Spurious Correlations and Fake Patterns: How Arithmetic Manufactures Proof

Introduction

In 1994, the journal Statistical Science published a paper reporting that the names and death dates of a list of famous rabbis were encoded in the book of Genesis, spelled out at fixed intervals — take every 12th letter, or every 50th, or every 477th, and a name appears. The statistics were not obviously broken. The journal was peer-reviewed. The authors were serious people, and the result was presented as a discovery about the text.

Five years later, four mathematicians re-ran the experiment. They found the same kind of result in a Hebrew translation of War and Peace, using the same protocol applied to a text nobody considered sacred [mckay-1999]. Whatever the experiment was measuring, it was not measuring the book of Genesis.

That episode is the engine of this guide. Its lesson is not "those people were stupid." The lesson is that arithmetic is not a detector of truth. Division manufactures precision from any two numbers. Exponentiation manufactures drama. A ratio between two measured quantities always exists, because measuring produces quantities — and any ratio you can compute, you can also name. Give the operation enough freedom and it will hand you a number that looks like a message.

Why does the number feel like proof? Three properties do most of the work. Precision: decimals read as measurement, and measurement reads as care. Reproducibility: anyone who runs the same steps gets the same answer, which feels like a law of nature but is only a property of arithmetic. Impersonality: the calculation does not appear to have an author, so it does not appear to have a motive. None of those three properties is evidence about the world. Together they are enough to make a coincidence look like a finding.

The rest of this guide is organized as a demolition manual rather than a sermon. First, the recipe: four levers that turn any measured fact into any number you like, worked end to end on the Great Pyramid. Second, the same recipe in the wild — numerology, gematria, Bible codes, sacred geometry — plus a live specimen you can operate yourself in about ten seconds. Third, why the output feels like discovery even to careful people. Fourth, a four-question test you can run in a conversation. Finally, the places where this argument stops working, because a guide that only teaches suspicion is just a different way of being wrong.

One note before starting: almost nobody in this story is lying. Numerology, gematria and sacred geometry were built by people who believed them, and most of them were doing something recognizably human — looking for order in a world that does not volunteer it. The recipe is worth learning precisely because it does not require a fraud to operate it.

The Recipe: Four Levers and One Target

Every fabricated numerical "discovery" follows the same skeleton. There is a real measurement, a conversion, an operation, and a target. Four knobs get turned, and one of them decides the answer in advance.

1. An arbitrary base. The measurement itself: a length, a date, a count, a name. It has to be real, or at least checkable, because that is what makes the final number feel grounded.
2. An exotic unit. The base gets re-expressed in a unit you choose — varas, cubits, furlongs, letters, "sacred inches", megahertz. Changing units changes the number without changing the world.
3. An accommodated operation. Multiply, divide, square, take a root, raise to a power, sum the digits. Each one produces a different number from identical inputs.
4. A target chosen in advance — or afterwards. Either you fix the number you want and search for operations that reach it, or you fix the operations and search for a number that feels meaningful. The freedom has to live at one end of the pipeline. It is never absent.

Here is the canonical example, assembled step by step. Every individual step is true, or close enough that arguing about it would miss the point.

Step 1 — a real measurement. The Great Pyramid of Giza has a base side of roughly 230.33 metres [giza-dimensions]. Different surveys give values between about 230.25 and 230.45 m, because the original casing stones are gone and the corners are eroded. Keep that range in mind; it will matter later.

Step 2 — an exotic unit. The vara is a Spanish length unit of about 0.8359 m, used in Spain and its American territories from the 16th century onward. It has no connection whatsoever to the Old Kingdom. Convert anyway:

Step 3 — a biological ruler. Call a dromedary's ordinary walking stride 1.96 m, which is 2.345 varas. Divide the pyramid by the camel:

Step 4 — a cosmological multiplier. The sun crosses 15 degrees of sky per hour, so in fifteen minutes it covers 3.75 degrees of arc. Multiply:

Step 5 — the announcement. Four hundred and forty. Which is the frequency of the A above middle C, in hertz. Which means the pyramid encodes concert pitch.

Stop and look at what just happened, because the arithmetic is flawless. 230.33 divided by 0.8359 is 275.5. Divide by 2.345 and you get 117.5. Times 3.75 gives 440.6. Nothing was miscalculated. The problem is not in the arithmetic; it is in the connection between the arithmetic and the claim.

Now the four levers, named:

Lever one, the base. 230.33 m is a modern metric average of a ruined structure. The builders did not know the metre. If you prefer 230.25 or 230.45, you get a slightly different answer, and you would pick whichever one lands better.

Lever two, the unit. The vara exists to convert. So does the cubit, the span, the furlong, the "pyramid inch". With five plausible units and a stride, you have five shots at a memorable number before you have done anything dishonest.

Lever three, the operation. Here is the part the published versions always omit: the operation is also a choice. Raise 117.5 to the power 3.75 and you get about 58 million. Divide it by 3.75 and you get 31. Add 3.75 and you get 121. The same three ingredients support at least four "discoveries", and only one of them resembles a musical note. The recipe reports the operation that worked.

Lever four, the target. In this example the target came first: 440 is famous, so the operations were tuned to reach it. In mystical systems the same freedom sits at the other end — the operations are fixed by tradition and the target is chosen after seeing what comes out. Either arrangement is the same trick with the receipt filed differently.

Now the detail that should end the conversation, and that the recipe cannot survive: hertz is cycles per second. The second is a modern unit, defined since 1967 by counting oscillations of a caesium atom. Concert pitch itself is an orchestral convention of the twentieth century — before that, "A" was whatever the local organ said it was. The pyramid was built around 2560 BCE, roughly 4,500 years before anyone could measure a frequency, in a civilisation with no electricity and no seconds. The number 440 is dimensionless: it is a ratio of a length to a length, multiplied by an angle in degrees. Attaching "megahertz" to it is not a measurement. It is a costume.

One more thing. Camel strides are not a constant of nature; they vary with the animal, the gait, the load and the sand. Run the band from 1.90 m to 1.98 m and the "discovery" sweeps from about 455 down to about 436. The target sits comfortably inside that band. So the claim is not even wrong in an interesting way — it is underdetermined, and the uncertainty was spent on making the number pretty.

The Same Recipe in the Wild

The Giza example is a demonstration built for this guide. The genre is older, larger, and in most cases entirely sincere. Four traditions are worth knowing, because they are the same machine with different paint.

Numerology maps letters to digits. The Pythagorean version assigns A=1 through I=9, then starts over at J=1, so the alphabet folds into nine buckets:

where i is the position of the letter in the alphabet. Sum the letters of a name, reduce the total to one digit, read the meaning. Three separate choices are hiding in that sentence: which alphabet (Latin, Hebrew, Cyrillic, and their transliterations give different sums), which reduction rule (digital root, master numbers kept, vowels only, consonants only), and which form of the name (full, short, maiden, married, stage). Nine possible outputs, dozens of possible inputs, and the reading is supplied by the person who recognises themselves in it.

Gematria is the older and more disciplined cousin, and it does not need a mapping table at all: in Hebrew and in Greek, letters are numbers. Aleph is 1, bet is 2, tav is 400. A word already has a value. The freedom moves elsewhere — into spelling (plene or defective), into which word boundaries you count, into whether you add, multiply or compose values, and into which verse you decide is the relevant one. Each of those decisions is another draw from the machine.

Bible codes industrialise the process. Instead of words, take letters at equal intervals: every n-th letter, in either direction, across the whole text. The search space is enormous — text length times skip range times direction times spelling variants — and "near" is a threshold somebody has to set. In the 1994 experiment, the list of rabbis, the spelling of their names, and the definition of proximity were all settled in ways that later turned out to be doing the work [mckay-1999]. The tell was not the arithmetic. It was that the protocol was tuned until the text cooperated.

Sacred geometry skips letters and goes straight to architecture. Take two lengths from a building, form a ratio, compare it to a constant you like: π, φ, √2, 2π, the fine-structure constant. Six candidate ratios, four units, five constants — that is a hundred and twenty comparisons before anyone has lied. Some will land within one percent, and the one that lands gets published. The Great Pyramid is a generous donor here, because its outer casing is gone: you may measure the base you can see or the base you infer, the height as built or the height as eroded, and both choices are defensible.

The modern version is not mysticism; it is a Tuesday in a research lab. The same flexibility — which outcome to report, which covariates to include, when to stop collecting data, which subgroup to highlight — produces "significant" findings from noise. Simmons, Nelson and Simonsohn demonstrated it by showing that a completely impossible hypothesis (that listening to a song makes you older) could be made to pass conventional significance tests through undisclosed analytical choices [simmons-2011], and Head and colleagues documented how widespread the practice is across scientific fields [head-2015]. No sacred text required. The lever is the same one.

Look at the last column. That is the whole taxonomy. In the mystical systems the operations are inherited and the target is discovered; in the modern and contrived versions the target is named first and the operations are reverse-engineered. Either way, one end of the pipeline is free, and the freedom is what generates the result.

A live exhibit. The Love Calculator on this site is the machine with the casing off, and it is worth two minutes because you can operate it yourself. It offers four methods for the same pair of names. Letter Match counts the unique letters two names share and divides by the unique letters in the longer name. Vowel Harmony compares vowel-to-letter ratios. Numerology applies the A=1 to I=9 mapping above, sums each name and reduces it. FLAMES reduces the pair to a count and uses that count to index a six-letter word: Friends, Lovers, Affectionate, Marriage, Enemies or Siblings. Romeo and Juliet score about 16.7% under Letter Match and about 70% under Vowel Harmony.

Same two names. Same calculator. Two very different verdicts, because the mapping is a design decision, not a measurement. That is precisely the gematria move: the letters are converted by a rule somebody chose, and the rule determines the answer. The difference is honesty — this calculator is labelled entertainment, and its own documentation says the algorithms are arbitrary heuristics. Most of the genre forgets to print that label. Use it as a specimen: try the pair of names you care about, then try two random strings of letters. A percentage and a message appear in both cases, delivered with exactly the same confidence.

Why a Fabricated Number Feels Like Discovery

If the recipe is this transparent, why does it work on intelligent people — including the people who build it?

The pattern detector is not optional. Human perception is tuned to find structure, and it fires on noise as readily as on signal. Psychology has a name for the disposition: apophenia, the tendency to perceive meaningful connections among unrelated phenomena, studied as a general bias toward false positives rather than as a defect of a few minds [blain-2020]. You cannot switch it off, and you would not want to: the same machinery that finds faces in clouds finds the actual patterns that keep us alive.

Confirmation closes the loop. Once a candidate number exists, attention moves to the confirming instances — the other measurements that "fit", the other words with the same value, the other verses that mention the same thing. The misses are not collected, because nobody enumerated the search. This is the same selective accounting that makes gambling systems feel reliable: the wins are memorable and the losses are explained away.

Multiple comparisons do the arithmetic. This is the part that can be counted, so count it. If you test one hypothesis at the conventional α = 0.05, you have a 5% chance of a false positive. Test several independent hypotheses and the chance that at least one looks significant climbs fast:

Twenty comparisons is an afternoon with a spreadsheet. A sacred-geometry survey, a gematria reading that tries four spellings, and a Bible-code search over a few thousand skip distances are all running dozens or thousands of comparisons. None of them reports the denominator, and the denominator is the whole story.

What a p-value is not matters here too. The American Statistical Association's statement is explicit that a p-value does not measure the probability that a hypothesis is true, nor the size or the importance of an effect [asa-2016]. It measures how surprising the data would be if a specific model were true. Reversing that conditional — "the data are surprising, therefore the model is false, therefore my story is true" — is the single most common error in popular statistics, and it is the exact error the recipe depends on.

Scale makes it worse. Ioannidis showed that when studies are small, effects are small, the analytical flexibility is large, and many teams chase the same questions, most published research findings will be false [ioannidis-2005]. That is not a claim about the honesty of scientists; it is a statement about what happens when selection operates on noisy measurements at scale. The mystical version of the same process — many texts, many measurements, many interpreters, one memorable hit — has been running for centuries.

And the arithmetic is genuinely reproducible. This is the last piece of the illusion, and the strongest. Run the Giza steps again tomorrow and you get 440.6 again. That stability feels like corroboration. But reproducibility is a property of division, not of pyramids: any two numbers you can measure will keep producing the same quotient forever, and that tells you nothing about whether the quotient means anything. Diaconis and Mosteller made the point precisely in their work on coincidences: the question is never "how improbable is this event", it is "how improbable is this class of events" — because the class is what you were really sampling from [diaconis-mosteller-1989].

How to Test a Claim: Four Questions

The checklist is deliberately short enough to use while someone is still talking. It works on a pyramid ratio, a numerology reading, a headline about breakfast, and a dashboard that discovered a trend.

1. Was the target chosen before the search? Ask the counterfactual: would the claim's author have been equally impressed by 441? By 438? By 0.44? If the answer is no, then the target did the work, and the arithmetic was a search, not a test. Preregistration is the formal version of this question, and its absence is not a moral failing — it just means the finding is a hypothesis, not a result.

2. Is the unit the natural unit of the phenomenon? Would the people who built the thing have used that unit, and would they have had that concept at all? A vara on the Giza plateau fails this test before breakfast; hertz fails it permanently. The test also catches the reverse error: a unit that is unusual but intrinsic — natural units in physics, for instance — is not automatically suspect. Unconventional is not the same as arbitrary.

3. How many combinations were tried? Ask for the denominator, plainly: how many ratios, units, spellings, subgroups, outcomes, or stopping points did you look at before this one? If the answer is "we didn't count", assume at least twenty and read the table above. An honest large denominator is not disqualifying — a finding can survive a wide search if it is then tested on fresh data. Concealing the denominator is what turns exploration into fabrication.

4. Does it need rounding, truncation, or a favourable variant to land? Recompute the band. If the underlying measurement is uncertain by a few percent and the hit survives only at one edge of that range, the hit belongs to the input, not to the object. 440.6 is not 440; it is 440.6, and the distance between them is where the choice lives.

Two extensions are worth carrying around. First, the negative control: apply the exact same protocol to a text, building or dataset where the effect should be absent. McKay's team did this with War and Peace and the code followed them there [mckay-1999]. If your method finds the pattern everywhere, it is not finding the pattern. Second, ask whether the claim is a derivation or an observation. A proof is valid or invalid on its own terms; a numerical claim about the world has to survive contact with the world. Confusing the two is how "the math checks out" becomes a substitute for "the evidence checks out".

A claim that survives all four questions is not proven. It is merely worth the next test, which is the most any honest claim can offer.

Limitations: Where This Argument Does Not Apply

A demolition manual that leaves you unable to believe anything is not an improvement on credulity. Four boundaries matter.

Genuine correlations exist, and they survive the checklist. Smoking and lung cancer, CO₂ and global temperature, vaccination and antibody titres — these are not numerology with better branding. What separates them is not the sophistication of the arithmetic but the structure of the evidence: effects measured on prespecified outcomes, replicated out of sample, robust to reasonable changes in units and model specification, and predicted in advance. The recipe in this guide describes claims that were never subjected to any of that. It says nothing about claims that were.

Exploration is legitimate. Most real discoveries begin with someone noticing a pattern that was not the point of the study. Looking at data to generate a hypothesis is not p-hacking; publishing it as though it were a confirmed finding is. The remedy is a fresh test on fresh data, not abstention.

Mathematics is not arbitrary. Within mathematics, a proof is not a pattern claim: it is checked against axioms and rules, and it is either valid or it is not. What is arbitrary is the bridge between a number and a fact about the world — the choice of unit, the choice of measurement, the choice of what to compare. Confusing the two produces both errors at once: credulity toward numbers that were chosen to fit, and nihilism toward mathematics itself, as if the Pythagorean theorem were a matter of opinion.

Improbable events really do happen. The birthday paradox is not a trick of units or a rounding artifact; it is a correct probability about a class of events, and it is why shared birthdays in a room of twenty-three people are ordinary. The lesson of the recipe is not that improbable things are impossible. It is that the probability has to be computed for the event class you actually sampled, and that you have to count what you looked at [diaconis-mosteller-1989].

Nothing on this page is a probability error, and separating it from one is the whole point. A coincidence is an event that happened and was noticed afterwards, so its odds can be computed honestly - that is what the It's Not a Coincidence guide does with the birthday paradox, the law of truly large numbers and Littlewood's law. Here the odds were never the issue. The target was chosen first, and the derivation was assembled to reach it. If the number survives the question "how many opportunities were there?" you are dealing with a coincidence; if it only survives "was the target chosen before or after the search?" you are dealing with a fabrication.

Two more honest caveats, specific to this page. First, the pyramid example is constructed here, not quoted from a published author: the measurements are real, the unit is real, the solar arc is real, but the assembly is ours, and it exists to show the method rather than to accuse anyone. If you meet a version of it in the wild with a straight face and a citation, the four questions will handle it. Second, the checklist is a filter, not a court. It will reject some true claims that happen to be poorly presented, and it can be gamed by anyone willing to run a fresh preregistered study — which is exactly the behaviour we want to reward, and the reason the filter is worth applying anyway.

The line worth drawing is consequences. A love calculator that produces a number for two names is a party game, and its own documentation says so; the machinery is visible, the stakes are zero, and the amusement is real. The same machinery in a hiring decision, a medical choice, a loan application or a relationship ultimatum is a different object entirely. Arbitrary arithmetic is fine as entertainment. It is not fine as a basis for decisions that change someone's life.

Frequently Asked Questions

❓ Does the Great Pyramid really encode the note A at 440 hertz?

✅ No. The example in this guide is constructed to demonstrate a method, not quoted from a researcher. It uses a real base length, a real historical unit and a real solar arc, then multiplies them together until the number 440 appears. Hertz measures cycles per second, a unit defined in the twentieth century, so it cannot describe a building from 2560 BCE.

❓ What is a spurious correlation?

✅ A statistical relationship that appears in data but does not reflect a causal connection between the two things measured. It can arise from chance, from a shared third factor, or from the way the data were selected and grouped. With enough variables, some pairs will always line up.

❓ Is numerology always wrong?

✅ As a factual claim about the world, yes: letter-to-number mappings are chosen by people, so the results are design decisions rather than measurements. As a cultural practice and a source of entertainment, it is not a factual claim at all. The problem begins when a mapping starts influencing decisions about money, health, hiring or relationships.

❓ What is the difference between numerology and gematria?

✅ Numerology usually assigns values to letters through a table, such as A=1 through I=9 in the Pythagorean system. Gematria works with alphabets where letters already carry numeric values, as in Hebrew and Greek, so a word has a value without any conversion table. Both then rely on the interpreter's choices about spelling, boundaries and operations.

❓ Why do so many measurements come out close to the golden ratio?

✅ Because a ratio of two lengths you are free to choose can be steered toward almost any constant, and because with five or six candidate constants and several ways to measure each length, one comparison will land close by chance. Getting within one percent is easy when you have a hundred attempts and report only the successful one.

❓ What is p-hacking in one sentence?

✅ P-hacking is the practice of trying multiple analyses, subgroups or stopping points and reporting only the one that crosses a significance threshold, which converts a search into a fake confirmation.

❓ Does a p-value tell me the probability that my hypothesis is true?

✅ No. The American Statistical Association states explicitly that a p-value is not the probability that the hypothesis is true, nor a measure of the size or importance of an effect. It describes how surprising the data would be if a specified model were correct.

❓ If the math is correct, how can the conclusion be wrong?

✅ Because correct arithmetic only guarantees that the output follows from the inputs. If the unit was chosen to fit, the target was named after the search, or the denominator was never counted, every step can be flawless and the conclusion still says nothing about the world.

❓ What is apophenia?

✅ Apophenia is the tendency to perceive meaningful connections among unrelated phenomena, including patterns in noise. It is best understood as a general bias toward false positives in pattern detection rather than a rare disorder, and it is the perceptual half of what makes fabricated numerology feel convincing.

❓ Can the Bible code be tested scientifically?

✅ Yes, and it was. Researchers extracted letter sequences at fixed intervals from Genesis and compared hit rates against control texts. When the protocol was applied to a Hebrew translation of War and Peace, comparable results appeared, which indicates the effect came from the flexibility of the method rather than from the text.

❓ How do I check a claim a friend sends me?

✅ Ask four questions: was the target chosen before the search, is the unit natural to the phenomenon, how many combinations were tried, and does it need rounding or a favourable variant to land. If the answers are vague, the claim is a hypothesis rather than a finding.

❓ Isn't math supposed to be objective?

✅ Mathematics is objective about what follows from what. It is silent about whether the quantities you fed it were the right ones, whether the units mean anything, or whether the comparison was chosen in advance. Objectivity in the calculation does not transfer to the story built on top of it.