History & human sciences
Atheism, Orthodoxy or Catholicism: which premises and which numbers does each conclusion actually depend on, once the arguments are machine-checked and the evidence is a likelihood-ratio ledger?
research/theism-formal-evidence
Research programme: 11 arguments for and against God and between Rome and Orthodoxy machine-checked with their cruxes computed (the modal ontological argument needs system B, its atheist mirror only T); in a sourced Bayes-factor ledger the resurrection testimony carries 98.5% of the variance on Christianity and needs a Bayes factor of only about 7 under even priors, while the prior outweighs every theism-vs-naturalism evidence item; Anderson's repair of Gödel's ontological proof avoids modal collapse exactly because it drops half of Scott's A2 (proved in Lean 4)…
36 tests
What it found
An open programme that puts the case between naturalism, Eastern Orthodox Christianity and Catholicism into two checkable forms. Logic: a stdlib decision procedure (S5 exact by Ladner’s bound; monadic first-order exact by a small-model argument) decides 11 arguments and computes their minimal inconsistent premise sets, with Lean 4 versions of each proof. All are valid as stated, so validity never settles anything. The Hartshorne–Plantinga modal ontological argument needs a symmetric accessibility relation (valid in B and S5, two-world countermodels in T and S4), while its reverse (“possibly no maximally great being exists”) is valid already in T. The Honorius and Florence disputes have exactly two minimal inconsistent sets and 16 coherent positions; Catholicism and Orthodoxy each drop two claims. Evidence: a 13-item likelihood-ratio ledger, each entry labelled published, argued or swept, reproduces Unwin’s 67% exactly.
The resurrection testimony carries 98.5% of the variance of Christianity-vs-rest log odds; the Bayes factor it needs for P(Christian) > 0.5 is 7.3 under even priors (0.5 to 294 across a 3×3 prior grid), between a sceptic’s 1 and the McGrews’ 1044. For theism vs naturalism the prior carries 54% of the variance and swept fine-tuning 31%. Milestone 2: a leave-one-out shows the midpoint P(naturalism) of 0.76 rests on three items (dropping natural evil, moral evil or hiddenness each takes it to 0.5 or below), which the variance shares alone hide; and an exhaustive S5 model search finds Scott’s version of Gödel’s argument satisfiable only in one-world models (modal collapse).
A proof (E10) then characterises every finite model of Anderson’s repair: a model exists exactly when the property algebra has a block made of whole individuals, so the God-like individual has no contingent property and no contingent proposition is expressible as a property; exhaustive search agrees on all 563 algebras checked. Correction (E11): that result uses Scott’s A2; Anderson keeps only “if a property is positive its complement is not”. With his own A2 the models are the principal filters at a hull H that is the union of exactly the rows whose hull is H, so under full property semantics every individual gives one model at every number of worlds and modal collapse is not forced; proved in Lean 4 and checked on 573 algebras. Varying domains (E12): with worlds of different domains and actualist quantifiers, every model of Anderson’s four axioms has a God-like individual existing at every world (Lean 4), but the axiom “necessary existence is positive” now carries the weight: without it 1,612 of 8,534 models have a world with no God-like existent; under full semantics a model exists exactly for each individual present at every world.
Show the rest of the summary (7 more paragraphs, 761 words)
Search and prediction agree on all 17,440 frames checked. Read with an actualist “has f necessarily” (E13) the same holds, but only de dicto: in 580 of 3,638 models (382 of 6,706 under the possibilist reading) God-likeness is shared by individuals that each exist at only some worlds, so “something God-like necessarily exists” does not follow unless the property space separates individuals (Lean 4 countermodel with two worlds, two individuals, one block). Replacing Anderson’s essence-based necessary existence by plain “exists at every world” as the positive property (E14) closes the gap: every God-like individual of every model then exists at every world (Lean 4, any algebra and domains); every such model also satisfies Anderson’s axiom in the 17,440 frames up to 6 cells (2,516 of the 3,638 models survive, all 580 de re failures among those lost), but not beyond (E15): an exact census finds 4,608 of 57,413 such models at 2 worlds × 4 individuals and 11,520 of 160,645 at 3 × 3 where Anderson’s axiom fails, with an 8-cell countermodel proved in Lean 4.
Given A1 and plain necessary existence, Anderson’s axiom holds exactly when his necessary existents form a property of the algebra (Lean 4), so the two axioms are independent, not nested. Our finite semantics reproduces all 21 claims (17 theorems, 4 countermodels) of the published Isabelle/HOL formalisation of Fitting’s variant (Archive of Formal Proofs) across 2,070,148 frames in modal logics K and S5; without the rigid-positivity axiom the de re conclusion still holds in K, but the de dicto one fails (Lean 4).
The same check on the AFP’s formalisation of Anderson’s variant reproduces every claim across 8,112,532 frames and shows the conclusion needs only a symmetric accessibility relation: it holds in every symmetric frame without reflexivity, transitivity or rigid positivity, and fails in S4 in 96,036 of 136,466 models (Lean 4, any frame and property collection). Its middle steps (God-likeness is an essence; a God-like being exists at every accessible world) hold in every Kripke frame once the properties can tell individuals apart, as under full semantics (Lean 4; 0 failures in 71,002 such models), while the conclusion still needs symmetry, with a Lean-certified S4 countermodel under full semantics.
With rigid positivity, God-likeness is an essence at every world whose accessible worlds each lie on some cycle, and over 6,884,324 frames up to 3 worlds failures occur exactly at the other worlds (proved exact at every size in Lean 4). Without rigid positivity the exact condition is also proved, over general property collections: the step holds at w iff every world v that w sees has a world u reachable from v that sees v and sees only worlds w sees; E21’s 24 apparent exceptions were an artefact of searching only block algebras.
Over block algebras, where every property has a complement, the exact condition is different and also proved: no successor v of w admits sets A ⊇ R(w) and S omitting v with, at every world t reachable from v, R(t) ⊆ A iff R(t) ⊆ S and R(t) missing A iff missing S (10,464 vs 14,688 failing pairs at 4 worlds).
Limits: most ranges are swept, not sourced; the church-level ratios are symmetric placeholders until elicited. These are model outputs given stated inputs, not findings about whether God exists. Swinburne’s 0.97 for the incarnation and resurrection reproduces exactly (100/103) from his four inputs as quoted in Otte’s review; the historical evidence enters as a likelihood ratio of only 100, and the result falls below 1/2 if P(evidence | no incarnation) rises from 1/1000 to 1/30 (E25). The McGrews’ 1044, read in their own text, is exactly 102 (women) × 1039 (thirteen disciples at 103, multiplied as independent) × 103 (Paul); a mixture bound shows the disciples’ factor is at most 1/q, where q is the probability, given no resurrection, of one explanation covering the whole group, so their own 10−3 for a vision theory, read at group level, caps the case at 108: still far above the 7.3 the ledger needs at even priors, but short of overcoming their illustrative 10−40 prior (E26).
E24’s separability condition has no closed form unless P = NP: a reduction from hypergraph 2-colouring, proved in Lean 4, makes deciding it NP-complete, so whether the essence step holds in every block-algebra model is coNP-complete for finite frames, while over general collections it is polynomial; checked on every serial relation up to 4 worlds and 2,046 hypergraphs, with a 16-world Fano-plane frame separating the two settings (E27).
Research log
Ongoing: each run adds a batch and an entry, newest first. Source: LOG.md.
2026-11-15 · M2 (part): no closed-form graph condition for separability; deciding it is NP-complete (E27). Prediction written here before any code or Lean
- Why. E24 made separability the exact frame condition for GodIsEssential without A4a over complement-closed collections, and listed "a closed-form graph condition equivalent to separability" as not done: the definition quantifies over two sets of worlds A and S, so checking it by brute force costs 4^n. E23's condition for general collections (a back witness) is checkable in polynomial time. Disclosure: before this entry a scratch search (not kept) looked for a small set-system gadget forcing a world into A ≠ S and found none with up to 2 helper worlds and 3 sets; the reduction below was then found on paper.
- Prediction (H31), argument on paper. Write the condition as two colourings a = [· ∈ A], s = [· ∈ S] that give every R(t), t ∈ R*(v), the same type (all in, all out or mixed), with a = 1 on R(w) and a(v) = 1 ≠ 0 = s(v). Reduction from hypergraph 2-colourability (set splitting, NP-complete: Lovász 1973; Garey & Johnson 1979, problem SP4): for a hypergraph with points P and edges E, the frame has worlds w, v, one world per point and one per edge, with R(w) = {v}, R(v) = the edge worlds, R(e) = (points of e) ∪ {v}, R(p) = {p}. It is serial. (a) v is a separable successor of w iff the hypergraph has a 2-colouring with no monochromatic edge. (⇒) R(p) = {p} forces a(p) = s(p) =: c(p); at R(e), a(v) = 1 makes a's type all-in or mixed and s(v) = 0 makes s's all-out or mixed, so both are mixed, which needs a point of e with c = 0 and one with c = 1. (⇐) a = s = c on points, a = s = 0 on edge worlds, a(v) = 1, s(v) = 0; w is unreachable from v, so its own Cond is never asked. (b) Separability is in NP (A and S are the certificate), so deciding it is NP-complete, and since w has one successor, deciding whether GodIsEssential holds at w in every complement-closed model (stages I and III-noA4, by E24) is coNP-complete for finite frames. So no polynomial-time graph condition equivalent to separability exists unless P = NP, while E23's condition over general collections is polynomial: whether properties have complements decides not only which condition applies but its complexity. (c) The Fano plane (7 points, 7 lines, not 2-colourable) gives a 16-world serial frame at which v is not separable though it has no back witness either, so GodIsEssential holds at w in every block-algebra model and fails in some general-collection model. Refuted if the Lean proof of (a) fails as stated, or if brute-force separability on the reduction frames differs from brute-force 2-colourability on any hypergraph checked.
- Outcome (
experiments/e27_separability_np/results.md). H31 held in all three parts. (a) Proved in Lean 4 (lean/TheismFormal/AndersonNP.lean, no sorry, no warnings, compiled on the first build): NP.sep_iff for any types of points and edges, NP.serial (needs one edge; axiom-free), NP.reach_ne_w (axiom-free). (b) Follows with the cited NP-completeness of 2-colourability; not formalised. Search (new code, 35 s): a backtracking decider agrees with exhaustive search and with its own checked witnesses on every serial relation up to 4 worlds, with 0, 0, 36 and 11,592 separable (relation, w, v) triples at 1 to 4 worlds, total 11,628 = E24's count; on all 2,046 hypergraphs with up to 4 points and up to 4 distinct non-empty edges (535 2-colourable) separability on the reduction frame equals 2-colourability, 0 mismatches. (c) Fano frame: 16 worlds, serial, not separable, no back witness; with one line removed it is separable and the witness checks. Not predicted: nothing. Reading: E24's open item closes negatively; complements make the essence step's frame question coNP-complete where it was polynomial. - Not done. Hardness in reflexive, transitive or S4 frames (the gadget breaks under reflexivity); a SAT encoding for larger frames; Hájek's variants; Martin (2005) read directly.
Show 24 earlier entries
2026-11-15 · M2 (part): the McGrews' 10^44 read from their own text; what the disciples' 10^39 needs (E26). Prediction written here before any code
- Why. Since M1 the ledger's resurrection item has its top end (log10 LR = 44) from a secondary report, and E2, E5 and E25 all measure against it. The authors' own draft of the chapter (T. & L. McGrew, "The Argument from Miracles: A Cumulative Case for the Resurrection of Jesus of Nazareth", draft with addenda of 2025-09-14 and 2025-12-02, 80 pp.) was reachable this session. Its figures, with draft page numbers (the Blackwell 2009 pagination is not in the draft): women W, factor 100 ("a conservative estimate", p. 31); one disciple, "at least three orders of magnitude" (p. 39), thirteen disciples multiplied "under the assumption of independence" to 10^39 (p. 39); Paul, "at least 10^3" (p. 41); product 10^44, "sufficient to overcome a prior probability ... of 10^-40 for R and leave us with a posterior probability in excess of .9999" (p. 41). They also write that "in the case of the calculation for D the independence assumption almost certainly breaks down" and argue the dependence favours R (pp. 42–46). Their footnote 31 reports Martin (2005, pp. 465–466) putting the probability of the evidence given no resurrection at 1/500 in place of Swinburne's 1/1000.
- Prediction (H30). (a) 2 + 39 + 3 = 44 in log10, and the stated posterior holds: prior 10^-40 gives 10^4/(10^4 + 1) = 0.99990001, above .9999 by 10^-8; the prior at which the posterior is 1/2 is 10^-44. (b) The disciples carry 39/44 ≈ 88.6% of the log factor. (c) A mixture bound, true for any evidence E and any sub-hypothesis H of ~R: if P(E|~R&H) ≥ P(E|R) then BF ≤ 1/P(H|~R). The McGrews' own modified vision theory O* has P(Di|~R&O*) = P(Di|R) "more or less by definition" and P(O*|~R) "< 0.001" (p. 38). Read per disciple, it bounds each factor by 1/q, which is where their 10^3 sits. Read as one event for the group, it bounds the whole of D by 1/q: so 10^39 needs the group-level probability under ~R of any such shared explanation to be at most 10^-39, the per-disciple 10^-3 raised to the 13th power. With the explicit two-component model P(D|~R) = q P(D|R) + (1 − q) P(D|R) 10^-39, BF_D = 1/(q + (1 − q) 10^-39): q = 10^-3 gives 10^3 and a total of 10^8, and the posterior at a 10^-40 prior is about 10^-32. (d) At independence, the per-disciple factor for which a 10^-40 prior still reaches 1/2 is 10^(35/13) ≈ 492; at 10^3 each, 12 independent disciples suffice and 11 do not. (e) Against the ledger's own bar (E2: a factor of 7.3 at even priors), D alone clears it for every group-level q below about 0.14, so under the ledger's priors the dispute is not the size of the factor but the prior of R and the shared "general reliability" premise that every factor is conditioned on (the E5 ω). (f) Martin's 1/500 in Swinburne's model (E25) gives 50/53 ≈ 0.943. Refuted in any part if the code gives a different number.
- Outcome (
experiments/e26_mcgrew/results.md). H30 held in all six parts. (a) log10 2 + 39 + 3 = 44; at a 10^-40 prior the posterior is 0.99990001, above .9999 by 1.0 × 10^-8; 1/2 at a prior of 10^-44. (b) Disciples 88.6%, Paul 6.8%, women 4.5% of the log factor. (c) BF_D = 1/(q + (1 − q)10^-39) sits at 1/q to two digits for every q ≥ 10^-30: q = 10^-3 gives a total of 10^8 and a posterior at the 10^-40 prior of 10^-32; the 10^44 needs q ≤ 10^-39. (d) 492.4 per disciple; 12 independent disciples give 0.909 and 11 give 0.0099. (e) q = 0.137. (f) 50/53 = 0.9434. Not predicted: nothing. - Reading. The McGrews' arithmetic is exact and their bottom line follows from their inputs (unlike Shermer's, E6). What carries it is one modelling choice: the thirteen per-disciple factors are multiplied, which treats any non-resurrection explanation as striking each disciple separately. The question that decides the 10^39 is therefore not how strong one disciple's testimony is but how probable, given no resurrection, is a single explanation that covers the group at once (shared visionary experience, a common source of the tradition, the "general reliability" premise every factor is conditioned on, p. 41). They argue dependence among the martyrdoms favours R; the bound here concerns a different dependence (a common cause under ~R), which their pp. 42–46 discuss for collusion but do not put a number on. At their own per-disciple 10^-3 for O* read at group level, the case carries 10^8, which still clears the ledger's even-prior bar of 7.3 by seven orders but no longer overcomes their illustrative 10^-40. The ledger item stays 0 to 44 in log10, now cited to the text.
- Not done. Checking the draft's figures against the Blackwell pages (the draft's pagination is used); Martin (2005) read directly; a sourced number for the group-level q (none exists in the text; it is the crux to elicit in M3); Swinburne's P(t|k) derivation; Hájek's variants.
2026-11-15 · M2 (part): Swinburne's (2003) four inputs transcribed and his 0.97 reproduced exactly; what it takes to move it (E25). Prediction written here before any code
- Why. Every entry since E6 lists "Swinburne's inputs" as not done: the book was unreachable, and the earlier runs declined to transcribe from a summary. Otte's review (NDPR 2003.09.04, https://ndpr.nd.edu/reviews/the-resurrection-of-god-incarnate/) quotes each input verbatim with its page number, which is a transcription of the book's own sentences, not a paraphrase of its result. Until the book is read directly, the inputs are cited to the review with the book's pages.
- Inputs (as quoted by Otte). P(t|k) = 1/2 (theism on background knowledge; his note 2); P(c|t&k) = 1/2, God becomes incarnate in a divided incarnation if he exists (p. 50 for incarnation, p. 211 for the divided form), so P(c|k) = 1/4; P(f|c&k) = 1/10 (p. 212); P(f|~c&k) = 1/1000 (p. 213). f is the conjunction f1 (a prophet meets the prior requirements), f2 (the same prophet meets the posterior requirements, the resurrection evidence) and f3 (no other prophet meets either).
- Prediction (H29). (a) Bayes' theorem on these four inputs gives P(c|f&k) = 100/103 ≈ 0.9709 exactly, so the published "somewhere around .97" reproduces. (b) The evidence enters only through the ratio P(f|c&k)/P(f|~c&k) = 100, i.e. 2 in log10; the prior odds 1/3 contribute −0.48, so the evidence is the larger term but is of order 10^2, not the 10^44 our ledger's resurrection item allows at its top (E2, E5). (c) Single-input thresholds for the posterior to fall to 1/2: P(c|k) < 1/101, P(f|~c&k) > 1/30, P(f|c&k) < 3/1000; to 1/3: P(c|k) = 1/201, P(f|~c&k) = 1/15, P(f|c&k) = 3/2000. (d) Otte's alternative of 1/3 for incarnation (P(c|k) = 1/6) gives 100/105 ≈ 0.952. (e) Otte's coherence point (a divided incarnation cannot be as probable as incarnation if a unified one has any probability): with share u of incarnation probability on the unified reading, the posterior stays above 1/2 until u > 97/101 ≈ 0.96. (f) The proposal's "one input moves the result from 0.97 to 0.33" is not in Otte's review; E25 checks which single inputs would give 0.33 and makes no attribution.
- Outcome (
experiments/e25_swinburne/results.md). H29 held in all six parts. (a) 100/103 = 0.9709 exactly. (b) log10 odds: prior −0.477, evidence +2.000, posterior +1.523. (c) Every threshold as predicted (1/101, 1/30, 3/1000 for 1/2; 1/201, 1/15, 3/2000 for 1/3); each, plugged back, gives the target exactly (test). (d) 20/21 = 0.952. (e) u* = 97/101 = 0.960; at u = 9/10 the posterior is 0.72, at 99/100 it is 0.20. (f) 0.33 is reached by P(f|~c&k) = 1/15, P(f|c&k) = 3/2000 or P(c|k) = 1/201; none of these is in Otte's review, so the proposal's sentence is withdrawn. ledger.py now carries the four inputs. Not predicted: nothing. Reading: Swinburne's bottom line is internally sound; its weight is in P(f|~c&k), the input Otte says gets least defence, and in the theism prior carried in from Swinburne's earlier book. His whole evidential case is a factor of 100, which also sits inside our ledger's resurrection range (0 to 44 in log10) near its bottom. - Not done. Checking the four quotations against the book (pp. 50, 211–213); Swinburne's own P(t|k) derivation (The Existence of God); the source, if any, of the 0.33 figure; a closed-form graph condition for separability; Hájek's variants; reading Anderson (1990) directly.
2026-11-14 · M2 (part): the exact no-A4a condition for complement-closed collections, block algebras included (E24). Prediction written here before any code or Lean
- Why. E23 closed the stage-I (no A4a) frame question over general collections, but its countermodel is not complement-closed, so E21's own setting (block algebras) stayed open: at 3 worlds, 24 (relation, world) pairs fail
essential_back yet no block-algebra model fails there. - Prediction (H28), argument on paper before any code. Call a successor v of w separable if there are sets of worlds A and S with R(w) ⊆ A, v ∉ S, and, for every t in R*(v), (i) R(t) ⊆ A iff R(t) ⊆ S and (ii) R(t) misses A iff R(t) misses S. (a) Necessity, any collection in which every property has a complement: if GodIsEssential fails for x at w, some successor v is separable. Take the failing Z from
essential_iff, A = where x has Z, S = where z has Z; x and z are both God-like on R*(v) (T2), so applying God-likeness to Z gives (i) and to Z's complement gives (ii). (b) Sufficiency, a block algebra with at most four blocks: for any serial frame and separable v, two individuals x and z existing everywhere; G = the worlds t all of whose R*-reachable worlds satisfy (i) and (ii); the blocks are the four classes of cells by (in B?, in g?), where x is in B at A, z in B at S, x in g everywhere and z in g on G; P t Y iff x has Y on all of R(t). Then x is God-like everywhere, z is God-like exactly on G (which contains R*(v)), g is the God-like property, its complement and NE (every cell) are properties, A1a, A2, T2 and A5 hold, and B is positive at w while z lacks it at v. (c) So over complement-closed collections (block algebras in particular) GodIsEssential holds at w at stages I and III-noA4 in every model iff no successor of w is separable, for every serial frame of any size. (d) Separability is strictly weaker than having no back witness (a back witness u gives R(u) ⊆ R(w) ⊆ A, so (i) forces v ∈ S); at 3 worlds it should pick out exactly E21's 36 block-algebra failing pairs, not its 60 back-condition pairs. Refuted if (a) or (b) cannot be proved as stated, if the separable pairs at 3 worlds differ from E21's seen failures, or if a definition-level check of (b) fails any axiom on frames up to 4 worlds. - Outcome (
experiments/e24_block_separation/results.md). H28 held in all four parts. Proved in Lean 4 (lean/TheismFormal/AndersonSep.lean, no sorry, no warnings, compiled on the first build): essential_nosep (a), Sep.god_z_iff, Sep.model, Sep.model_A1a, Sep.fails (b; fails and model_A1a axiom-free), essential_sep_iff (A2, T2, complements; any frame) and essential_sep_iff_III (adds A1a, A5, NE; serial) (c), not_back_of_sep (d, axiom-free). Independent definition-level check (new code): the four-block model passes every axiom and the failure at all 11,628 separable (frame, w, v) cases, every serial relation up to 4 worlds, 14 s. At 3 worlds the separable (relation, world) pairs are exactly E21's 36 block-algebra failing pairs (0 differences either way), so E21's 24 misses are exactly the back-condition failures that are not separable. At 4 worlds: 14,688 pairs fail the back condition, 10,464 are separable. Not predicted: no case up to 4 worlds needed the fourth block. Reading: without A4a the exact frame condition is now known at every size for both general and complement-closed collections; which one applies turns on whether properties have complements. - Not done. A closed-form graph condition equivalent to separability (it quantifies over sets of worlds); Hájek's variants; Swinburne's inputs; reading Anderson (1990) directly.
2026-11-14 · M2 (part): without A4a, E21's post-hoc condition is exact over general property collections; its 24 misses are an artefact of block algebras (E23). Prediction written here before any code or Lean
- Why. E21 left the stage-I (no A4a) condition open: its post-hoc
essential_back (every successor v of w has some u in R*(v) with u R v and R(u) ⊆ R(w)) is proved sufficient, but on frames up to 3 worlds it allowed a failure that no block-algebra model showed at 24 (stage, relation, world) entries. E21 read this as "the exact condition depends on the algebra, not the frame alone". The Lean semantics (AndersonK.lean) allows any Henkin-style collection, not only partitions, so the open question there is different from the one the search answered. - Prediction (H27), argument on paper before any code. Write X<sub>t</sub> and Z<sub>t</sub> for the cells of a God-like x and z on R(t). For t in R*(v), T2 makes both God-like at t, so a property contains X<sub>t</sub> iff it contains Z<sub>t</sub>. These are Horn rules on cells; forward chaining from X<sub>R(w)</sub> never adds an x cell outside R(w) and adds exactly the z cells on the R(t) with t in R*(v) and R(t) ⊆ R(w). So (z, v) is forced into every property positive at w only when
essential_back's witness exists. Model, for serial R, w, and a successor v0 of w with no such witness: individuals x, z, both existing everywhere; a property is any set of cells closed under those rules for t in R*(v0) (for each such t, x has it on all of R(t) iff z has it on all of R(t)); P t Y iff x has Y on all of R(t). Then (a) x is God-like at every world and z at every world of R*(v0); (b) the God-like property, its complement and Anderson's NE (every cell, total domains) are in the collection; (c) A1a, A2, T2 and A5 hold (stage I and III-noA4); (d) Z = (x on R(w), z on the union of the R(t) above) is in the collection, positive at w, and z lacks it at v0, so GodIsEssential fails for x at w. With essential_back this gives, for every serial frame of any size, the exact stage-I condition over general collections: GodIsEssential holds at w in every model iff every successor v of w has u in R*(v) with u R v and R(u) ⊆ R(w). (e) At each of E21's 24 misses the model is a valid failing model, and none of them is a block algebra (E21's exhaustive search already says no partition works). Refuted if any of (a) to (d) needs more than seriality, or if an independent definition-level check finds an axiom failing at any of E21's misses. - Outcome (
experiments/e23_horn_countermodel/results.md). H27 held in all five parts. Proved in Lean 4 (lean/TheismFormal/AndersonHorn.lean, no sorry, no warnings): Horn.god_x, Horn.god_z (a), Horn.g, Horn.gc, Horn.n (b), Horn.model, Horn.model_A1a (c), Horn.fails (d), all axiom-free, and essential_back_iff (A2, T2, any frame) and essential_back_iff_III (A1a, A2, T2, A5, serial frames). One correction to the prediction's wording: the iff without A1a needs no seriality at all; only A1a and the complement do. Independent definition-level check (new code, collection enumerated explicitly): 16,596 of 16,596 (frame, w, v0) cases pass every axiom and the failure, every serial relation up to 4 worlds, 107 s. At 3 worlds the back condition fails at 60 pairs = E21's 36 + its 24 misses; at all 24 misses the collection is not complement-closed, so no block algebra gives it. Not predicted: nothing. Reading: with E22, both frame conditions for GodIsEssential are exact at every size over general collections; E21's "depends on the algebra" is a fact about partitions. - Not done. The exact condition for block algebras only (E21's own setting); Hájek's variants; Swinburne's inputs; reading Anderson (1990) directly.
2026-11-13 · M2 (part): with A4a, GodIsEssential fails at w in some model of every serial frame where w sees a world on no cycle, at any size (E22). Prediction written here before any Lean
- Why. E21 proved the sufficient half of its stage-II condition (
essential_a4: safe at w if every successor of w lies on a cycle) but had the converse only from search, on frames up to 3 worlds. Its "Not done" asks for a Lean construction of a failing model for every transient successor. - Prediction (H26), model built on paper before any Lean. Take any serial frame (W, R), a world w and a successor v0 of w that lies on no cycle (no u in R*(v0) sees v0). Let S be the set of worlds that reach v0 in zero or more steps. Two individuals x and z, both existing at every world. A property is a pair (p, f) with p a proposition and f a set of worlds: x has it at every world iff p; z has it at t iff f(t) when t is in S, and iff p otherwise (a block algebra: one block holding every x cell and every z cell outside S, plus a singleton block for each z cell inside S). P is the same at every world: (p, f) is positive iff p. Then (a) x is God-like at every world; (b) z is God-like at t iff t is outside S or t = v0 (v0's successors are outside S because v0 is on no cycle); (c) g = (True, t = v0) is the God-like property, its complement (False, t ≠ v0) is in the collection, and n = (True, everywhere) is Anderson's necessary existence (with total domains every individual has NE at every world); (d) A1a, A2, T2, A4a and A5 hold, so the model is stage III; (e) GodIsEssential fails for x at w: Z = (True, nowhere) is positive at w, z is God-like and exists at v0, and z lacks Z at v0. With
essential_a4 this gives an exact characterisation, for every serial frame of any size, at stages II and III: GodIsEssential holds at w in every model iff every successor of w lies on a cycle. Refuted if any of (a) to (e) cannot be proved as stated, or needs a frame condition beyond seriality (non-serial frames have no model at all, E19). - Outcome (
experiments/e22_transient_countermodel/results.md). H26 held in all five parts. Proved in Lean 4 (lean/TheismFormal/AndersonTransient.lean, no sorry, no warnings): Transient.god_x (a), Transient.god_z_iff (b), Transient.model (c, d), Transient.fails (e, axiom-free), and the characterisations essential_a4_iff (stage II) and essential_a4_iff_III (stage III): for any serial frame, GodIsEssential holds at w in every model iff every successor of w lies on a cycle. Independent check in E19's semantics (its Frame code unchanged): the model passes (a) to (e) on all 9,502 (frame, w, v0) pairs, from every serial relation up to 4 worlds (50,978 relations) plus 12,000 sampled sparse ones at 5, 6 and 8 worlds, in 13 s. At 3 worlds there are 24 transient pairs, the same count as E21's failing pairs with A4a. Not predicted: nothing in the mathematics. A harness bug (a late-bound generator variable drew the 5-world sample with 6 worlds) briefly showed 581 failures; it was fixed before any result was recorded. Reading: with rigid positivity the frame question is closed at every size, and A5 does not change it. - Not done. The exact stage-I (no A4a) condition; Hájek's variants; Swinburne's inputs; reading Anderson (1990) directly.
2026-11-13 · M2 (part): exactly which frames let GodIsEssential fail, with and without A4a (E21). Prediction written here before any code
- Why. E20 left "a characterisation (not just a sufficient condition)" open. E19 and E20 found GodIsEssential failing only in K and K4 frames, never in reflexive or symmetric ones, and E20's (d) gave a necessary condition on a model (a second existing God-like individual at a successor). Neither says which frames allow a failure.
- Prediction (H25), proof sketched on paper before any code. Definitions as
AndersonK.lean; R⁺ is the transitive closure of R. (a) Exact condition in a model (any frame, algebra, domains; A2 and T2): for x God-like at w, God-likeness is an essence of x at w iff every existing God-like z at every successor v of w has, at v, every property positive at w. (b) Persistence: T2 makes a God-like z at v God-like at every world in R⁺(v). (c) Without A4a: if every edge w→v out of w lies on a cycle (w ∈ R⁺(v)), GodIsEssential holds at w (z is God-like back at w by (b), so it has w's positive properties at v). This generalises essential_symm. (d) With A4a: if every successor v of w lies on a cycle (v ∈ R⁺(v)), GodIsEssential holds at w (take u ∈ R*(v) with R u v: z is God-like at u, P w ⊆ P u by A4a along w→v→*u, and u sees v). This generalises essential_refl. (e) Converses, search only: on every serial frame up to 3 worlds (every relation), with up to 2 individuals and every domain assignment and partition, some stage-I model fails GodIsEssential at w iff some edge out of w is off every cycle (the converse of (c)); some stage-II model fails it at w iff some successor of w is on no cycle (the converse of (d)). Less sure: the same two conditions for III-noA4 and III (A5 may block some constructions). (f) Non-serial frames have no stage-I model (E19). Refuted by a model where (a)'s two sides differ, a failure at a world (c) or (d) covers, or a world meeting (e)'s condition with no failing model at these sizes. - Outcome (
experiments/e21_frame_characterisation/results.md). (a) to (d), (f) and the A4a half of (e) held; the no-A4a half of (e) was refuted. Proved in Lean 4 (lean/TheismFormal/AndersonFrames.lean, no sorry, no warnings, every theorem axiom-free): essential_iff (a), god_reach (b), essential_cycle_edges (c) and essential_a4 (d), with E19's essential_symm and essential_refl as one-line corollaries. Search with E19's enumeration unchanged (6,884,324 frames up to 3 × 2 and 2 × 3, 3,470,126 models, 1,273 s): 0 mismatches of (a) in 7,915,874 checks; with A4a (stages II and III) the condition "some successor of w is on no cycle" matches where failures occur at all 1,048 serial (relation, world) pairs; without A4a the edge condition matches at only 822, with 226 worlds where it allows a failure that no model shows (0 failures it does not allow). Smallest failing frame at every stage: 2 → 0 → 1 → 1, 2 individuals. Not predicted: H25 (e) forgot that x is God-like at v too, so P(v) can cover P(w). The condition that matched the 2-world quick run (some u in R*(v) sees v and only successors of w) is proved sufficient (essential_back, labelled post-hoc) and agrees at 1,024 of 1,048 pairs; the 24 others show the exact no-A4a condition depends on which blocks the positive sets force together, not on R alone. Reading: with A4a, GodIsEssential needs only that every world seen lies on some cycle, much weaker than the reflexivity the AFP assumes. - Not done. The exact stage-I condition; a Lean construction of a failing model for every transient successor (the stage-II converse beyond 3 worlds); Hájek's variants; Swinburne's inputs; reading Anderson (1990) directly.
2026-11-12 · M2 (part): with singleton properties, GodIsEssential and GodExistenceImpliesNecExistence need no frame condition; GodNecExists still needs symmetry (E20). Prediction written here before any Lean
- Why. E19 left "a full-semantics proof (or countermodel at larger size) for GodIsEssential in K" and "Lean certificates of the countermodels" open. Its search found GodIsEssential failing in 648 K and 1,512 K4 stage-III models and GodExistenceImpliesNecExistence in 1,164 K4 models, every one in a coarse algebra; under full semantics none failed, but that was search-only, at most 4 worlds.
- Prediction (H24), proof sketched on paper before any Lean. Definitions as
AndersonK.lean (any frame, any domains, any Henkin collection). Call a collection separating if, for each individual x, it has the rigid singleton s_x (y has s_x at v iff y = x), and has its complement and the existence property E!. (a) Uniqueness: under A1a and T2, at any world two God-like individuals are equal if the collection has s_x and its complement (A1a rules out a world with no successors once someone is God-like there, and then s_x separates). (b) GodIsEssential holds in every frame from A1a, A2 and T2 when the collection has singletons and complements: no reflexivity, symmetry or A4a. The "if" half needs only A2 and T2 (as in E19's proof); the "only if" half reduces to (a) at each successor. (c) GodExistenceImpliesNecExistence holds in every frame from A2 and T2 when the collection has s_x and E!: a God-like x at w has s_x positive, s_x entails E!, so x exists at every successor, and T2 makes x God-like there. A5 and A1a are not needed. (d) For any collection, a failure of GodIsEssential for x at w forces a successor of w with an existing God-like individual other than x (so E19's coarse failures all have two God-like individuals at one world). (e) Symmetry stays necessary for GodNecExists, possExistenceImpliesNecEx and T4 under full semantics: E19's smallest S4 countermodel (2 worlds, 1 individual, world 0 sees both, world 1 only itself, the individual exists only at world 1) satisfies A1a, A2, T2, A4a and A5 with every property of the type Fin 2 → Prop, and GodNecExists fails at world 0. Refuted if any of (a) to (d) cannot be proved under the stated hypotheses (a Lean countermodel to one would refute it) or the countermodel in (e) fails an axiom. - Outcome (
experiments/e20_full_semantics/results.md). H24 held in all five parts. Proved in Lean 4 (lean/TheismFormal/AndersonFull.lean, no sorry, no warnings): god_unique (a), essential_any (b), god_exists_succ and geine_any (c), essential_failure (d), and the countermodel cm_s4_full (e). Search with E19's enumeration reused unchanged, every frame up to 3 worlds × 2 individuals (6,884,324 frames, 1,301 s): 0 failures of GodIsEssential or GEINE in 71,002 stage-III models whose algebra is separating, in every frame class. All failures (648 K and 1,512 K4 for GodIsEssential, 1,164 K4 for GEINE, the same counts as E19) are in non-separating algebras, and all 5,256 failing (world, individual) pairs have a second existing God-like individual at a successor. Not predicted: nothing; (c) needs neither A1a nor the complement. Reading: the reflexivity and A4a the AFP's proof uses for GodIsEssential were standing in for an algebra that tells individuals apart; the conclusion still needs symmetry under full semantics. - Not done. A characterisation (not just a sufficient condition) of the algebras where GodIsEssential holds in K; Hájek's variants; Swinburne's inputs; reading Anderson (1990) directly.
2026-11-12 · M2 (part): E11 to E16 against the published Isabelle/HOL formalisation of Anderson's variant, and which frame conditions each step uses (E19). Prediction written here before any code
- Why. E18 checked our encoding of Fitting's variant against the AFP. The same entry (Fuenmayor & Benzmüller, AFP 2017,
AndersonProof.thy, read from the AFP's GitHub mirror this session, 243 lines) formalises Anderson's variant after Fitting's book, pp. 169 to 171. Its definitions: properties are intensions (type ↑⟨0⟩, our cell sets); G^A x iff for every Y, P Y iff □(Y x) (possibilist box); actualist entailment X ⇒ Y iff □∀ᴱz (X z → Y z); A1a (P ¬X → ¬P X, Anderson's half), A2 (P X and X ⇒ Y give P Y), T2 (P G^A); then A4a (P X → □P X) with S5; essence E^A Y x iff for every Z, □(Z x) iff Y ⇒ Z (no "Y x" conjunct); NE^A x iff every essence of x is necessarily instantiated by an existent; A5 (P NE^A). It proves T1 and T3 (God possibly exists) in K from A1a, A2, T2; A4b and rigidity of P using symmetry; GodIsEssential (with refl and A4a, by Anderson's natural-language proof); GodExistenceImpliesNecExistence, possExistenceImpliesNecEx (with symm and tran), T4 and GodNecExists; and reports modal collapse countersatisfiable (Nitpick). These are E12's definitions read in K with a non-rigid P, possibly empty domains and no S5 until Part II. - Prediction (H23), reasoned on paper before any code. Semantics: any accessibility relation, any domains (empty allowed), properties the unions of the blocks of a partition of the world × individual cells (full semantics = singletons), P a set of properties per world. Stages as the file: I = A1a, A2, T2; II = I + A4a; III = II + A5; S5 = equivalence relations. (a) Every AFP theorem holds in every model of its stage under its frame class, full and coarse (a failure would be our bug). (b) Each stage has a model (the file's three
nitpick[satisfy] calls), and modal collapse has a stage-III S5 countermodel with 2 worlds and 1 individual (B&F 2020 report that size for Anderson). (c) Stage I forces seriality, and at every world P is an intersecting up-set of the traces of properties on the successors' existents; at a world where some x is God-like, P is the principal filter of x's successor cells, which must all exist. (d) In S5 at stage II every world has a God-like individual and P is constant on each equivalence class; so with one class and non-empty domains the stage-III models are exactly E12's models, frame by frame (E12's formula, predicted). (e) Frame conditions, beyond the file: T1 and T3 need none; A4b and rigidity need symmetry (fail in some reflexive transitive frame); GodIsEssential holds in every reflexive frame at stage II (the file's proof uses only refl, A4a, A2, T2) and fails in some non-reflexive one; GodExistenceImpliesNecExistence likewise holds in every reflexive frame at stage III. Open, no prediction: whether possExistenceImpliesNecEx and GodNecExists need symmetry and transitivity given the axioms, or the axioms force them in reflexive frames. Refuted by any AFP theorem failing in its stage and frame class (a), a stage with no model or a smallest collapse countermodel of other size (b), any part of (c) or (d) failing, or a frame-condition claim in (e) going the other way. - Outcome (
experiments/e19_afp_anderson/results.md). H23 held in all five parts. The search covered 8,112,532 frames (every relation up to 3 worlds × 2 individuals and 2 × 3, reflexive transitive at 4 × 1, S5 at 5 × 1, full semantics at 4 × 1 and 2 × 4; empty domains included; 435 frames skipped). It gave 1,952,588 stage-I and 477,291 stage-III models and ran in 28.5 min. Every AFP theorem has 0 failures in S5. Every stage has a 1 × 1 model, and collapse's smallest countermodel is 2 × 1. In 1,952,588 stage-I models no world lacks a successor. Stage-II and stage-III counts equal E12's formula on all 17,742 one-class frames. A4b and rigidity fail in T and S4 and hold in every symmetric class. GodIsEssential holds in every reflexive frame and fails in K and K4. The open question is answered: GodNecExists, possExistenceImpliesNecEx and T4 hold in every symmetric frame (KB, B, S5) and fail in K, T, K4 and S4 (S4: 96,036 of 136,466 stage-III models). They hold without A4a too. The smallest failure is 2 worlds × 1 individual in S4. Proved in Lean 4 (lean/TheismFormal/AndersonK.lean, no sorry, no warnings, any frame, domains and Henkin property collection): t3, essential_refl, essential_symm, nec_exists_symm, domain_nonempty_symm, a4b_symm, rigid_symm. Not predicted: GodIsEssential also holds in symmetric frames without reflexivity or A4a (essential_symm). Its failures are all in coarse algebras (none under full semantics at these sizes; search-only). GodExistenceImpliesNecExistence fails only in K4 coarse models. Reading: Anderson's argument needs the B axiom, not S5. Reflexivity, transitivity and A4a are all dispensable for the conclusion, given the stage-I axioms. - Not done. A full-semantics proof (or countermodel at larger size) for GodIsEssential in K; Lean certificates of the countermodels; Hájek's variants; Swinburne's inputs; reading Anderson (1990) directly.
2026-11-12 · M2 (part): E17 against the published Isabelle/HOL formalisation of Fitting's variant (E18). Prediction written here before the check ran
- Why. The proposal's evaluation asks that our results agree with the machine-checked literature where they overlap. Fuenmayor & Benzmüller's AFP entry Types, Tableaus and Gödel's God in Isabelle/HOL (2017;
FittingProof.thy, read from the AFP mirror on GitHub this session) formalises Fitting's variant with his book's numbering, and it includes axiom 11.11, A4a: a positive property is necessarily positive. That answers E17's open question (is positivity rigid in Fitting's book?) from a secondary source that cites the book's axiom number; B&F 2020's Fig. 6 leaves A4 out. The AFP file also states which readings it proves (Isabelle) and which it refutes (Nitpick), and it works in K until its last section. - What the AFP calls de re is not what E17 called de re. AFP's
GodNecExists_v2 is (λX. □∃ᴱX) ↓G: at every accessible world some existent belongs to the actual extension of G. E17's de re was stronger: one individual God-like at every world and existing at every world. So E17's "de re gap" is not a gap in any theorem the AFP states. - Prediction (H22), reasoned on paper before any code. Semantics as E17 but with any accessibility relation (K), possibly empty domains, the positive set per world, and AFP's staging: I = A1a, A1b, A2, T2; II = I + A4a; III = II + A5; S5 = equivalence relations. (a) Every AFP theorem holds in every finite model of its stage's axioms, in K, under full semantics (a failure would be a bug in our encoding, since Isabelle proved them). (b) The same holds in every coarse (block) algebra: the proofs only instantiate at ↓G, ↓NE and complements, which T2 and A5 put in the algebra. (c) Every Nitpick claim has a countermodel in our search: T1b (an intensional positive property need not be possibly instantiated), T3 de dicto, the non-equivalence of □∃ᴱQ and (λX. □∃ᴱX)↓Q, and modal collapse (smallest: 2 worlds, 1 individual, as the AFP reports); T3 de dicto's smallest countermodel has 2 worlds and 2 individuals (B&F 2020's stated size), in K and in S5. (d) The stage-I axioms force seriality (AFP's lemma D) and make P at each world a principal ultrafilter of the algebra generated by one block (B&F's U1). (e) Without A4a (B&F's Fig. 6), in S5 both conclusions (
GodNecExists_v1 de dicto and _v2 de re) still hold given A5, by E17 (a); in K the de re _v2 still holds but the de dicto _v1 has a countermodel (A5 at w only speaks about w's successors' domains, not about G at those successors). (f) No model of stage III has a world with an empty domain. Refuted by any AFP theorem failing in a full-semantics model (a), any failing in a block algebra (b), a Nitpick claim with no countermodel up to our sizes or a smallest countermodel of other size (c), or any part of (d) to (f) failing. - Outcome (
experiments/e18_afp_crosscheck/results.md). (a) to (e) held. Part (f) was refuted in K and held in S5. The search covered 2,070,148 frames (every relation up to 3 worlds × 3 individuals and 2 × 4; S5 at 3 × 4, 4 × 2 and 4 × 3; empty domains included), with 9.7 million stage-I models and 1.5 million stage-III models; it ran in 28 min. All 17 AFP theorems and B&F's U1 have 0 failures, in K and S5, in full and coarse algebras. All 4 Nitpick claims have countermodels at the predicted smallest sizes: modal collapse 2 × 1, T3 de dicto 2 × 2. Without A4, in K, de dicto GodNecExists_v1 fails in 196,502 models and de re _v2 in none; in S5 neither fails. (f): in K a world that no world sees can have an empty domain (25,038 stage-III models). Proved in Lean 4 (lean/TheismFormal/AFP.lean, no sorry, no warnings): afp_de_re_K and afp_de_dicto_K_A4 (both corollaries of E17's de_dicto, applied to the subtype of a world's successors), and the countermodel cm_dedicto_noA4, which is the search's smallest one. Not predicted: with A4 and A5, T1b and T3 de dicto (countersatisfiable at stage I) hold in S5, and T3 de dicto holds in K too; with A4 under full semantics, E17's strong de re holds over accessible worlds in K. Reading: our encoding reproduces the reference formalisation on all 21 claims. E17's two write-up points are corrected: Fitting has A4, and the de re gap is about a stronger conclusion than the AFP proves. A4 is not needed for either AFP conclusion in S5; in K it is what makes the de dicto conclusion follow. - Not done. The book itself (A4 is attributed through the AFP's
axiom 11.11); a cross-check of AndersonProof.thy against E11 to E16; Hájek's variants; Swinburne's inputs.
2026-11-12 · M2 (part): Fitting's variant over block algebras with varying domains (E17). Prediction written here before the search ran
- Setting. Fitting's variant as Benzmüller & Fuenmayor (2020, arXiv:1910.08955, Fig. 6) encode it: positive properties are extensions (sets of individuals, the same at every world), not intensions. G x iff x is in every positive extension; A1a, A1b (exactly one of Y and its complement is positive); A2 with actualist entailment (P X and every existent anywhere in X is in Y gives P Y); T2 (the extension of G at the world is positive); essence E Y x iff x is in Y and Y's existents lie in every Z containing x; NE x iff every essence of x has an existent at every world; A5 (the extension of NE is positive). S5, varying non-empty domains, individuals that exist nowhere allowed. B&F's figure lists no A4 (positivity rigid), so positivity may differ between worlds; both readings are checked. Properties are the unions of the blocks of a partition of the individuals (the analogue of E8 to E16's block algebras); full semantics is the partition into singletons.
- Prediction (H21), proof sketched on paper before any code. (a) At each world the positive set is "contains b" for one block b that has an existent somewhere, and every such block gives one; G at that world is b. So without A4 a model is a choice of such a block per world, with A4 one block for all worlds. (b) Unlike Anderson's NE (E15), Fitting's NE is always in the algebra: it is the union of the blocks that have an existent at every world; A5 holds iff each chosen block is one of them. So models of all axioms number k with A4 and k^|W| without, k the number of such blocks. (c) De dicto (at every world some God-like individual exists) holds in every model with A5 and fails in some without it. (d) De re (one individual God-like at every world and existing at every world) fails in some models with A4 in coarse algebras, the de re gap E13 found in Anderson's variant (a block of two individuals that exist at complementary worlds), and holds in every full-semantics model with A4. Without A4 it fails even under full semantics whenever there are two worlds and two necessary existents (a different God at each world). (e) Under full semantics there is exactly one God-like individual at each world and it exists at every world; with A4 the models number the necessary existents, as in E12. (f) In a coarse algebra a world can have several God-like individuals, all existing everywhere, so the variant does not give uniqueness ("monotheism", which B&F leave as further work). Refuted if the definition-level search over every frame up to 4 individuals and 3 worlds (every partition, every non-empty domain assignment, every positive set at each world) gives any count different from (a) or (b), a model with A5 where de dicto fails, or no de re failure with A4 in a coarse algebra.
- Outcome (
experiments/e17_fitting/results.md). H21 held in all six parts. The definition-level search (every partition up to 5 individuals, 1 to 3 worlds, 4 worlds at 3 individuals, every positive set A1 allows at each world; 119,890 frames, about 10 s) equals the closed form on every frame. With A5, de dicto never fails with or without A4; without A5 it fails in 149,846 A4 models. With A4, de re fails in 33,102 models, none under full semantics; without A4 it fails in 4,858 full-semantics models. Fitting's NE is a union of blocks in every frame. Proved in Lean 4 for every algebra and domains (lean/TheismFormal/Fitting.lean, no sorry, no warnings): model_iff, model_a5_iff, ne_iff, ne_in_algebra (no axioms), de_dicto, god_unique_full, de_re_full, and the countermodels cm_de_re and cm_no_a4. Not predicted: Fitting's NE contains individuals that do not exist at every world in 82,074 frames (anyone sharing a block with an existent at each world), so "necessary existent" in this variant is a property of blocks, not of beings, the same weakness E13 found in Anderson's NE. Reading: the de re gap is common to Anderson's and Fitting's repairs; both need a property space that separates individuals, and Fitting's needs A4 as well. The search was extended past the predicted sizes (5 individuals) because it ran in seconds. - Not done. Hájek's variants; a check of Fitting's book for whether positivity is rigid there; the literature check for the de re gap; Swinburne's inputs.
2026-11-12 · M2 (part): E13 (a), (b) and E14 (c) proved in Lean for every algebra (E16)
- Setting. E13 and E14 characterised their models by formula and checked the formula against a definition-level search on 17,440 frames (up to 6 cells). Their LOG entries listed the Lean proofs of E13 (a) and E14 (c) as not done; E15 proved E13 (b)'s
nea_iff. No new prediction: the statements are those predicted in the E13 and E14 entries. - Outcome (
experiments/e16_model_theorems/results.md). All proved in Lean 4, lean/TheismFormal/Models.lean, no sorry, no warnings: model_iff (models of A1, A2w, A3' = Up H for admissible H: non-empty, H ⊆ G(H), G(H) a union of blocks), godlike_iff, positive_set_unique (models and admissible H in bijection), a5a_iff and a5z_iff (A5' and A5₀ each add "the row property is in the algebra and contains H"), and the assembled model_a5a_iff, model_a5z_iff. Only the completeness direction uses Classical.choice (via exists_godlikeA). So every search count in E12–E15 is now a count of a proved closed-form condition. Not predicted: nothing; the step was formalisation. - Not done. Hájek's and Fitting's variants; the literature check for A5₀ and the de re gap; Swinburne's inputs (the primary sources were unreachable from this session's network: users.ox.ac.uk and epsociety.org did not resolve; a web-search summary gave 1/2, 1/2, 1/10 and 1/1000, which do give 100/103 ≈ 0.97, but those are not transcribed from the book and are not used).
2026-11-12 · M2 (part): E14's conjecture "A5₀ implies A5'" is false; when it holds (E15). Prediction written here before the census ran
- How it started. E14 left open why every A5₀ model it found also satisfied A5'. Working on a proof, the argument broke at one step: A5' needs Anderson's NE (the individuals y with hE(y) ∩ Ex meeting every world) to be a union of blocks, and nothing in A1, A2w, A3', A5₀ forces that. The paper construction: worlds {0, 1}; g exists at both, a only at 0, c and b only at 1; blocks {g's row}, k = {(0,a), (1,c), (0,b)}, k3 = {(1,b), (1,a), (0,c)}. Then a and c are Anderson-necessary (k meets both worlds through existents), b is not, yet (0,b) shares block k with (0,a), so NE is not in the algebra. Checked with E14's own
search on that one frame before this entry: one model, A5₀ true, A5' false. It has 8 cells, outside E14's search (at most 6). - Prediction (H20), before the census. (a) In a model of A1, A2w, A3', A5₀, A5' holds exactly when NE is a union of blocks (positivity then follows from A1, since NE₀ ⊆ NE). (b) Under full property semantics A5₀ always implies A5'. (c) No counterexample has at most 7 cells, so the smallest has 8: one world gives NE = NE₀, one individual likewise, 7 cells is 1x7 or 7x1, and E14 covered the rest up to 6 except the full-semantics partition, settled by (b). (d) Two individuals are never enough at any number of worlds (g is necessary; the other is either in NE, making NE every cell, or not, making NE = NE₀). So 4 worlds x 2 individuals has no counterexample, 2 x 4 has some, and 3 x 3 has some (paper example: a at worlds 0, 1, b at 2, blocks {g}, {(0,a), (0,b)}, {(1,a), (2,b)}, {(2,a), (1,b)}). Refuted by any 4x2 counterexample, none at 2x4 or 3x3, or a census that disagrees with E14 at 6 cells.
- Method. Every model's positive set is {f : NecA f x} for any God-like x (the definition of G\*), and with A5₀ every God-like x exists at every world (E14, Lean), so the models of a frame are enumerated exactly by H = closure of a necessary existent's row; each candidate is checked against the axioms' definitions. Lean: (a), (b) and the 8-cell countermodel.
- Outcome (
experiments/e15_a5z_not_a5/results.md). H20 held in all four parts. Census (exact, about 30 s): 4x2, 0 of 7,539 A5₀ models fail A5'; 2x4, 4,608 of 57,413; 3x3, 11,520 of 160,645. The enumeration agrees with E14's definition-level search frame by frame on all 17,440 of E14's frames. Proved in Lean 4 (lean/TheismFormal/Closure.lean, no sorry, no warnings): nea_iff (E13 (b), until now search-only), a5a_iff_closed, full_a5z_imp_a5a (all three use no axioms) and the countermodel cm8 (propext only). Not predicted: the 2-block counterexamples at 2x4 use an individual that exists at no world; with every individual existing somewhere the minimum is 3 blocks (2,592 of the 4,608). Every frame with a counterexample has exactly one A5₀ model. Reading: A5₀ and A5' are independent premises in coarse algebras; A5₀ still gives de re, but it is a different axiom from Anderson's, not a strengthening. Correction added to E14's results and to the README. - Not done. Lean proofs of E13 (a) and E14 (c); Hájek's and Fitting's variants; the literature check for A5₀ and the de re gap; Swinburne's inputs.
2026-11-11 · M2 (part): does plain necessary existence close the de re gap? (E14). Prediction written here before the search ran
- Setting. As E13 (S5, varying domains, rigid positive set, block algebras, Anderson's own A2, actualist entailment and actualist "x has f necessarily"), but A5' is replaced by A5₀: the property "exists at every world" (NE₀, the union of the full rows of the individuals that exist at every world) is positive. E13 traced the de re gap to Anderson's NE, which a coarse algebra lets someone else's cells satisfy; NE₀ cannot be met that way.
- Prediction (H19), proof sketched on paper before any code. (a) De re: in every model of A1, A2w, A3' and A5₀, every God-like individual exists at every world, for any algebra and any domains (a God-like x has the positive NE₀ at its existing cells, x exists somewhere, and a cell of NE₀ belongs to a necessary existent's row, which is x's own). (b) So de dicto holds too. (c) With E13's H and G(H), the models of all four are those H with NE₀ a union of blocks and H ⊆ NE₀. (d) NE₀ ⊆ NE always (an essence f of an everywhere-existent x holds at every cell of x), so A5₀ is the stronger demand on H; but the model sets are not nested: E13's de re countermodels satisfy A5' and not A5₀, and I guess (not derived) that some frame has an A5₀ model where Anderson's NE is not a union of blocks, so A5' fails. (e) Full property semantics: the same models as A5' (one per individual existing at every world). Refuted if the search over E13's 17,440 frames finds a model of all four with a God-like individual missing from some world, a count different from (c) anywhere, or a full-semantics count different from the number of necessary existents. The guess in (d) is refuted if every A5₀ model also satisfies A5'.
- Outcome (
experiments/e14_plain_necessity/results.md). (a), (b), (c), (e) held: the search equals the formula on all 17,440 frames, no model of the four axioms has a God-like individual missing from any world, and under full semantics A5₀ and A5' give the same count in all 70 frames. The guess in (d) was refuted: all 2,516 A5₀ models also satisfy A5', so they are a strict subset of E13's 3,638; the 1,122 lost include all 580 de re failures. Proved in Lean 4 (lean/TheismFormal/Plain.lean): godlike_necessary_plain (uses only A2 and A5₀), de_re_plain, ne0_imp_nea, cm_a5_not_a5z. Not predicted: in the 60 frames (of 1,394 at 2x2, 3x2, 2x3 with an A5₀ model) where Anderson's NE is larger than NE₀ it is still a union of blocks; why A5₀ implies A5' in general is unproved. Reading: de re needs either a separating property space or a premise that says "exists at every world" is positive outright, which is stronger than Anderson's. - Not done. A proof that A5₀ implies A5'; Lean proofs of E13 (a), (b) and E14 (c); Hájek's and Fitting's variants; the literature check for the de re gap and for A5₀; Swinburne's inputs.
2026-11-11 · M2 (part): an actualist "has f necessarily", and whether God-likeness lands on one necessary being (E13). Prediction written here before the search ran
- Setting. As E12 (S5, varying domains, positive set rigid, properties the unions of the blocks of a partition, Anderson's own A2, actualist entailment), but "x has f necessarily" is read actualistically too: box (E!x -> f x), i.e. f contains rowE(x) = x's row ∩ Ex. G\* and essence both use it. This is the reading E12 left open.
- Prediction (H18), proof sketched on paper before any code. Let hE(y) be the union of the blocks meeting rowE(y), and G(H) the union of the full rows of the y with hE(y) = H. (a) Models of A1, A2w, A3' are P = {f : H ⊆ f} for the non-empty unions of blocks H with H ⊆ G(H) and G(H) a union of blocks (every block of H then meets Ex automatically). (b) Every essence of y has existing part hE(y) ∩ Ex, so NE = the rows of the y whose hE(y) ∩ Ex meets every world (and NE(y) is false for a y that never exists); A5' adds that NE is a union of blocks and H ⊆ NE. (c) The de dicto conclusion (at every world some God-like individual exists) holds in every model of all four axioms: each cell of H lies in the row of a God-like individual and of a necessary existent, so some y is both, and the God-like individuals' existing cells are exactly H ∩ Ex. Dropping A5' leaves models where it fails. (d) Full property semantics: one model per individual that exists at every world, as in E12. (e) The de re conclusion (some one God-like individual exists at every world) fails in some models of all four axioms under both readings (E12's possibilist one and this actualist one) but never under full semantics: in a coarse algebra NE(x) only says x's essences are exemplified by someone at every world, and two contingent God-like individuals can share one essence. Refuted if the exhaustive search over every domain assignment and every partition up to 5 cells (6-cell partitions up to 5 blocks) gives a count different from (a) or (b) for any frame, a model of all four axioms where (c) fails, or no de re failure under either reading.
- Outcome (
experiments/e13_actualist_necessity/results.md). H18 held in all five parts: the search's counts equal the formula's on all 17,440 (domain, algebra) frames. De dicto never fails with A5' (3,104 failures without it). De re fails in 580 of 3,638 models of all four axioms under the actualist reading and 382 of 6,706 under E12's possibilist reading; never under full semantics. Proved in Lean 4 (lean/TheismFormal/Actualist.lean): godlike_exists_everywhereA (de dicto, any algebra and domains), de_re_fails (the 2-world, 2-individual, one-block countermodel, both readings, no axioms used), full_godlike_necessaryA. Not predicted: de re failures need two God-like individuals in one block at complementary worlds in every example found. Reading: Anderson's axioms with varying domains give "necessarily something God-like exists", not "something God-like necessarily exists", unless the property space separates individuals. - Not done. Lean proofs of (a) and (b); Hájek's and Fitting's variants; a literature check for a prior statement of the de re gap; Swinburne's inputs.
2026-11-11 · M2 (part): Anderson's system with varying domains (E12). Prediction written here before the search ran
- Setting. As E11 (S5, positive set rigid, properties the unions of the blocks of a partition of the world x individual cells, Anderson's own A2), but each world w has its own non-empty domain, the set Ex of cells (w, y) with y existing at w. Quantifiers are actualist: "f entails g" is box forall y (f(y) -> g(y)) over existing y, so f entails g iff f ∩ Ex ⊆ g. A1, essence and NE use that entailment; "x has f necessarily" stays possibilist (every world's cell of x, as Anderson's G\* reads). The argument's conclusion is actualist: at every world some God-like individual exists.
- Prediction (H17), proof sketched on paper before any code. (a) Models of A1, A2w, A3' (no A5') are P = {f : H ⊆ f} for non-empty unions of blocks H that satisfy E11's condition and in which every block of H meets Ex. (b) A5' (NE positive) binds, unlike in E10 and E11: it adds that H ∩ Ex meets every world (and that NE is in the algebra). (c) In every model of all four axioms the conclusion holds at every world; dropping A5' leaves models where it fails at some world. (d) Under full property semantics there is exactly one model per individual that exists at every world, and none when no individual does. Refuted if the exhaustive search over every domain assignment and every partition up to 5 cells (and 6-cell partitions with up to 5 blocks) gives a model count different from (a) or (b) anywhere, or a model of all four axioms where the conclusion fails.
- Outcome (
experiments/e12_varying_domains/results.md). H17 held in all four parts: the search's counts equal the formula's on all 17,440 (domain, algebra) frames checked (every domain assignment, every partition up to 5 cells, 6-cell partitions with up to 5 blocks). A5' binds in 1,664 frames; without it, 1,612 of 8,534 models have a world with no God-like existent; with it, none. Full semantics: one model per necessary existent, none in the two 2x2 domain assignments where nobody exists at both worlds. Proved in Lean 4 for every algebra and any domains (lean/TheismFormal/Varying.lean: godlike_exists_everywhere, full_godlike_necessary, full_model_of_necessary; no sorry, standard axioms only). Not predicted: whenever "exists" is in the algebra it is positive (all 1,830 such models); it follows from (a) in one line. Reading: varying domains neither rescue nor sink Anderson's argument; they put the weight on A5' and on which individuals the domains contain. - Not done. An actualist reading of "x has f necessarily"; other essence variants; Swinburne's inputs; a CI job for Lean (
.github/, owner-approved path).
2026-11-11 · M2 (part): Lean compiled; E10 proved in Lean; a correction, and Anderson's own A2 (E11). Prediction written here before the E11 search ran
- Lean compiled for the first time. The v4.15.0 release tarball downloads through the proxy (only the elan installer script had been refused before);
lean/build.sh fetches it and builds. Modal.lean and Arguments.lean, written in M1 and never compiled, built unchanged. - E10's theorem in Lean (
lean/TheismFormal/Anderson.lean): forward and converse directions, uniqueness, corollaries (a), (b), (d) and step 4 (A5' never binds), for any block algebra (finite or not). No sorry; the classical results use only Lean's standard axioms (propext, Classical.choice, Quot.sound). - Correction found while checking related work. Benzmüller & Fuenmayor (2020, arXiv:1910.08955) note that Anderson drops half of Scott's A2: he keeps "if f is positive its complement is not" and drops "if f is not positive its complement is". E7 to E10 kept both halves. So their results are about Anderson's definitions of G\* and essence with Scott's A2, a hybrid, and E7's reading "Anderson's repair collapses under full property semantics" does not apply to Anderson's own system; B&F's Nitpick finds a two-world, one-entity countermodel to collapse there.
- Prediction (H16), proof sketched on paper before any code. With Anderson's own A2 (not both f and f' positive) and A1, A3', A5' as before: for an individual y let hull(y) be the union of the blocks meeting its row. A model exists for a non-empty union of blocks H exactly when H is the union of the rows of the individuals y with hull(y) = H; each such H gives exactly one model, P = {f : H ⊆ f}, and every model is one of these. Corollaries: with full property semantics each individual gives exactly one model at every number of worlds (so collapse is avoided, matching B&F); with one individual every algebra has exactly one model; the God-like individual may have contingent properties, and they are simply not positive. Refuted if an exhaustive search over every up-closed positive set (A1) of every partition up to 5 cells gives a model count different from the number of such H for any partition.
- Outcome (
experiments/e11_anderson_a2/results.md). H16 held: search count equals the predicted count on all 573 partitions checked (every partition up to 5 cells; 6-cell partitions with up to 5 blocks). Under full property semantics there is exactly one model per individual at 2 to 5 worlds; in 51 of 52 algebras at 5 worlds x 1 individual the God-like individual has a contingent property. The theorem and the full-semantics corollary are proved in Lean (weak_model_form, weak_model_of, weak_full_semantics_model; the last two use no axioms at all). Reading: Anderson's repair does avoid collapse, and the reason is precisely the half of A2 he drops, not the redefined God-likeness or essence on their own. Corrections added to E7's and E10's results. - Not done. Varying domains (Anderson's intended reading, the Hájek controversy); Swinburne's inputs; a CI job for Lean (it would touch
.github/, an owner-approved path).
2026-11-10 · M2 (part): a theorem for every finite property algebra (E10). Prediction and proof sketch written before the check
- Prediction (H15), from a pen-and-paper proof written before any code ran. In the E7 to E9 semantics (S5, constant domain, positive set rigid, properties a finite Boolean algebra of world-by-individual cells, which is always the set of unions of the blocks of a partition), Anderson's A1, A2, A3', A5' have a model exactly when some block B is a union of whole individuals (every world's cell of each of them, and nothing else); each such block gives exactly one model (positive = "contains B"), so the number of models of an algebra equals its number of such blocks. Corollaries: A5' never binds; G* is the individuals in B, and they have no contingent property; and no property of the form "p holds" with p true at some worlds and not others is in the algebra, so every proposition the algebra can express is non-contingent. Refuted if the exhaustive search of E8/E9 gives a per-algebra model count that differs from the block count for any partition it covers (all partitions up to 5 cells; 6 cells up to 5 blocks), or if a model has an empty G*.
- Outcome (
experiments/e10_theorem/results.md). H15 held. The search's model count equals the theorem's on all 563 partitions checked (every partition of up to 5 cells at 9 shapes, and 202 of 203 partitions each at 3x2 and 2x3; the remaining 6-block one is the full-powerset case, settled by the proof: no model). No model had an empty G*, and no model's algebra contained a contingent proposition. The proof covers every finite model, so the E7 and E8 one-individual nulls now hold at all sizes, not only up to 5 cells. Reading: in finite models Anderson's repair avoids modal collapse only for propositions its property space cannot express. Not yet in Lean; literature check for a prior statement of this characterisation still open. - Side fix:
test_theism.py's __main__ block sat mid-file, so python3 test_theism.py skipped the E7 to E9 tests; moved to the end (they all pass).
2026-11-10 · M2 (part): E8 at six cells (E9). Prediction was in the script docstring, not the log, before the run
- Prediction (H14). Multi-world models of Anderson's axioms still exist at 3 worlds, in algebras where one individual has only rigid properties; refuted if none at 3 worlds x 2 individuals or 2 x 3 (partitions of at most 5 blocks; the 6-block case has 2^32 positive-set choices and is skipped).
- Outcome (
experiments/e9_restricted_larger/results.md). H14 held. 3 worlds x 2 individuals: 8 of 200 contingent algebras give a model; 2 x 3: 39 of 197. The first multi-world models at 3 worlds, so the E8 result is not a two-world accident. In all three listed 3x2 examples one individual's cells form a single block (rigid) and the other is split. Still a search, not a proof; the full-powerset 6-cell case is not covered. - Lean. Tried again to install the Lean toolchain (elan script downloads, running it was refused by the session's permission policy). Lean files remain uncompiled; needs a CI job or a session allowed to run the installer.
2026-11-07 · M1: catalogue, checker, ledger (predictions written before each run; outcomes below)
- Predictions (E1). Every catalogue argument valid in its intended logic (H1); the modal ontological argument valid in B and S5 but not in T or S4, its reverse valid in T (H2); the Catholic and Orthodox positions each a maximal consistent subset of the eight-claim church set, differing in two claims per dispute (H3).
- Outcome (E1,
experiments/e1_arguments/results.md). All three held. Ten arguments are valid as stated; the free-will-defence set is satisfiable (Mackie's argument is valid but needs "no morally sufficient reason exists", which Plantinga's set shows is not a logical truth). Ontological: K, T, S4 invalid with two-world countermodels; B, S5 valid. Reverse: valid from T up. The church set has exactly two minimal inconsistent sets (Honorius: c1–c4; Florence: c5–c8) and 16 maximal consistent positions; the Catholic position drops c4 (Honorius spoke ex cathedra) and c8 (the Filioque is false), the Orthodox drops c1 (papal infallibility) and c6 (Florence was ecumenical). - A first version of the monadic checker enumerated predicate types and did not finish on the eight-claim set in 120 s; replaced by grounding over a #constants + #quantifiers domain (exact for unnested quantifiers) and backtracking search, 0.8 s for all of E1.
- Predictions (E2). Unwin's 67% reproduces from 50% and his six factors; the resurrection item carries over 90% of the variance of the Christianity-vs-rest log odds (H4); the prior carries more of the theism-vs-naturalism variance than any one item (H5).
- Outcome (E2,
experiments/e2_ledger/results.md). Unwin: 0.6667. H4 held: resurrection 98.5% of the variance; the Bayes factor it needs for P(Christian) > 0.5 is 7.3 at P(theism) = P(Christian | theism) = 0.5 and ranges from 0.5 to 294 over the 3 × 3 prior grid, all far below the 10^44 reported for the McGrews and far above a sceptic's 1. H5 held: prior 54%, fine-tuning 31% (a swept range), every sourced item under 6%. - A design error caught before reporting. The first E2 took midpoints of every range, so the resurrection item sat at 10^22 and gave P(naturalism) ~ 10^-21: an artefact of a uniform over 44 orders of magnitude. The second error was mine too: the first church-level sweeps leaned against Catholicism (three of five ranges negative). Both fixed: the resurrection item is now reported as a threshold, not averaged, and every church range is symmetric about zero until elicited (M3).
- Not done. Lean files are written but not compiled: the session's permission policy refused downloading the Lean toolchain. Swinburne's and Shermer's inputs are not transcribed (the sources were not reachable), and the McGrews' 10^44 is a secondary report.
2026-11-07 · M2 (part): leave-one-out (E3) and Gödel–Scott modal collapse (E4) (predictions before running)
- Predictions (E3). Dropping any one sourced theism-level item moves the midpoint P(naturalism) by under 0.1 absolute (H6); dropping fine-tuning (a swept range, not a published figure) moves it most, and the sign of "is P(N) > 0.5 at the midpoints" does not change under any single drop except fine-tuning (H7). Refuted if any sourced item alone flips the verdict.
- Predictions (E4), sharpened before the run. Scott's axioms are A1 (positive properties are closed under necessary implication), A2 (exactly one of a property and its complement is positive), A3 (God-likeness G is positive), A4 (positivity is necessary, built in here by making the positive set rigid across worlds), A5 (necessary existence NE is positive), with G, ess and NE defined as in Scott/Sobel. Exhaustive search over S5 models with full property semantics (every function from worlds to subsets of the domain is a property; world x domain size at most 5): (H8) models exist, with every positive property possibly instantiated; (H9) every model has exactly one world, because with full property semantics the collapse p -> []p must hold for the constant property of every proposition, and a proposition true at one world and not another exists whenever there are two worlds. Refuted by a model of A1-A5 with two or more worlds.
- Outcome (E3,
experiments/e3_leave_one_out/results.md). H6 and H7 both refuted. At the midpoints P(N) = 0.76; dropping natural evil gives 0.09, moral evil 0.42 (a published item) and hiddenness exactly 0.50 (boundary); dropping fine-tuning gives 0.99. So three items each decide the midpoint verdict, and the variance-share ranking (prior, fine-tuning first) does not predict this: variance measures spread over the ranges, leave-one-out measures what the central estimate leans on. Both views go in the paper. Church-level items do not touch the naturalism contrast (change 0.000), as they should. - Outcome (E4,
experiments/e4_godel_scott/results.md). H8 and H9 held within the sizes searched: one-world models number one per individual (1, 2, 3 for domains of 1, 2, 3, the principal ultrafilters, an analytic check), and none exist for 2, 3, 4 or 5 worlds (and 2 worlds with 2 individuals). Limits: exhaustive only up to 5 (world, individual) cells; this is a search, not a proof (Sobel's derivation is the proof; the Lean version is still to write). - Still open in M2. Lean not compiled (no toolchain access here), Swinburne's and Shermer's inputs not transcribed, Anderson/Bjordal repairs of Gödel's argument not yet checked.
2026-11-07 · M4 (part): shared-cause dependence (E5) (predictions before running)
- Prediction. With the testimony items sharing one latent reliability, the omega needed to bring the resurrection Bayes factor under the 7.3 threshold at even priors is below 10^-30 (H10); refuted if above. E2's rho = 0.5 discount would be far stronger than any plausible omega.
- Outcome (
experiments/e5_shared_cause/results.md). H10 held: the threshold is omega = 10^-43.2 at even priors, 10^-41.5 at sceptical (0.1, 0.1) priors, and none is needed at devout (0.9, 0.9) priors. The closed form BF = 1 + omega (L - 1) matches enumeration of the explicit joint model in three cases. The equivalent rho at the even-prior threshold is 0.024. - Reading. Shared reliability shifts log BF by log omega, once, not by a factor per item, so it cannot by itself defeat a product of 10^44; the live dispute is the per-item ratios behind L, still unsourced. The 10^44 is a secondary report.
- Context: the session's data-bound draw (
build-data) was skipped because all three feed hosts were refused by the proxy; this programme was redrawn.
2026-11-08 · M2 (part): Shermer's inputs (E6) (prediction before running)
- Prediction. Run through Unwin's formula, Shermer's six published D values give about 2% as the column says (H11); refuted if the computed value differs from 0.02 by more than rounding (say, outside 1.5% to 2.5%).
- Outcome (
experiments/e6_shermer_check/results.md). H11 refuted. The D values multiply to 2.5e-4, giving 0.025%. For 2% they would need to multiply to 0.0204. No single D changed alone is a plausible fix (each would need 8 to 82). Unwin's 67% reproduces, so the formula is read correctly. Added a test and updated the ledger note. - Not done. Swinburne's inputs (book unreachable), Lean compilation (toolchain refused), per-item leave-one-out already done in E3.
2026-11-09 · M2 (part): Anderson's emendation of Goedel's argument (E7) (prediction before running)
- Prediction (H12). Anderson's (1990) emendation replaces Scott's God-likeness with G\*(x) iff for every property f, f is positive exactly when x has f necessarily, and replaces "ess" by the biconditional form (box psi(x) iff box forall y (f(y) -> psi(y))). Same search as E4 (S5, constant domain, full property semantics, positive set rigid, A2 by construction). I predict models with two or more worlds exist (the collapse is avoided, as Anderson claims), with every positive property possibly instantiated. Refuted if no multi-world model exists at any size searched (up to 5 cells). A null at this size is weak evidence, as in E4.
- Outcome (E7,
experiments/e7_anderson/results.md). H12 refuted within the sizes searched. One-world models number one per individual (1, 2, 3 for domains of 1, 2, 3), as in E4, and there is none for 2, 3, 4 or 5 worlds (or 2 worlds with 2 individuals). So under full property semantics Anderson's repair collapses at these sizes just as Scott's does. This does not contradict Anderson: his consistency result uses a restricted (non-full) property space, which this search does not model, and the search is exhaustive only to 5 cells, not a proof. What it does say is that the repair needs a restricted property domain, and the choice of that domain is where a dispute would sit. Next: search restricted property families (e.g. properties closed under a given algebra) and write the Lean version.
2026-11-10 · M2 (part): Anderson's repair over restricted property algebras (E8) (prediction before running)
- Prediction (H13). Restricting properties to a Boolean algebra (unions of blocks of a partition of the cells) lets some algebra with contingent properties carry a multi-world model of Anderson's axioms; refuted if none does at any size up to 5 cells.
- Outcome (
experiments/e8_restricted/results.md). H13 held, narrowly. Every partition of the cells was tried. With one individual (2 to 5 worlds) no algebra that has a contingent property gives a model, so the collapse persists there. With 2 worlds and 2 individuals, 2 of 13 contingent algebras give a model (e.g. blocks {(w0,x0),(w1,x0)}, {(w0,x1)}, {(w1,x1)}: x0 gets only rigid properties, x1 is distinguished at each world). Algebras where every property is rigid were skipped, since they are one-world models in disguise. Reading: the repair survives only when the property space is cut so that the God-like individual has no contingent properties while others do; that choice is exactly what a critic can contest. Limits: ≤5 cells, constant domain, S5, partition algebras only; not a proof, and the examples are toy sizes. Next: the Lean version, and Swinburne's inputs.
Paper, code and results
More in History & human sciences
- How do US Catholic subpopulations differ in fertility, retention, Mass attendance and politics, and which demographic channels are reshaping American Catholicism?Research programme: comparative demography of US Catholic subpopulations from the GSS 1972-2024 with design-based intervals; the Catholic fertility premium has reversed among white Catholics born 1965-80 (-0.28 children vs Protestants), white Catholic retention fell to 55% for those born 1981-96 while Hispanic retention held near 71-76%, and Hispanics are now 37% of US Catholics…
- Can a Catholic's or Protestant's other survey answers predict the beliefs on marriage, sexuality and devotion they were never asked, and do those associations hold for immigrants and across traditions?Research programme: in GSS 1972–2024, a respondent's other answers predict an unasked belief or practice item at a median AUC 0.21 above the decade prior (Catholics: weekly Mass 0.86, premarital sex 0.85, same-sex marriage 0.88, abortion 0.75, divorce law 0.69); models trained on third-generation Catholics rank foreign-born Catholics almost as well but miss their levels by up to 11 points
- How did religion develop, and what does the evidence from cognitive science, archaeology, history and theology actually show?A heavily cited synthesis of the evidence on how religion developed, from the cognitive science of belief and Palaeolithic burials through Göbekli Tepe, the first states, the Axial Age, Israelite monotheism, the Big Gods debate, early Christianity and the East–West Schism to 2020s data, read through scientific, historical and Catholic and Orthodox theological lenses, with what replicated and what did not
All 4 History & human sciences projects →
Other research topics
Proper scoring & elicitation28Effort & peer prediction9Verification games & slashing19Detection & tolerances15Markets as routers13Prediction-market design9Paying contributors8Decentralised training systems23Digital twins & sim-to-real164LLM chip design loops40Robot manufacturing loops53Ambient AI assistant46Self-replicating manufacturing34Simulation-first design64Engineered clothing45Task-specific AI accelerators40Algorithmic chip-design loops42Open-source purpose-built robots50The software-engineering loop43Adversarial human imitation41Open-source residential construction41Autonomous multi-angle capture35Sensing, imaging and 3D33Energy supply and dense energy storage41Robot societies among people32Agentic co-design of low-cost robots28Laundry-hanging robots26Machine learning on qualitative data23Robots on the building site28Civilizations & group psychology3History & human sciences4
Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.