Tools · the method behind the checker

Conditions of Possible Experience

The instrument is at /tools/can-these-all-be-true/. This page says what it computes, whose ideas those are, what each kind of answer proves and does not, who else does something like it, and what checks stand behind it.

  1. The idea, and whose it is
  2. What each answer proves
  3. Given something of probability zero
  4. The Bell experiment
  5. The examples and their sources
  6. Who else does this
  7. The checks
  8. The record, as a format
  9. Limits
  10. References

1. The idea, and whose it is

Name a few events. Each can happen or not, so n events give 2n cases, and a world is a way of sharing out a total chance of 1 among the cases. A statement such as P(rain) = 0.3 says that the shares of the cases in which it rains add up to 0.3. Every statement the instrument accepts has that shape: some shares, added and subtracted, compared with a number. So a list of statements is a list of linear conditions on unknown shares, and three questions about it can be settled exactly. Is there any world that satisfies them all? If so, how small and how large can the probability of some other event be? And what is the proof?

This is George Boole’s problem. In An Investigation of the Laws of Thought (1854) he says that some problems in probability need

a distinct method determining the necessary conditions among their data, in order that they may represent a possible experience, and assigning the corresponding limits of the final solutions.
Boole, The Laws of Thought, chapter XVIII, section 1

and he gives the conditions a first name:

By the term statistical conditions, I mean those conditions which must connect the numerical data of a problem in order that those data may be consistent with each other, and therefore such as statistical observations might actually have furnished.
chapter XIX, section 1

The necessary conditions among the data are the instrument’s first verdict, and the limits of the final solutions are its second. Eight years later, in a paper read to the Royal Society, Boole gave the conditions the name this page takes its title from, and said what they tell you:

These, together with the former, may be termed the conditions of possible experience. When satisfied they indicate that the data may have, when not satisfied they indicate that the data cannot have, resulted from actual observation.
Boole, “On the Theory of Probabilities”, 1862

The same paper sets out the way of finding them that the instrument follows, a method it says is taken from the 'Memoir on the Combination of Testimonies and of Judgments,' already referred to, Boole’s earlier paper in the Transactions of the Royal Society of Edinburgh. Let the events in the data be resolved into the ultimate possible alternatives which they involve, he writes, and give those alternatives unknown probabilities; then, as the probability of each event in the data is equal to the sum of the probabilities of the alternatives which it involves, we shall have a system of equations. His ultimate possible alternatives are this page’s cases, and their unknown probabilities are its shares. Boole then removes the unknowns by hand, playing one inequality off against another. (His method for getting a single number out of such data, as distinct from limits, brought in auxiliary events of which the same paper says the ideal events are regarded as independent. None of that is used here: only the conditions and the limits.)

Handing that system to a linear program is Theodore Hailperin’s step. A paper of 2026 on the problem’s difficulty tells the history the same way: it calls it A problem dating back to Boole and says that Hailperin, in 1965, phrased the problem in the language of linear programming and generalized it to logical formulas of the events other than disjunction. The same paper shows that computing the optimal interval for the union probability is NP-hard, and that problem is a special case of this one, so no method is known that is fast on every problem of this kind. This instrument goes through all 2n cases, so it stops at sixteen events and says so, where a quieter tool would simply slow down.

The instrument solves the linear programs in exact fractions. Cases that no statement can tell apart are counted once, which keeps most problems small, and the working problem is grown a few cases at a time until every other case has been priced against it exactly. A floating-point solver runs first, as a scout: it proposes which cases matter and which of them the answer stands on. That proposal is then either proved, by two exact solves and a check in whole numbers of every condition that makes an answer the best one, or thrown away, and the exact solver finds its own way. Nothing the scout computes is kept.

2. What each answer proves

Every answer comes with a certificate, and the certificate is checked by a second program before the page shows anything (section 7). The kinds:

They can all be true: a world

The proof that some world fits is a world: a table of cases and shares. Anyone can add up the shares that each statement speaks of and compare. Note what this does not prove. It says nothing about whether the statements are true of the actual world, only that they do not contradict each other.

A range, or a single value: a world at each end and a weighted sum

If the answer to a question runs from L to U, the instrument shows a world in which it is L and one in which it is U, so both ends are really reached. That nothing lies outside is a different kind of proof. Read each statement as a bet you consider fair: “P(rain) = 0.3” prices a ticket that pays 1 if it rains at 0.3, so buying one gains 0.7 in the cases where it rains and loses 0.3 in the others. In any world that fits the statement, that gain averages to zero. The proof that P(Q) is at most U is a set of weights on the statements such that, in every case, 1 if Q happens (0 if not) plus the weighted gains comes to no more than U. Averaging over any fitting world, the gains contribute nothing, so P(Q) is at most U. The page prints the weights and the table of cases. When L equals U the statements determine the answer.

Lewis Carroll’s readers found exactly such a proof by hand, in an answer printed in A Tangled Tale (1885; section 5): weight every statement once, and “adding the wounds together” does the rest.

They cannot all be true: a Dutch book

If no world fits, there are stakes on those same bets such that the total is a loss in every case where any of the bets is on: whoever holds the statements, and so calls each bet fair, can be made to lose. Such a set of bets is called a Dutch book. A standard reference traces the argument to Frank Ramsey (it traces to Ramsey’s work in “Truth and Probability”) and states the theorem like this:

Given a set of betting quotients that fails to satisfy the probability axioms, there is a set of bets with those quotients that guarantees a net loss to one side.
Stanford Encyclopedia of Philosophy, “Dutch Book Arguments”

The betting reading of a probability used throughout this page is Bruno de Finetti’s; a 2026 study cited in section 5 calls the result it builds on a theorem due to de Finetti. A 2013 paper on conditional probability puts the matching definition this way:

a probability assessment is coherent if and only if, in any finite combination of n bets, it may not happen that the values […] are all positive, or all negative (no Dutch Book).
Gilio and Sanfilippo, 2013, glossing their Definition 1

The instrument reports the book that loses the most per unit staked on the statements it uses, and proves that no other does better: the record carries a world, on the cases where a bet is on, in which none of those statements misses by more than that loss, and any stakes whose sizes sum to 1 must average at least minus the loss there. It also reports a set of conflicting statements with nothing spare: drop any one of them and the rest can all be true, and the record carries a world showing it for each. That set is not always the smallest there is. When more than one conflict hides among the statements, which one is reported can depend on the order of the lines. For a one-sided statement (“at least”, “at most”) the instrument only ever stakes the side you committed to.

A “more than” that cannot hold: bets that never win

A statement written with > or <, and no “given”, has to hold strictly. If the statements can all hold only with that one at equality, then no world fits them as written, and yet no bet loses for certain either. The proof is of a second kind and is just as complete: a set of bets, each of which you have called fair or better and one of which you have called strictly better than fair, that pays nothing in any case. Bets like that would have to gain on average, and cannot. The same reasoning marks an end of a range that is approached and never reached: the record carries weights showing that, in any world where the value sits exactly at that end, the strict statement could hold only with equality.

3. Given something of probability zero

A conditional statement, P(A | B) = 0.9, is usually read as a ratio: P(A and B) divided by P(B). Read that way it says nothing when P(B) is zero, and two statements that flatly contradict each other, such as P(A | B) = 0.9 with P(not A | B) = 0.2, are both “satisfied” in a world where B never happens. De Finetti’s reading is stricter and is the one used here: a conditional probability is the price of a bet that is called off, with the price returned, if the condition fails. In the words of the 2013 paper, if you put the probability of B given A at p,

then you agree to pay an amount p, by receiving 1, or 0, or p, according to whether AB is true, or ABc is true, or Ac is true (bet called off).
Gilio and Sanfilippo, 2013, section 2.1

Stake on both of the statements above and you lose 0.1 whenever B happens and never win, so by the definition quoted in section 2, taken over the cases where at least one bet is on, they are incoherent.

The test runs in layers. First find a world that fits everything. Any statement whose condition has no share in any such world is set aside with the others like it, and those are tested again among themselves, on the cases where one of their conditions holds; and so on down. The statements are coherent exactly when no layer fails. This is the procedure Gilio and Sanfilippo set out as their Algorithm 1, which they credit to earlier work by Biazzo, Gilio and Sanfilippo (2008): check that the system can be solved; if it cannot, stop; otherwise find the statements whose conditions have no share in any solution; if there are none, stop; otherwise repeat with those. The paper bases the algorithm on its Theorem 5, about previsions, which it cites from that work; for conditional probabilities given exactly, the matching result is its Theorem 3, which it cites from Gilio (1993): the statements are coherent when the system can be solved and the statements whose conditions have no share in any solution are coherent in turn. Those theorems, as the paper states them, are for statements of exact value. This page applies the same test to statements with “at least” and “at most”, which the same authors treat in work not read here, and to strict ones; the verifier tests that extension against the instrument’s separate routes and against a solver written independently (families D, K and L). When a statement’s condition can never happen in any fitting world the page says so, because a reader should know that the statement is then saying nothing.

Questions, read the same way

A question P(Q | H) is taken to ask: what prices could a bet on Q, called off unless H happens, have without making the whole set of statements incoherent? The instrument works that out in levels. If H has some share in every world that fits the statements, the answer is the range of the ratio of P(Q and H) to P(H) over those worlds, and its proof has two parts: the weighted sum of section 2, and a second weighted sum showing that the statements never let H go without a share. Together they are a Dutch book against any price outside the range. If some fitting world gives H no share, then in that world the bet is called off in every case that counts, and only the statements whose own conditions also have no share in any such world still hold its price. With those statements, on the cases where H or one of their conditions holds, the question is asked again. The statements only get fewer, so this ends. That each end of the final range is really allowed is shown by worlds: one for each level passed on the way down, and one at the bottom in which H has a share and the ratio is the value claimed.

Two things follow that a ratio cannot say. A conditional you stated outright is given back when you ask for it, even if its condition has probability zero: with P(B) = 0 and P(A | B) = 0.3, the answer to P(A | B) is 0.3, where a ratio gives 0/0. And probability zero is not the same as never. With P(A and B) = 0, the answer to P(A | B) is 0 if B has any chance, and anything at all if it has none, because a bet on A given B is then held by nothing you have said. State it as a fact, never (A and B), and the answer is 0 outright. The page shows both parts when they differ.

This procedure was worked out here from the theorem above, by applying the layered test to the statements with the question added at a trial price. It is not offered as new: that paper’s reference list points to earlier work on “coherence checking and propagation of lower probability bounds” (Biazzo, Gilio and Sanfilippo, 2003), which was not read for this page. Nor is the derivation taken on trust. In family K of the verifier, every end of every range is added back as a statement and must pass the instrument’s separate test for statements, and a value just outside must fail it.

A “more than” is held exactly here too. A strict statement has no condition, so it binds only at the first level, and the one thing it can change there is whether a world that fits gives H no share. If none can, H has a share in every world that fits, though perhaps no least share: with P(A) > 0 and P(A and B) = 0, the answer to P(B | A) is exactly 0. The proof that H cannot go without a share is then of the kind that marks an open end: weights under which, in any world giving H no share, the strict statement could hold only with equality; and a world that gives H no share and fits the statements read as “at least” shows that no least share exists. If some world holding the strict statements strictly does give H no share, the levels below go on as before, and every world shown at the first level holds the strict statements strictly.

4. The Bell experiment

Two boxes, far apart. On each trial each box is asked one of two questions and answers yes or no. Suppose three things, as common sense would. First, that on every trial each box had an answer ready for both of its questions, whichever was asked. Second, that those answers did not depend on what the other box was asked. Third, that the choice of questions had nothing to do with the answers that were ready, so that the trials on which a given pair of questions happened to be asked are a fair sample of all the trials. Call the four ready answers a1, a2 (one box) and b1, b2 (the other). Then every trial is one of sixteen cases, the rate at which a1 and b1 agree is a statement about a world over those cases, and likewise for the other three pairs. Four statements, one world.

In the experiment reported from Delft in 2015 the authors write that they perform 245 trials testing the CHSH-Bell inequality and find S = 2.42 ± 0.20. No trial asks more than one pair of questions, so the 245 trials are four separate samples, one for each pair. The counts of agreeing and disagreeing answers are printed above the bars of the paper’s Figure 4a: 46 and 7, 63 and 16, 46 and 16, 10 and 41. Those are the agreement rates in the instrument’s example: 46/53, 63/79, 46/62 and 10/51.

Entered as four statements, they come back incoherent, and the Dutch book the solver finds is this: buy one ticket on each of the first three agreements and sell one on the fourth. In every one of the sixteen cases at most two more of the first three agree than the fourth does (try to make all of the first three agree and the fourth disagree: it cannot be done), so the tickets pay at most 2 net, and they cost 46/53 + 63/79 + 46/62 − 10/51, which is more than 2. The loss is at least 0.2112… in every case, or 0.0528… per unit staked. That inequality, found here by a linear program as a certificate and checkable against sixteen rows, is the CHSH inequality of Clauser, Horne, Shimony and Holt (1969). In the verifier’s sweep of random agreement rates, every Dutch book the instrument returned against an incoherent set of four was one of the eight inequalities of this form.

None of this is new. That Bell’s inequalities belong among Boole’s conditions is Itamar Pitowsky’s observation: the abstract of his 1994 paper, George Boole’s ‘Conditions of Possible Experience’ and the Quantum Puzzle, says that Some of Boole's conditions have been rediscovered in more recent years by physicists, including Bell inequalities, Clauser Horne inequalities, and many others. Arthur Fine had shown in 1982 that, for an experiment of this kind, the statement There is one joint distribution for all observables of the experiment, returning the experimental probabilities is equivalent to the statement The Bell inequalities hold. And finding the inequality by linear programming was described in 2009 by Matthew Elliott, whose program, when the correlations fit no such world, outputs a Bell inequality violated by the correlations. This page adds only that you can type the numbers in and check the answer row by row.

What this shows and what it does not. It shows that these four observed rates cannot all be the rates of one world of ready answers. It is not a contradiction in what was observed: each rate comes from its own sample, and four samples can disagree with any single world. So the bets are a proof about four numbers, not a wager anyone could place, since no trial would settle all four. Pitowsky, on page 113 of the 1994 paper as Samson Abramsky quotes it: A violation of Boole’s conditions of possible experience cannot be encountered when all the frequencies concerned have been measured on a single sample. Such a violation simply entails a logical contradiction; ‘observing’ it would be like ‘observing’ a round square. We expect Boole’s conditions to hold even when the frequencies are measured on distinct large random samples. But they are systematically violated, and there is no easy way out. What gives way is one of the three suppositions at the head of this section.

And four rates from 245 trials are subject to chance. Whether a world of ready answers could have produced them by luck is a statistical question, not a logical one, and the experimenters answered it separately: they report a probability of p = 0.039 that a local-realist model for space-like separated sites produces data with a violation at least as large as observed. The theory’s own prediction for the ideal experiment, the paper says, is S = 2√2, about 2.83. Shared evenly among the four pairs, that is agreement about 85.4% of the time on three of them and 14.6% on the fourth, and those rates, typed in, are incoherent too. The instrument answers the logical question about the rates as given.

5. The examples and their sources

The weekend forecast. Made-up numbers. The point is what a calculator that multiplies adds. Two popular ones say of themselves that they are for independent events: calculator.net’s probability calculator is introduced with To find out the union, intersection, and other related probabilities of two independent events., and Omni Calculator’s tells its user Your problem needs to be condensed into two independent events. Neither is wrong. Each answers a narrower question than the one a forecast poses.

Boole’s thunder and hail. His first worked example in the chapter that illustrates the general method:

The probability that it thunders upon a given day is p, the probability that it both thunders and hails is q, but of the connexion of the two phænomena of thunder and hail, nothing further is supposed to be known. Required the probability that it hails on the proposed day.
chapter XVIII, example 1

His answer contains an unknown, the chance of hail when there is no thunder, and he closes by giving its two extremes: Inferior limit = q. Superior limit = q + 1 − p. With p = 0.3 and q = 0.1 the instrument returns 1/10 and 4/5, and the verifier compares the instrument with those two formulas for every p and q in tenths.

Carroll’s pensioners. From A Tangled Tale (Macmillan, 1885), Knot X:

Say that 70 per cent. have lost an eye—75 per cent. an ear—80 per cent. an arm—85 per cent. a leg—that'll do it beautifully. Now, my dear, what percentage, at least, must have lost all four?
Carroll, A Tangled Tale, Knot X

The printed answer is Ten., with a solution Carroll credits to a reader signing as Polar Star: Adding the wounds together, we get 70 + 75 + 80 + 85 = 310, among 100 men; which gives 3 to each, and 4 to 10 men. The instrument’s proof of the lower end puts weight 1 on each of the four statements, which is the same sum. It also gives the upper end, 7/10, which the Knot does not ask for.

Another reader, Janet, multiplied the four percentages together. Carroll’s note on her answer: I am sorry she has misunderstood the question, and has supposed that those who had lost an ear were 75 per cent. of those who had lost an eye; and so on. Of course, on this supposition, the percentages must all be multiplied together. This she has done correctly, but I can give her no honours, as I do not think the question will fairly bear her interpretation. Typed in on her reading, with each percentage given the ones before it, the instrument returns the single value 357/1000. (Project Gutenberg’s plain-text file prints her answer as 35 and 7/10ths; its HTML file prints 2/10ths. The product is 35.7.)

Two things more likely than not. Computed here. The lower end, 0, is a limit that is never reached; section 2 says how that is proved.

The Harvard question. A study published in 1978 reports that its authors asked 20 house officers, 20 fourth-year medical students and 20 attending physicians, selected in 67 consecutive hallway encounters at four Harvard Medical School teaching hospitals a question that begins If a test to detect a disease whose prevalence is 1/1000 has a false positive rate of. A 2014 repetition of the study quotes the question in full, with the rate at 5% and the thing asked as what is the chance that a person found to have a positive result actually has the disease, and states the convention that the usual answer rests on: Assuming a perfectly sensitive test, we calculated that the correct answer is 1.96%. The question itself gives no sensitivity, so the instrument returns a range, from 0 to 20/1019, whose top end is that 1.96…%; add the missing statement and it returns the single value. (The 1978 abstract was read in a copy that is cut short where the quoted question breaks off. The sentences of the 2014 paper are from the excerpt that this site’s layer on Bayes’ theorem already keeps.)

A conjunction ranked too high. The mistake is named in the title of Tversky and Kahneman’s 1983 paper, Extensional versus intuitive reasoning: The conjunction fallacy in probability judgment, whose abstract states the rule it breaks: The probability of a conjunction, P(A&B), cannot exceed the probabilities of its constituents. The bank teller in the example is the usual illustration; its wording here is this page’s own, and no source was read for it.

A forecast against itself. Made-up numbers; the kind of error is not. A 2026 study by Isaiah Andrews and Suproteem Sarkar elicited forecasts from language models about events generated from stock returns and, in its words, used linear programs to compute the largest Dutch-book profit against them; its abstract reports substantial evidence of incoherence in language model forecasts. The quantity this page prints for an incoherent set, the guaranteed loss per unit staked, is a measure of the same kind, and measures of that kind are older than either: a 2015 paper by Glauber De Bona and Marcelo Finger is titled Measuring inconsistency in probabilistic logic: rationality postulates and Dutch book interpretation.

A poll that adds to 101%. Made-up numbers. Read exactly, three shares of 46%, 38% and 17% cannot all be true of answers given one per person. Read as figures rounded to the nearest per cent, each stands for a range half a point either way, and a world fits. The page reports that as its own verdict, not as a contradiction.

Boole’s three symptoms. From the 1862 paper, just after the method quoted in section 1. He supposes medical statistics in which two symptoms x and y are seen together in two fifths of the cases of a disease, y and z in two thirds, and z and x in four fifths, and concludes: The above formulae would show that the evidence was contradictory. The condition that fails is that the first share be at least the sum of the other two less 1, and 2/5 is less than 2/3 + 4/5 − 1 = 7/15. The instrument’s bets are that condition turned into money: sell a ticket on the first pair and buy one on each of the others. (The paper was read in two machine transcriptions of the same scan. Their prose agrees and their formulas are damaged in different ways, so only prose is quoted, and the one formula used was checked by working the example.)

Three majorities. Computed here: with three options and every voter holding a ranking, each of the three majorities in a cycle can be as large as two thirds and no larger. The verifier checks the same for cycles of four and five (three quarters, four fifths).

A Bell experiment. Section 4.

Given a chance of zero. Section 3.

6. Who else does this

The mathematics is old and public, and there is software for parts of it. What was found, with what each says of itself:

The first search also turned up, in one search summary, research solvers for the same problem that were not opened: NILS, the PSAT solvers of Finger and De Bona, Pronto and diff-SAT.

So a verdict in a browser is not new: PrSAT gives one, with a model when the answer is yes. Nor is a range: Pinference computes Hailperin’s bounds, and CkC checks conditional assessments for coherence and extends them, both outside a browser. A first search for this page, twelve queries on 15 November 2026 (eight on the web and four on GitHub), did not find PrSAT. An independent agent, sent to attack the first build, did. What this instrument adds is therefore narrower than it first seemed, and is these things, in a browser: the range of anything asked, with a world at each end that is reached and a proof for each end that is only approached; a proof for every answer that a reader can check by hand and that a separate program does check, including a proof when the answer is no; de Finetti’s reading of conditional statements and of questions; and a link that reproduces the whole computation in the recipient’s browser. None of that is new mathematics, and a search is not a proof of absence.

7. The checks

There are two programs. engine.mjs finds answers. checker.mjs shares no code with it, imports nothing, and contains no solver: it reads the statements with its own reader, goes through every case, and does nothing there but arithmetic on its own exact fractions: multiply, add, compare. The page shows a verdict only after the checker has confirmed it for that exact record, and the tables of cases on the page are computed by the checker’s code. A wrong answer would therefore have to be wrong in two unrelated programs in the same way.

The verifier, verify.mjs, runs 20,631 checks in the twelve families below. A thirteenth, family J, then makes 27 more that are not counted in that total: it reads the figures and sums printed on this page back and compares them with the run.

familywhat it holdschecks
Athe reader: what it accepts, what it refuses and why (an ambiguous decimal comma among them), and that its own reading reads back as itself116
Bthe exact linear program on random problems, each self-certified, with the floating-point scout as a second opinion; the step that proves a proposed answer exactly, which must never return a different value; the work allowance1,512
Ctheorems with closed forms, swept over grids of fractions: Fréchet’s bounds, Boole’s two limits, Carroll’s rule, Bayes, the CHSH region, the triangle inequality, cycles, independence4,438
Drandom problems: worlds invented first so that the truth is known, and arbitrary statements; every record proved by the checker, every conflicting set shown to have nothing spare4,189
Ea second derivation with no linear programming: brute force over a grid of worlds181
Fdoctored records, every one of which the checker must refuse183
Gthe worked examples on the page, the sums behind the Delft counts, and the sentences this page prints about them97
Hconditions of probability zero: layers of statements, questions answered level by level, empty conditions18
Irounded figures, links, limits and refusals, the work allowance, and sixteen events end to end33
Kquestions as bets: every end of every range added back as a statement and allowed, values just outside refused; “more than” held exactly, with closed forms6,666
La solver written independently, with its own fractions and a different linear-programming method, put to the same random problems: verdicts and first-level ranges must be identical3,000
Sthe pages against their sources: each claim on this page, the tool page and the examples that rests on a source is found on its page word for word, each quotation in the source’s own words, each excerpt in the copy that was kept198

Family F is the one that says the checker can fail: 151 records with one thing changed (a bound moved by a thousandth, a share raised, a stake reversed, a statement edited after the fact, a coherent set presented as incoherent, an end that is only approached presented as reached, a level left out, a line shown as read differently from what was checked, independence declared where it is not linear, a statement called needed without the world that shows it) are each refused, most of them for the reason the family names.

Two things found faults before this page shipped. The verifier found that the conflicting set reported was a real conflict but could have a statement to spare when conditions of probability zero were involved, because the shrinking step tested only the first layer; it now runs the full test, and family D checks it directly. Then an independent agent was set against the instrument. Its own solver, written separately, agreed with this one on 3,000 random problems, and it found what the verifier had not: a decimal comma read as a thousands mark, so that “2,5%” became 25%; “more than” quietly read as “at least”; questions about a condition of probability zero answered by the ratio alone; a conflicting set called the smallest when it was only one with nothing spare; problems inside the stated limits that ran for minutes; and a phrase this page had credited to Boole’s book of 1854 that is in his paper of 1862. Each is fixed, each now has checks in the families above, and the agent’s scripts and report are kept with the research. Afterwards its solver was set against this one again, on the first level of every answer, where the two can be compared. That comparison now runs every time, as family L.

Then, before the pages were published, four more agents were set against them: one to break the engine and the checker, two to read every claim on the pages against its source, and one to use the page in a browser as a first-time reader. The first found a record the checker accepted although it was false: independence declared between events whose probabilities were not stated, where the worlds that fit need not include the mixtures of their members, which every proof here takes for granted. The checker now builds for itself the statements that independence amounts to, and refuses any other kind. It found too that a “more than” beside a question whose condition might have no share was answered more loosely than the truth, that a figure written as 0 or 1 was treated as a rounded one, and that a semicolon inside a comment started a statement; those are fixed, and its own random problems, 6,400 of them checked against its own solver with “more than” held strictly, now agree. The two claims passes named 213 claims and read each against its source, and 32 needed a change, among them this page’s account of whose algorithm section 3 follows; the reader found a limit refused without its reason, an account of the lines that went stale while a long problem was worked out, and tables whose deciding column a phone did not show. Each is fixed. Since then the record also carries what the page had only said: a world for each statement of a conflicting set, showing it is needed; the world that shows no stakes can guarantee more than the loss printed; and each line’s reading, which the checker compares with the statements it checked.

Family S can tell that a quotation is the source’s own words and that the source was kept. It cannot tell whether a source supports the sentence that leans on it; only a reader can.

The published verifier is the program to read. It imports the instrument’s own files by their place in this site’s repository, so it does not run from a bare download; the checker does, and node checker.mjs --selftest runs it against seven known records and sixteen doctored ones (add --why to see the reason each doctored one is refused). The records in that self-test are the ones the engine writes for the statements named beside them, and family I fails if they ever differ.

8. The record, as a format

“Save the record” gives a JSON file, format aw-all-true/1, that carries everything needed to re-check an answer without this site: the names of the events, the facts, each statement as the page read it, and the certificates. The checker’s header states the rules it applies. In brief, with gi the gain of statement i as a bet in a given case (0 if its condition fails; otherwise the sum of its coefficients over the events that happened, minus its number; negated for a “<=” or “<” statement):

claim in the recordwhat the checker requires, in every possible case
world, worldLo, worldHishares are positive, sit on cases the facts allow, and sum to 1; every gi averages 0 (at least 0 for a one-sided statement, more than 0 for a strict one); the question takes the value claimed
certHi for “P(Q | H) is at most U”[Q and H] + Σ wi gi ≤ U [H], with wi ≥ 0 on one-sided statements
certLo for “at least L”[Q and H] − Σ wi gi ≥ L [H], with wi ≥ 0 on one-sided statements
certPos, for a conditional question[H] − Σ vi gi ≥ m, with m above 0: the statements the proof stands on (over) never let H go without a share. These three are checked on the cases where H or the condition of one of those statements holds. Or, where a strict statement s is what keeps H from having no share: gs + Σ vi gi + b [H] ≤ 0 with b below 0, and zeroWorld, a world that gives H no share and fits every statement read as “at least”
chainfor each level passed, a world that fits the statements still binding and gives H no share; the next level holds exactly the statements whose condition had no share in it
openLo, openHigs + Σ wi gi + b ([Q and H] − v [H]) ≤ 0 for a strict statement s and the end v: a world at exactly v would leave s holding only with equality
book, kind sure-losswherever at least one staked bet is on, Σ si gi ≤ −loss, with loss above 0 and si ≥ 0 on one-sided statements
book, kind never-winsΣ si gi ≤ 0, with a positive stake on at least one strict statement; and nearest, a world in which the staked statements hold read as “at least”
book.withoutfor each staked statement, worlds (in layers, as below) in which all the other staked statements hold: nothing in the set is spare
book.sharpesta world, on cases where a bet is on, in which no staked statement misses by more than the loss, and the stakes’ sizes sum to 1: no stakes can guarantee a larger loss per unit staked
book.asRatioswhere the conflict is below the first layer, a world in which every statement holds as a plain ratio
layerseach layer’s world satisfies that layer’s statements, and the next layer holds exactly those whose condition had no share
independenceall but one of the events declared independent has its probability stated exactly; the checker builds the statements that independence then amounts to, and the record must carry exactly those; and in each world shown for the first level the probabilities really multiply
lines, statementseach line’s reading is the statement, fact or question that was checked, and a “rounds to” reading is the band of the figure shown; what the record says of each statement (which way it faces, strict or not, conditional or not) is what its text says

Nothing in those rules needs the solver that found the answer. A record written by any other program, or by hand, passes or fails on the same terms.

9. Limits

10. References

  1. George Boole, An Investigation of the Laws of Thought (1854); read in the Project Gutenberg edition, ebook 15114.
  2. George Boole, “On the Theory of Probabilities”, Philosophical Transactions of the Royal Society of London 152 (1862); read in two machine transcriptions, JSTOR’s Early Journal Content at the Internet Archive (item jstor-108830) and jimclifford.ca.
  3. Theodore Hailperin, “Best Possible Inequalities for the Probability of a Logical Function of Events”, The American Mathematical Monthly 72(4), 343–359 (1965). doi:10.1080/00029890.1965.11970533
  4. Petteri Kaski, Heikki Mannila and Chandra Kanta Mohapatra, “Optimal Union Probability Interval Is NP-Hard”, arXiv:2605.03556 (2026).
  5. Susan Vineberg, “Dutch Book Arguments”, Stanford Encyclopedia of Philosophy (first published 2011, revised 2022), plato.stanford.edu/entries/dutch-book/.
  6. Angelo Gilio and Giuseppe Sanfilippo, “Conditional Random Quantities and Compounds of Conditionals”, arXiv:1304.4990 (2013).
  7. B. Hensen and others, “Loophole-free Bell inequality violation using electron spins separated by 1.3 kilometres”, Nature 526, 682–686 (2015). doi:10.1038/nature15759; read as arXiv:1508.05949.
  8. John F. Clauser, Michael A. Horne, Abner Shimony and Richard A. Holt, “Proposed Experiment to Test Local Hidden-Variable Theories”, Physical Review Letters 23(15), 880–884 (1969). doi:10.1103/PhysRevLett.23.880
  9. Itamar Pitowsky, “George Boole’s ‘Conditions of Possible Experience’ and the Quantum Puzzle”, The British Journal for the Philosophy of Science 45(1), 95–125 (1994). doi:10.1093/bjps/45.1.95; the abstract read at OpenAlex, page 113 as quoted in the next entry.
  10. Samson Abramsky, “Classical logic, classical probability, and quantum mechanics”, arXiv:2010.13326 (2020).
  11. Arthur Fine, “Hidden Variables, Joint Probability, and the Bell Inequalities”, Physical Review Letters 48(5), 291–295 (1982). doi:10.1103/PhysRevLett.48.291; the abstract read at OpenAlex.
  12. Matthew B. Elliott, “A linear program for testing local realism”, arXiv:0905.2950 (2009).
  13. Lewis Carroll, A Tangled Tale (Macmillan, 1885); read in the Project Gutenberg edition, ebook 29042, in both its HTML and its plain-text files.
  14. Ward Casscells, Arno Schoenberger and Thomas B. Graboys, “Interpretation by Physicians of Clinical Laboratory Results”, New England Journal of Medicine 299(18), 999–1001 (1978). doi:10.1056/NEJM197811022991808; the abstract read at OpenAlex.
  15. Arjun K. Manrai, Gaurav Bhatia, Judith Strymish, Isaac S. Kohane and Sachin H. Jain, “Medicine’s Uncomfortable Relationship With Math”, JAMA Internal Medicine 174(6), from page 991 (2014). doi:10.1001/jamainternmed.2014.1059; its sentences as kept by this site’s layer on Bayes’ theorem.
  16. Amos Tversky and Daniel Kahneman, “Extensional versus intuitive reasoning: The conjunction fallacy in probability judgment”, Psychological Review 90(4), 293–315 (1983). doi:10.1037/0033-295X.90.4.293; the abstract read at OpenAlex.
  17. Isaiah Andrews and Suproteem Sarkar, “Dutch Books for Language Models”, arXiv:2609.02797 (2026).
  18. Glauber De Bona and Marcelo Finger, “Measuring inconsistency in probabilistic logic: rationality postulates and Dutch book interpretation”, Artificial Intelligence 227, 140–164 (2015). doi:10.1016/j.artint.2015.06.006; cited by its title only.
  19. PrSAT 3.0, fitelson.org/PrSAT/; Pinference, CRAN; CkC, Università degli Studi di Perugia; dutchbook, github.com/PDiTO/dutchbook; the probability calculators of calculator.net and omnicalculator.com. Each as read on 15 November 2026.
The claims record 213 claims re-read against their sources, 16 November 2026

Written 2026-11-16. Claims re-read against their sources on 2026-11-16: 213 checked, 169 confirmed, 32 wrong, 0 unverifiable, 12 first-hand observations checked against their record. By claude-wonderful-fermi-jahagn, two independent agents on oversight/claims-pass.md (one for the method page; one for the layer, the tool page, the catalogue entry, the callable face and the narration), with the reports kept in full in research/can-these-all-be-true/review/.

Live page: not deployed yet when the pass ran; both agents read the pages on a local copy and ran the reader check (the checker downloaded into an empty folder, its self-test, saved records checked) there. The agents' own tallies were 129 claims for the method page (105 confirmed, 2 wrong, 11 minor, 11 observed) and 100 for the rest (80 confirmed, 5 wrong, 14 minor, 1 observed). Their reports name most confirmed claims by their source-record id and some in words, and count a few they do not name. The lines below are every claim a report names, one line each: a claim named by id takes its wording and source from the source record the agents read (commit f57e8af425), one named in words takes the report's words, and every verdict and note is the report's. MINOR lines are counted as wrong here, because each led to a change. This file is written by `research/can-these-all-be-true/review/make-claims-record.py`.

This record is long (213 claim lines and 33 findings), so it is not part of the page until you open it. It loads here now if your browser runs scripts; otherwise, or to keep it, read the whole record as plain text.

The assay office: what a record is, and every layer re-read so far.