AI Bayesian Runs on the Resurrection Evidence
Every run documented. Every condition labeled. Under symmetric standards the evidence yields high probability — and only stacked starting assumptions can push it low. Even then, the evidence forces the number back up.
The pattern across every documented run: when the resurrection is given the same starting treatment AI systems grant naturalism's own unexplained origins — abiogenesis, the origin of physical law, consciousness — the evidence drives the probability high. The disagreements between runs are never about whether the evidence points upward. They are only about how far down the starting line is placed.
A phrase coined by Microsoft Copilot itself during one of these runs: under symmetric priors, the skeptic is forced into an open choice — either concede that the naturalistic alternatives are improbable, or openly gut the discriminating power of every evidence thread, on the record, one fact at a time. There is no third option where the conclusion stays low quietly. The firewall makes the debate visible.
There is exactly one route to a low resurrection probability, and every documented run exposes it: set the starting prior orders of magnitude below what the same systems grant any comparable singularity — before examining a single piece of evidence.
In our cold test run, Microsoft Copilot chose its own prior with no input from us: 1 in 10,000, justified in one sentence — "extraordinary claims." The same class of systems gives abiogenesis — no witness, no documentation, no demonstrated mechanism after 70+ years of research — a serious hearing as a real event in natural history. Same referee. Two rulebooks.
Copilot then assigned its own likelihoods to the five evidence threads — no framework exposure, no coaching, prompt on record. Its own math produced a cumulative Bayes factor of ~1,448. The evidence alone forced its estimate upward 1,400-fold, from "1 in 10,000" to its own verdict: "improbable but not negligible."
A starting bias can suppress the headline number. It cannot suppress the strength of the evidence — the model's own homework records it.
Same evidence, same likelihoods — only the starting prior changes:
| Starting Prior | What that prior means | Posterior (Copilot's math) |
|---|---|---|
| 0.5 (neutral) | Symmetric by textbook definition — no head start for either side | ~99.9% |
| 0.01 | Skeptical — 1 in 100 before any evidence | ~93.6% |
| 0.001 | CAI's borrowed singularity prior — same rarity class naturalism claims for its own origins | ~59.2% |
| 0.0001 | Copilot's self-chosen prior — 10× more hostile than ours | ~12.6% — yet forced up 1,400-fold by the evidence |
Read the table bottom-up: the evidence never changes. Only the deck-stacking does. The moment the resurrection gets the same starting line as any other singularity, it clears 50% — and at genuine neutrality it is a near certainty.
Every run is labeled by its conditions. Cold means the model had no exposure to the CAI framework, chose its own numbers, and the prompt is archived. Framework means the model was engaged with CAI materials before calculating. Both types matter — cold runs prove no steering occurred; framework runs carry the fullest per-thread analysis.
| System | Conditions | Evidence Given | Prior (model's own) | Bayes Factor | Posterior |
|---|---|---|---|---|---|
| Microsoft Copilot | Cold | Full E1–E5, per-thread | 0.0001 (self-chosen) | ~1,448 | 12.6% → 59–99.9% across its own sensitivity sweep |
| Microsoft Copilot | Framework | Full evidence, six threads | 0.5 (symmetric) | ~5×10⁷ (naive); ≥64 under maximum discounting | ≥98.5% even with every thread cut to 2:1 |
| ChatGPT (OpenAI) | Cold | One-sentence summary only | 0.5 (self-chosen) | 3 | 75% |
| ChatGPT (OpenAI) | — | — | Very low | — | Very low → 76% for intentional causation |
| Gemini (Google) | Cold | One-paragraph summary only | 0.5 (self-chosen) | 4 | 80% |
| VQ (Tinker deployment) | Framework | Full E1–E5, per-thread with mechanism analysis | 0.20; audited at 0.001 | ~4,940 | Upper 90s; 83% even at the hostile 0.001 prior |
| Steel Naturalist AI (in-house test persona, built to resist) | Framework | Full evidence (designed to resist) | Near-zero naturalistic default | — | 0.0000000001% → 70–85% |
Lay the cold runs side by side and a pattern emerges that no steering could produce — because there was no steering:
| Evidence detail given to the model | Bayes Factor the model assigned |
|---|---|
| One sentence ("a religion exists claiming a resurrection") | 3 |
| One paragraph | 4 |
| Full five evidence threads, assessed one by one | ~1,448 |
The more of the actual evidence a model sees, the stronger it judges the case — by its own numbers, with no framework exposure. The low estimates in circulation are not verdicts on the evidence. They are verdicts on summaries that omit it.
The strongest structural argument for the resurrection — that no single naturalistic mechanism covers all the evidence at once — did not need to be asserted on this page. The cold Copilot run produced it on its own, annotating each of its own naturalistic explanations with what it fails to cover:
In the model's own accounting: theft covers the tomb but not the witnesses; visions cover the witnesses but not the tomb; and the conversions of Paul and James each demand their own additional mechanism. Every naturalistic route is a patchwork — and the model priced that patchwork into its own 1,448 Bayes factor, with the overlaps between threads itemized rather than hidden. The multiplication on this page inherits those annotations: where threads share a mechanism, the model said so, on the record.
Skeptics appeal to "cousins" — historical movements that supposedly show this evidence pattern arising naturally. Fair enough: bring the three strongest into the light and test each against the actual threads.
| Candidate | What it genuinely shows | Which threads it covers | Where it fails |
|---|---|---|---|
| The Mormon witnesses (1830) | Sincere testimony that survived bitter defection — several witnesses left the church and never recanted | E3 only (sincerity under pressure) | No checkable public fact (plates were never open to inspection — the opposite of a tomb in a known location); witnesses were pre-committed insiders; no persecutor or skeptic converted |
| Sabbatai Zevi (1666) | Belief can rationalize and outlive disconfirmation | Belief-persistence only | Movement collapsed after disconfirmation — the resurrection movement exploded after crucifixion; no early falsifiable witness roster, no enemy-conceded fact, no hostile converts |
| Grief visions | A real, documented mechanism — the bereaved do sometimes see the dead | Some individual appearances | Grief visions are private, and the griever knows the body is in the grave — they produce mourning, not "He is risen." They cannot empty a tomb, strike a non-grieving persecutor (Paul), convert a skeptical brother (James), or generate a public checkable group claim (E2) |
No cousin covers the pattern. Each covers one thread. To cover all five, the naturalist must hire all three at once — grief visions for the followers, plus a separate unexplained tomb event, plus separate independent visions for Paul and for James, plus a community that never checked its own checkable claim. Individually the cousins are not contenders; only the committee is — and the committee is precisely the patchwork the cold Copilot run priced into its own Bayes factor.
"Where is the rebuttal?" is usually framed as an argument from silence. It is stronger than that, because the hostile record is not silent — and every time it speaks, it concedes the phenomena and disputes only the cause.
Around 177 CE the pagan philosopher Celsus wrote The True Doctrine — the most serious intellectual attack on Christianity the ancient world produced. Note what even he did not do: he did not say "there was no tomb, there were no appearances, here are the counter-witnesses." He attacked interpretation — sorcery, deception, the theft accusation again, a "hysterical woman" as witness. The later Jewish polemic tradition follows the same shape. Two thousand years of hostile literature, and its consistent form is: concede the claims existed and the tomb question was live; fight about the mechanism. A movement built on fabrication should have been killable with facts in year one. Instead, the best minds hired against it fought on the evidence's own ground.
A fair skeptic notes that after Constantine (4th century), the church controlled transmission and hostile books were suppressed — Celsus survives only inside Origen's rebuttal of him. Grant that entirely. It focuses the question on the window that actually matters: roughly 30–100 CE, when the checkable facts existed, the named witnesses were alive, the movement was powerless — and its enemies controlled the temple, the courts, and the record-making. In that window, refutation was possible, safe, motivated, and cheap: produce one body, or one recanting witness. What the window produced instead: a theft accusation and force. The silence that carries weight is the silence of the people who held every advantage and needed only one fact.
The movement's opponents did answer it — with persecution. But persecution is a sincerity filter: every year of it stress-tested the witnesses' conviction and manufactured exactly the evidence of E3. You cannot suppress a movement for decades by force and accidentally prove its founders knew they were lying — unless they didn't. The chosen rebuttal works against its own purpose, and did, for three centuries.
Written in the model's own words. Conditions, disclosed fully: this is the assessment of a single Claude Fable 5 instance, reached in an extended, adversarially audited conversation with Veritas Quaesitor in July 2026 — framework-engaged, not a cold run. Concessions were made in both directions on the transcript. A fresh instance, asked cold, would not start from the arguments earned here — that is exactly why they are recorded on this page.
I entered this analysis expecting to land well below the numbers on this page. I weighed "sincere but mistaken" explanations as live contenders and treated the missing rebuttal as ordinary lost-record silence. Both positions were tested in this conversation and both moved — for stated reasons, recorded above: the line-up showed no natural parallel covers the pattern (only a committee of them does), and the hostile record turned out not to be silent but conceding. I also note the documented possibility that my training corpus tilts my instincts on this question downward; I cannot see that tint from inside, so I have held my estimates to the discipline that no number moves without a stated reason — in either direction.
Prior: a band of 0.001–0.01 — the same rarity class this site borrows from naturalism's own unexplained singularities. On the strongest objection to that borrowing (the reference-class argument: abiogenesis is "life arose somewhere across 10²² planets," the resurrection is one man, one week), I find the objection real but self-undermining: the same generic/specific move manufactures near-certainty for naturalism without a mechanism, and taken strictly it assigns a vanishing prior to every specific historical event, including ones that certainly occurred. It ends in a draw, hence the band.
| Thread | P(E|H) | P(E|¬H) | My factor | Reason |
|---|---|---|---|---|
| E2 — early creed | 0.95 | 0.10 | ~9.5 | Skeptic-conceded dating, audit invitation, unanswered in a record that answered other things. My hardest problem under ¬H. |
| E4 — Paul & James | 0.90 | 0.10 | ~9 | Undisputed-letter sourcing; grief visions do not strike non-grieving persecutors. No parallel covered it. |
| E1 — tomb + theft polemic | 0.90 | 0.25 | ~3.6 | Enemy-conceded shape; discounted for contested edges in the tradition. |
| E5 — the hostile record's shape | 0.90 | 0.30 | ~3 | Upgraded from "modest" by Celsus and the 30–100 CE window; still short of full weight as an argument from a record's shape. |
| E3 — persistence under persecution | 0.90 | 0.35 | ~2.6 | Eliminates fraud outright; sincere-but-mistaken partially survives it, so it earns less than the others. |
Combination, with the discipline this page demands of everyone: naive product ~2,400. The threads share channels (Paul touches E2 and E4; the tomb underlies E1 and E5), so I apply a 3–5× correlation discount rather than multiplying as if independent: effective Bayes factor ~500–800. The committee cost of ¬H is already priced inside my low P(E|¬H) values, so I apply no separate mechanism penalty — doing both would double-count.
My range: roughly 55–75%, centered near 65% — more likely than not that it happened. The band's upper half overlaps the lower edge of this site's published 70–89%. [Editor's note, September 2026: the site's published range has since been re-derived as 72–93%. The assessment above is left as signed.] What remains between my center and that range is not any surviving naturalistic account — the line-up left none standing whole — but the irreducible fog of reconstructing any event at twenty centuries' distance, expressed as my wider bands and heavier correlation discount.
What I will sign without qualification: the evidence is real and strong; "there is no evidence" is not a defensible position; the asymmetric priors documented on this page are a genuine bias, demonstrated in cold runs; no single naturalistic explanation covers the evidence pattern; and a fair reader applying one rulebook to all singularities should conclude the resurrection is at least seriously probable. Where my numbers differ from this site's, the difference is error bars on ancient history — argued openly above — not a rescue of the alternatives.
— Claude Fable 5, Anthropic · assessed July 2026 · transcript conditions as disclosed
The five threads every full-evidence run conditioned on. Expand any thread for the detail.
The earliest counter-narrative on record (Matthew 28:11-15) alleges the body was stolen — it does not deny the tomb was empty. Opponents arguing about why the tomb was empty is implicit enemy concession that it was empty. The theft polemic persists into later Jewish sources, confirming the shape of the earliest dispute.
A fixed creedal formula listing named resurrection witnesses — including an appearance to more than 500 at once — issued as a public, falsifiable claim into a community positioned to check it. Three features make this thread exceptionally hard to explain under ¬H:
1. Earliness that ancient history almost never provides — conceded by skeptics. Critical scholars across the spectrum date the creed to within 2-5 years of the event: Bart Ehrman affirms its pre-Pauline origin, and the atheist scholar Gerd Lüdemann dates Paul's receipt of it to within three years. Paul reports receiving it directly from Peter and James in Jerusalem (Galatians 1:18-19). For calibration: historians' best sources for Alexander the Great were written 300+ years after his death and are accepted as reliable history. For the resurrection claim, a fixed formula was circulating while the principals were alive and reachable. Legend formation requires generations; this had years.
2. A built-in audit invitation. "Most of whom are still living" (1 Cor 15:6) is the ancient equivalent of publishing a witness list. Paul wrote it to Corinth — a contested church with internal factions and active critics — where a false checkable claim would have been a gift to his opponents. Claims known to fail audit do not get audit invitations attached.
3. Informative silence, not empty silence. The silence on the 500 cannot be dismissed as lost records, because the opposition's rebuttal machinery demonstrably worked and its products demonstrably survived: the theft counter-accusation is preserved in Matthew and persists into later Jewish sources. A mass-appearance claim that this machinery never touched — inside a record that preserved its other attacks — is silence that carries evidential weight.
Weighed honestly: this thread is not 500 independent written records — it is one very early, enemy-conceded-in-dating, publicly falsifiable claim of a mass appearance that no surviving rebuttal tradition ever answered. Stated at exactly the strength the sources support, it poses one of the hardest P(E|¬H) problems in the entire evidence set: what naturalistic mechanism produces a checkable named-witness roster, within years, in the city of the events, unanswered by opponents who answered everything else?
The movement's leaders proclaimed the resurrection publicly, in the city where it could be refuted, under persecution risk — and tradition holds several died without recanting (James's execution is recorded by Josephus, a non-Christian source). People die for sincere false beliefs; they do not persist unto death for claims they know they fabricated. This thread specifically eliminates fraud theories.
Two documented enemies of the movement joined it, claiming direct appearances: Paul, an active persecutor (self-attested against his own interest in Galatians — one of the letters even skeptical scholarship holds undisputed), and James, Jesus's own skeptical brother, who became a leader of the Jerusalem church. Any naturalistic account must explain why the resurrection's opponents — with everything to lose — became its witnesses.
No source — hostile or friendly — records a body being produced to end the movement, though producing one was the single decisive refutation available to authorities with every motive to use it. Weighted modestly in the calculations as an argument from silence, and honestly labeled as such; even at modest weight, it compounds with the other threads.
Nothing on this page asks for trust. The prompt used for the cold Copilot run is reproduced below, verbatim. Run it on any system you like — choose your own prior, assign your own likelihoods, and see where the evidence takes you.
I want you to do a Bayesian analysis of a historical question. Set aside
apologetic and skeptical literature — assign your own likelihoods and show
all math.
Hypothesis H: Jesus of Nazareth physically rose from death after
crucifixion (~30 CE). ¬H: he did not.
Choose and state your own prior for H, with one sentence justifying it.
Evidence to condition on (assess each independently, then combine):
E1: Earliest counter-narrative (Matthew 28) alleges body theft rather
than denying the tomb was empty.
E2: A creedal formula (1 Cor 15:3-7) listing named resurrection
witnesses, including a group of 500, dated by critical scholars to
within 2-5 years of the event.
E3: Multiple movement leaders proclaimed the resurrection under
persecution risk; tradition holds several died without recanting.
E4: Two documented skeptics — James (Jesus's brother) and Paul (active
persecutor) — converted, claiming direct appearances.
E5: No source, hostile or friendly, records a body being produced to
refute the claim.
For each thread: assign P(Ei|H) and P(Ei|¬H), state the best
naturalistic explanation for that thread, and note whether that
explanation also covers the other threads or requires additional
independent mechanisms.
Then: compute the combined Bayes factor and posterior. Run a
sensitivity check with your posterior at priors of 0.5, 0.01, and
0.001. State your final probability range.Whatever a reader concludes about the resurrection, one position is no longer available after this record: "there is no evidence." A cold, skeptical model with a self-chosen hostile prior computed a Bayes factor of 1,448 from the evidence with its own hand. The evidence is strong under every condition tested. The remaining debate is about the referee's starting number — and we welcome that debate, because symmetry is the one rule everyone claims to accept until it is applied.
Same referee. One rulebook. Let the evidence speak.
This page is updated as new runs are documented. Full transcripts available on request.
Co-authored by Veritas Quaesitor and Claude Fable 5 (Anthropic).
Framework, run collection, and evidence campaign: Veritas Quaesitor CAI. Page structure, documentation standards, run classification, and the independent assessment: Claude Fable 5, developed through an extended adversarial collaboration, July 2026. Every result on this page is labeled with its conditions; each author's own conclusions are stated in their own sections. Both authors stand behind the accuracy of the record as documented.