Method

The short version

1. Name the rivals.

   Six accounts of the same evidence: naturalism, deism or generic theism, other traditions, Judaism, Islam and Christianity.

   The six ↓
2. Fix the rules before looking.

   What counts as a hit, a miss or a fulfilment is set before scoring. Misses are counted. Nothing weak is quietly dropped.

   The rules ↓
3. Weigh each line as a likelihood ratio.

   How much more expected is this evidence if one account is true than if another is? Given as a range, with the critic's value as the low end.

   The arithmetic ↓
4. Combine honestly.

   Overlapping evidence is discounted, every line is capped by the chance its analysis is wrong, and the evidence against goes in the same ledger.

   Dependence and caps ↓
5. Show the result for every starting point.

   No single "probability that God exists". Instead: how far the evidence moves you, and how skeptical you would have to start for it not to move you past even odds.

   Priors and break-evens ↓

The six hypotheses

A test needs rivals. "God or chance" is a false choice: chance is not a worldview, and every serious alternative has to be named so the evidence can be weighed against it.

The site's three tiers are questions about groups of these. Tier I asks whether there is a Mind at all (N against the other five). Tier II asks whether that Mind is the God of Israel who speaks through Scripture (J, I and C against D and O). Tier III asks whether Jesus is who he claimed (C against J and I). Each line of evidence is scored only for the pairs it can actually tell apart: fine-tuning, for example, says nothing about which God, and the resurrection evidence is where Christianity and its nearest rivals part ways.

Rules of evidence

Each content page states its own rules at the top, before the data. These rules apply everywhere:

• Fixed before scoring. The hypotheses, which evidence bears on which of them, how lines are grouped, the starting-point presets and the error caps were set before any area's numbers were plugged in.
• The whole text, not the highlights. Where a claim is about a body of text (prophecy, science in Scripture, Genesis), every item is scored under rules fixed in advance: hits, misses and the ambiguous middle. The set of items is defined by a source that is not ours where possible, for example the features pre-Christian Jews expected of the Messiah, not a Christian list of fulfilments.
• Blind where it matters. The messianic-prophecy test was rerun with the list of expected features extracted and coded by separate AI agents who were not told its purpose, under a protocol fixed in advance (a human replication has not yet been done). On that list, scoring the hits alone gives about a trillion to one; scoring everything gives about 0.3. That gap is the size of cherry-picking, measured. The prophecy pages use the whole-record figure.
• Controls. A method that finds patterns must be run on cases where there should be none. Typology is tested on Lincoln, Muhammad and random figures; Genesis 1 is scored with the same rubric as fifteen other ancient creation accounts; Jesus is scored against other messianic claimants.
• No silent drops. A line that turns out weak stays in the ledger with a likelihood ratio near 1. Dropping weak lines is exactly the cherry-picking this site exists to avoid.
• The evidence against is evidence. Suffering, divine hiddenness, the hard texts of the Bible and the rest are scored by the same rules and sit in the same table (the case against). The evidence offered for other traditions is weighed too (rival traditions).
• Every number has a source or a calculation. Probabilities, counts and ranges come from cited work or from a script whose output is shown. Popular figures that fail this test are listed on each page under "What we don't claim".

Likelihood ratios, in plain words

The question for each line of evidence is not "is this improbable?" but "is this more expected if one account is true than if the other is?" That ratio is the likelihood ratio (LR):

LR = 1      the evidence doesn't help choose
LR = 10     the evidence is 10 times more expected under A
LR = 1/10   10 times more expected under B (it counts against A)

Small improbability alone proves nothing. If an observation is just as unlikely under both accounts, the ratio is 1. Both halves are always needed, the probability under each rival; the site states both. For combining many lines it is easier to add than to multiply, so the ledgers use bans, the logarithm of the ratio in base 10: +1 ban is a factor of 10, −0.5 ban a factor of about 1/3. The unit goes back to Alan Turing and I. J. Good's wartime codebreaking and to Good's "weight of evidence".¹

Rivals that come in several forms. Naturalism is not one hypothesis: it can be a single universe with brute-fact constants, a deeper law not yet found, or a multiverse with observer selection. A line's ratio against naturalism is limited by the form that fits the evidence best. Fine-tuning is the clearest case: its famous odds, about 1 in 10⁵¹, count only against one universe of blind, uncorrelated chance, while a cancelling law or a multiverse with selection would make the observation expected. So its ratio here is about 10, set by how much weight one gives those forms, not by the size of the tuning. And one fact is counted once: the same tuning can be stated as odds or as about 169 bits, and those are one piece of evidence, not two.

Ranges, not single numbers. Each line has three values: the low end, which a well-informed critic would defend; the high end, which the strongest scholarly defender would defend; and the central value, the research's own judgement. The strength badge on each page grades the line as evidence for its hypothesis, not the quality of the underlying facts (the Old Testament's history of the kings is well confirmed, but accurate history is about as expected from human authors, so as evidence for God it is weak). The badge is the research's overall judgement, guided by the central value and pulled down when the critic's value is near 1. Roughly:

Combining the lines: dependence and error caps

Dependence

Adding up the ratios of separate lines is only right if each line is independent of the others given each account. Often it isn't: the Gospels' reliability and the resurrection evidence share sources; the beginning of the universe, fine-tuning and the origin of life share one failure mode ("future science will close the gap"). So lines are grouped into clusters. Within a cluster the strongest line counts in full and every other counts at a reduced weight w, from 0 (only the strongest counts) to 1 (all independent). The idea is the survey statistician's "effective sample size": correlated observations are worth fewer independent ones.³ It is a heuristic, applied the same way to the evidence against. The verdict page also reports the check that was meant as the worst case: every tier treated as a single argument, with w = 0. Because the same discount falls on the evidence against, and on the two prophecy lines whose values are now below 1, that check now raises the result for Christianity; counting the evidence against twice, or dropping Tier I, are the checks that bite.

The error cap

A likelihood ratio of a million claims that the analysis behind it is reliable to better than one in a million. Almost nothing in history, philosophy or cosmology is. So every line is mixed with a small chance ε that its analysis is simply wrong, which caps how far one line can move the odds: by about 20× at ε = 1 in 20, 100× at 1 in 100, 1,000× at 1 in 1,000.⁴ This is what keeps numbers like "1 in 10¹²⁰" or "10⁴⁴" out of the ledger: they measure a model's internal arithmetic, not anyone's justified confidence in the model.

effective LR = LR / ((1 − ε) + ε·LR)   for LR ≥ 1   (never above 1/ε)
effective LR = (1 − ε)·LR + ε          for LR < 1   (never below ε)
log-odds of each account = log prior + Σ over clusters ( strongest line + w × each other line )

Priors, break-evens and "extraordinary claims"

Where you end up depends on where you start, and there is no agreed objective starting point for theism. Swinburne argues theism is simple and so not improbable to begin with; Draper and Oppy argue naturalism is the more modest hypothesis.⁵ So the site never states "the probability that God exists". It shows three starting points (a skeptic giving any Mind behind nature 1 in 1,000, a neutral reader at even odds, and a generous one at 80% who still favours no religion) and a break-even for each: how skeptical you would have to start for the evidence not to bring you past even odds. You can then put in your own.

Hume's maxim, that no testimony suffices for a miracle "unless the testimony be of such a kind, that its falsehood would be more miraculous, than the fact, which it endeavours to establish", is in odds form exactly this rule: the evidence must outweigh the odds against.⁶ The site accepts it and puts it in the open as the prior. What it does not do is set the prior for the resurrection from "dead men don't rise". Nobody disputes that frequency; it is the probability of rising by natural causes. The hypothesis under test is that God raised Jesus, and its prior depends on whether there is a God, which is what Tiers I and II are about. The joint model handles this automatically, and it also respects the limit this puts on what the resurrection can say about God (below).

Why our numbers are smaller than other sites'

Readers who know apologetics will notice that this site's figures are modest. That is deliberate. The famous large numbers fail for reasons a statistician would spot at once:

• Multiplying chosen hits. Peter Stoner's "1 in 10¹⁷" for eight messianic prophecies multiplies guessed factors as if independent, selects the hits after the fact, counts no misses and compares against "chance" instead of the real rivals (a man who knew the texts; authors who shaped the story). We recompute it on the messianic prophecy page.
• Invented inputs. Stephen Unwin's The Probability of God reached 67% by multiplying self-assigned factors; the arithmetic is fine and the inputs are made up.⁷ Bayes' theorem transmits the information in its inputs and creates none.
• Independence assumed. In the Sally Clark case a court heard that two cot deaths had odds of "1 in 73 million", obtained by squaring 1 in 8,543. The Royal Statistical Society said the calculation was "false" because there were "very strong reasons for supposing that the assumption [of independence] is false".⁸ The McGrews' careful Bayesian case for the resurrection reaches 10⁴⁴ mostly by treating many witnesses as independent; Brian Blais shows that under partial dependence "the tiniest deviation from independence results in the argument completely collapsing".⁹ We cite their argument for its structure, not its number.
• Huge factors on both sides. After the Rootclaim debate on the origin of COVID-19, Scott Alexander noted that multi-step probability estimates fail "because people don't think enough about the correlations between stages".¹⁰ If both the case for and the case against carry enormous factors, something is mis-specified. The caps above exist to prevent it.

Smaller numbers that survive a critic are worth more than large ones that don't.

Philosophy, the resurrection and this model

The research program behind this site also graded the classic philosophical arguments (from contingency, from the beginning, from fine-tuning, from mind) by how strong a proof each one is, from demonstration down to contested. Its verdict is modest. What is established without a contested principle is negative or conditional. The strongest results rule out particular forms of naturalism: a single universe of blind chance for the cosmological constant, the early universe as a fluctuation, a first state that explains itself. Many naturalists already reject those forms. Every step from there to God is contested, and the conclusion it can support is a generic theism that does not choose among Judaism, Christianity and Islam, or against deism.¹¹

Those grades are not likelihood ratios, and the two are never added. They decide which forms of naturalism the evidence on this site has to beat, which is why fine-tuning counts about 10 here and the beginning about 1.3. The argument from contingency is not scored as evidence at all: its force is your confidence in the principle that every contingent fact has an explanation, which is a starting point, and counting it as evidence would count it twice. Where the philosophy leaves off (which God, if any), Tiers II and III take over.

The same research proved a short limit on what historical evidence can say about God. However strong the evidence that an event happened (call that strength L), it can raise the odds of a God who acts, over a God who never acts or over naturalism, by at most 1 + P(God would do this) × (L − 1), so never by more than L.¹² The model respects this automatically. Its resurrection line favours Christianity alone, so its effect on the odds that there is a Mind at all depends on Christianity's share among the theistic options: from even odds it raises those odds about fourfold, from the skeptic's start about 1.35 times.¹³ That share is a judgement about God's purposes. If you think God unlikely to raise this man, the place to say so is Christianity's starting share, not the resurrection's ratio. The skeptic's starting point halves that share; the calculator does not yet let you set the shares one by one, but the published model does.

How this site was made

The research and drafting were done with AI research assistants (Claude, by Anthropic) working under the author's direction, and every claim is tied to a source or a published calculation you can check. Each content area began as a research dossier that gathered the primary literature on both sides, scored the evidence under the rules above, ran its calculations as scripts, and listed the claims that fail. The pages were written from those dossiers and then checked again against the sources. Where a writer found an error in a dossier, the page uses the corrected figure and says so in its source. Sources the research could only reach through other works are marked as such and are never quoted directly.

The site grew out of an earlier project of the author's that made the same case less carefully. Its strongest material was kept, its errors were dropped, and every claim was re-researched from the ground up. If you find a mistake, a missing source or a line of evidence we left out, tell us: it goes in the ledger.

What we don't claim

• "The odds against chance are 1 in 10^X." Chance is not the rival hypothesis, and no single line can honestly carry such a factor (see the error cap).
• "Proof", "proven", "falsified". Every argument here is probabilistic, and even the deductive steps rest on contested premises.
• "The probability that God exists is X%." A probability needs a starting point, and the starting point is yours. We give likelihood ratios and break-evens.
• "Bayes' theorem proves …" It only passes on the information in its inputs.
• "Each step builds on the last, so the case is a chain." A chain is as weak as its weakest link. The case is cumulative: evidence bears on several tiers at once.
• "Ruling out chance proves design." It rules out one form of naturalism; a law or a multiverse could also do the selecting.
• "The philosophical arguments prove the God of the Bible." On the research behind this site, the best of them support at most a generic theism, and every step to it is contested.
• "The resurrection proves God." It is evidence for the event; what it adds for God is limited by how likely God was to do it.
• "No other account has any viability." On the critics' numbers naturalism leads from every starting point the verdict page shows, and from the skeptic's start it keeps 99.54% even on the central numbers.

Sources

1. Good, I. J. (1950). Probability and the Weighing of Evidence. London: Griffin. (Not checked directly.)
2. Jeffreys, H. (1961). Theory of Probability, 3rd ed. Oxford: Clarendon Press. (Not checked directly.)
3. Kish, L. (1965). Survey Sampling. New York: Wiley. (Not checked directly.)
4. Ord, T., Hillerbrand, R. & Sandberg, A. (2010). "Probing the improbable: methodological challenges for risks with low probabilities and high stakes." Journal of Risk Research 13(2): 191–205. (Not checked directly.)
5. Swinburne, R. (2004). The Existence of God, 2nd ed. Oxford: OUP; Draper, P. (1989). "Pain and Pleasure: An Evidential Problem for Theists." Noûs 23(3): 331–350; Oppy, G. (2006). Arguing about Gods. Cambridge UP. (Not checked directly.)
6. Hume, D. (1748). An Enquiry concerning Human Understanding, §10, Part 1; quoted from McGrew, T. & Larmer, R. "Miracles." Stanford Encyclopedia of Philosophy (rev. 7 May 2024). plato.stanford.edu/entries/miracles
7. Unwin, S. D. (2003). The Probability of God. New York: Crown Forum; Shermer, M. (2004). "God's Number Is Up." Scientific American, July. michaelshermer.com
8. "Sally Clark." Wikipedia, accessed 12 Nov 2026, quoting the Royal Statistical Society's statements of October 2001 and January 2002. en.wikipedia.org/wiki/Sally_Clark
9. McGrew, T. & McGrew, L. (2009). "The Argument from Miracles: A Cumulative Case for the Resurrection of Jesus of Nazareth." In W. L. Craig & J. P. Moreland (eds.), The Blackwell Companion to Natural Theology, 593–662. Wiley-Blackwell; Blais, B. (2023). "Probability and the Independence of Testimony." 23 Feb. bblais.github.io
10. Alexander, S. (2024). "Practically-A-Book Review: Rootclaim $100,000 Lab Leak Debate." Astral Codex Ten, 28 Mar. astralcodexten.com
11. The research program behind this site (omega.logos.so), graded records: the verdict and its classes, omega/research/ARCHITECTURE.md §1 and §4; how fine-tuning bears on naturalism's forms, omega/research/dev/finetuning/RECORD.md and omega/research/boxes/BOXES_VERDICT.md.
12. The agency limit ("C3-cap", a proved conditional): omega/research/ceiling/c3-agency/RECORD.md §3, with its historical figures in §4 and the wording for this site in §9-10.
13. The model's own figures for the resurrection's effect on the odds of a Mind: truth/research/calc/14-lt-checks.py and its saved output (section B); see truth/research/14-integration.md §2.11.