along with the rest of the working repository. This page is the distillate of those entries and carries every rule they registered.

What you can check without the repository is the map itself. Open any claim on the flagship map and its detail panel shows the authored number, the computed one, and the source passage the number was read from. The trailing comment on a node in the Editor pane names the rubric class that produced its value, so a number on the map traces to a row in section 2, and the quote lines under a node give the sentence it was read from, with its locator in the book.

Numbers on this page were measured on 2026-09-11, 2026-09-23, 2026-09-27, 2026-09-28, 2026-10-07 and 2026-10-13 against the solve, and each says so where it appears; the dated record of each change in section 4 is the project's authoring log for that day.