W1. An error in the Cauchy derivation of P = e−S from A2+A3+A4,
    or a demonstrable counterexample against the axioms themselves
W2. The existence measure turns out non-normalisable: the Kraft conditions
    demonstrably violated, or Weil–Petersson volumes divergent
W3. A paradox in the framework that does not decompose into two layers
    (the both-and discipline of the philosophy page fails)
W4. The causal dichotomy (Theorem 10.4) leaks: a third option
    for actuality that is neither physical nor indexical

W3 deserves emphasis: the claim "every apparent contradiction is a layer confusion" is a methodological promise. One paradox that demonstrably needs both halves at the same level breaks it.

F1. Rokhlin or modular-invariance arguments turn out invalid
F2′. Cgauge from first principles ≠ value α⁻¹≈137 requires
    TRIGGERED Aug 2026: own 2-loop computation gives 11% deviation
    (old window ζ'~10³–10⁴ already refuted earlier, July 2026)
F3. GW background: smooth continuum, non-tensor polarisation,
    or > 42 discrete modes
F4. Discovery of a 4th chiral fermion generation
F5. Unwanted holonomy structures in precision observations

F2 was the most important criterion: the only one that yielded a number. Status August 2026: F2′ has been triggered. The framework's own 2-loop computation from the spectral unification boundary condition gives α⁻¹(m_e) = 152.2 against a measured 137.04: an 11% deviation, outside the framework's own criterion of ~10%. The α derivation is hereby retracted; not weakened, not reformulated. The criterion did its job twice: first the window fell (F2, July), then the derivation itself (F2′, August). What the computational programme delivered afterwards is in section 2 below.

V1. Convergence conjecture (11.7): simple laws that demonstrably
    admit no self-representation
V2. Effective dominance: a landscape sector where several
    unrelated configurations carry comparable μ-weight

The full computational programme of this section has also been written up as a scientific paper, published on this site: Selecting the Standard Model (Genis 2026). This testing page serves as its living prediction registry.

1. Unification route (F2′): SM content + spectral boundary condition
   → predicted α⁻¹(me) = 152.2; measured 137.04: 11% off.   FAILED
2. Selection route (F1*): the cheapest consistent geometry does not
   yield the observed forces, under all three frozen
   complexity measures.                                       FAILED
3. Weight-function route (f₀): EEML's e⁻ˢ as cutoff function gives
   parameter-free αU⁻¹ ∈ [1.3, 5.1]; required ≈ 46: factor 9.  FAILED

The pattern is consistent: countable structure (which charges exist, their ratios, the mixing angle 3/8 at the unification scale) does come out every time; continuous strengths do not. In EEML language: the strengths belong to the address (where we live in the distribution), not to the law. This completes the course set in v0.14: the three-layer never promised a point value for α, and three falsifications now confirm that quantitatively.

After the falsifications the selection question was posed again, with a condition EEML's indexical layer itself supplies: we observe, so the selected geometry must be able to carry observers. Concretely, and frozen in advance: an electron must be able to exist (light, uncoloured, charged), charged nuclei must be able to exist, and the observed forces must be reachable. The result, in steps each documented separately in the working documents:

• Of all cheaper candidate geometries, not one can make an electron. The bottleneck is the same every time: no light charged particle without colour. On the standard-model algebra the search reconstructs exactly the known particle content of one generation, including the correct hypercharges (1/3, −4/3, 2/3, −1, 2): not put in, but solved from the consistency conditions.
• One class of unification geometries (Pati-Salam type, with the lepton as a fourth colour) is cheaper and seemed to win. But that class mathematically forces the bottom quark and the tau lepton to have the same coupling at high energy, and the measured masses contradict that by about 35%. A theorem from the exact structural conditions shows the whole class shares this prediction, without exception and without an escape route in the explored universe. The measurements thereby eliminate the entire class.
• Because neutrinos have mass (measured since 1998), the bare standard-model content also drops out: it has no room for neutrino mass.

classified catalogue + anomaly freedom
+ "an electron must be able to exist"
+ "neutrinos have mass" + "bottom ≠ tau at high scale"
= the standard model with right-handed neutrinos, and nothing else

Within the explored universe of finite geometries (declared limits: at most three building blocks, matrix size up to four, bounded diagram size) this is the only candidate that passes all gates: mathematical consistency, observer conditions and the measurements. Not assumed, but left standing.

Guillotine prediction: neutrinos are Majorana particles. The right-handed neutrino sector the selection enforces admits a Majorana mass, and the structure requires it in three independent ways (symmetry breaking, coupling unification, the exact first-order condition). If experiments definitively rule out neutrinoless double beta decay (LEGEND-1000/nEXO class), this entire chain falls.

A window that fits: the heavy Majorana scale that yields the measured neutrino mass (~1.4·10¹⁴ GeV) lands in the middle of the independently predicted window of one and a half decades between the intermediate scale and the unification scale. Furthermore: no tree-level proton decay, no new gauge bosons within collider reach. Modest, but testable.

What this is not: not a derivation of α, nor of the masses themselves: the continuous numbers remain address. The challenger class can return if someone demonstrably computes threshold corrections above 13%; that route is documented and open. And the question of why there are three generations is not answered, but it now has a price: see the card below.

Generations: the number 3 is address, with a price tag. The cheapest valid observer address turns out to be one generation plus two right-handed neutrinos (leptogenesis supplies the required CP violation there; nothing anthropic needs the muon or the tau). Our measured three-generation address is exactly 10 bits more expensive, weight 1 in 1024. Any future explanation of "three" must therefore deliver exactly 10 bits; until then this stands honestly noted as the weakest spot of the address reading.

The dynamical cost term cannot supply those 10 bits. The naive vacuum reading of S_dyn would charge ~10²²⁸ bits per generation difference and is thereby refuted by our own existence; under the renormalised reading S_dyn is flat across the candidates. The complexity ranking above is thus the complete selection, and there is no hidden route to a point prediction through the dynamics.

Robustness. The outcome also stands in the extended universe with SU(5) embeddings (136 additional algebras, the lepton as a fifth colour): all 25 survivors carry the forced bottom-tau equality, the SU(5) cases additionally proton decay (Super-Kamiokande); no separate-lepton escape route exists there, and four-building-block algebras are analytically too expensive.

Typicality of α. Within the habitability window from the literature (α⁻¹ roughly between 59 and 685, from proton stability and galactic cooling; Adams, Sandora) the measured value sits at log position 0.34: comfortably interior, no edge. The narrow carbon window (±4%, Oberhummer et al.) costs the conditioning ~5 bits. Total address price of the framework across the two best-known open numbers: ~15 bits. Mild, quantified, and exactly what a distribution-plus-indexicality reading must be able to carry.

Extension (12 August): the agenda computed. Six follow-up phases, each frozen in advance. New on the law side: the spacetime dimension (3,1) is a singleton under the observer gates (orbits, complexity, predictability; classic arguments by Ehrenfest and Tegmark, here with status precision), and complex quantum theory follows from the same composition axiom that gave P = e⁻ˢ (real-valued quantum theory is experimentally excluded precisely under that composition assumption; Renou et al., Nature 2021). New on the prediction side: no WIMPs or other new coupled particles, and, after a documented internal conflict between two same-day registrations, the combined ledger that counts K-costs and typicality costs as one currency: it predicts a θ-solving, axion-like sector (the measured strong-CP zero would cost 34 bits as address; structure solving it costs only 10-20), giving dark matter a candidate explanation and making ADMX-class experiments referees in the framework's favour. On the address side: the cosmological constant priced at 3.5-9 bits (band over both literature windows, no choice after the fact: the best-known edge case remains an edge case). And the arrow of time is formalised as a measure preference for compressible initial conditions [motivated]. Full balance: working documents, phases 22-27.

The decision tree, frozen before nature chooses. Whether the Majorana guillotine can ever fall cleanly depends on the neutrino mass ordering, which will be measured in the coming years (JUNO). Hence, fixed in advance: detection of neutrinoless double beta decay = confirmed, regardless of ordering; inverted ordering plus a fully emptied band (~2035-2040) = chain refuted, publication obligation; normal ordering with the funnel = honestly undecidable, and that will not be massaged into something weaker. Also declared in advance: current cosmology pushes towards that difficult third branch, and that risk was accepted when the prediction was made.

And then the most elegant step: the framework places a weighted bet on the outcome. In the address reading the neutrino parameters are generic draws from the measure, and that idea already exists, tested, in physics: the anarchy scenario, which correctly predicted the large mixing angle θ₁₃ before its 2012 measurement. Three frozen predictions follow: the mass ordering is normal (95-99.9%; an inverted outcome costs 4-10 bits, booked in advance), the neutrino mass sum peaks near ~0.06 eV (where cosmology already points), and the Majorana mass m_ββ is small but not zero (~1-10 meV): null results at the next generation of experiments are therefore the expected outcome, without undermining the Majorana claim, with detection shifted to future sensitivity. Two of these statements (ordering, mass sum) are decidable this decade and are fixed before anyone knows the answer.

The objection answered here: "you chose gates that smell of our own world (an electron, charged nuclei), so the uniqueness of the standard model is gate-relative." Three answers, in ascending strength.

1. Reduction: three gates become one. The real gate is abstract and chemistry-free: G* = the member contains subsystems that can carry and process information stably, rewritably and durably (self-representation, exactly what the reading rule already asked). The concrete gates follow from it as a theorem sketch: many localisable bits require bound structures with a hard carrier; rewritability requires writing costs far below the carrier's binding energy, hence a second, soft binding layer; a soft layer over a hard one requires an unbroken long-range force with a light carrier not tied to the hard colour scale: a light, uncoloured, charged fermion, plus heavy charged anchors. The electron and the nuclei are thus not taste but the minimal implementation of "memory can exist somewhere".

2. Anchoring: the one remaining gate is self-validating. What remains as an assumption is: "at least one member passes G*". But every question about the justification of that assumption is asked by a system that itself carries and processes information stably: asking the question is the proof. This is the only assumption in the entire framework whose denial is performatively self-refuting: you cannot doubt it from an address that violates it. A Cartesian anchor, not a choice.

3. The back door, with three locks (addendum the same evening). The derivation in step 1 leaves one exotic route open: gravitationally bound nuclear matter (neutron-star-like substrates) has a marginal scale separation of its own (~10⁻³) and could in principle carry records without an electron. That route is now locked with three independent locks, each motivated. Lock 1, formation: such substrates only arise through gravitational collapse, and collapse requires gas to radiate its energy away; radiative cooling scales as 1/mass², so without a light charged particle a world cools a factor ~10⁶ more slowly (the classic Rees-Ostriker cooling criterion): the gas stays pressure-supported and the loophole world cannot build its own hardware. Lock 2, thermodynamics: even a given substrate is an isolated, monotonically cooling ball with a finite and shrinking free-energy budget; rewriting costs MeV per bit there, and no soft-photon economy surrounds it to drive computation sustainably: the memory freezes. Lock 3, the measure itself: even someone forcing locks 1 and 2 is left with substrates many orders rarer than chemistry (which arises for free on every cold planet); under the typicality reading a question-asker then overwhelmingly lives in the abundant class. Status: locked [motivated], not absolutely excluded [honest]: whoever wants to break the uniqueness of the standard model must force all three locks at once, and now knows their names.

Quantum 1: objective collapse will never be found. The Born derivation rests on unbroken unitarity; collapse models (GRW family) add parameters without gate necessity. Prediction: macro-superposition experiments (molecule interferometry, levitated nanoparticles, MAQRO class) keep confirming unitarity, however large the masses. [derived, conditional]

Quantum 2: quantum computers have no fundamental scaling limit. Same ground, and a real stake: a serious camp (around 't Hooft and the superdeterminists) predicts error-corrected quantum computers will fail at scale due to new physics. We take the other side: every failure is engineering, never natural law. Testable this decade against the scaling curves. [derived, conditional]

Cosmology 1: dark energy is a constant; the DESI hint of evolving w will fade. Current data suggest evolving dark energy at 2-4σ. Our Bayesian logic says: a quintessence field is extra structure without observer necessity (the same argument that struck the WIMPs), and Λ is an address coordinate, not dynamics. Prediction: the anomaly goes the way of most 3σ anomalies. A confirmed w ≠ −1 costs the framework its address reading of Λ. [motivated]

Cosmology 2, the boldest: no primordial tensor modes within reach. The smooth initial state is for us a measure preference (every history compresses to its simplest slice), not the product of an inflaton field: an inflaton is K-cost without gate necessity. Prediction: CMB-S4 and LiteBIRD find r below their detection limits. A measured primordial B-mode falsifies this reading of the arrow-of-time card. [bold, motivated]

Neutrino: the CP phase is generically large. The same anarchy reading that predicts the ordering says δ_CP is a generic draw, hence far from zero. DUNE and Hyper-K will measure it; the current hint (δ near −π/2) sits well. [motivated]

Meta: the axiom price list is stable. The measured 70.2% choice-dependence of mathlib stays within ±5 points as the corpus grows: axiom prices are properties of mathematics, not of the sample. Repeatable yearly with the archived scripts. [measurable]

Structure 1: the proton does not decay. This already appeared as an aside in the chain analysis and is now promoted to a headline prediction. The selection chain eliminated precisely the embeddings that carry proton decay (Pati-Salam via the bottom-tau measurement, SU(5) via the same theorem); what remained is the Standard Model with right-handed neutrinos as the final structure, without decay-mediating bosons. Hyper-Kamiokande (from 2027) is the referee. Failure condition: a single confirmed proton decay breaks not just this stake but the entire selection chain of section 2. [derived from the chain]

Structure 2: no light sterile neutrinos exist. The chain contains exactly three heavy right-handed neutrinos, and their scale falls in the independently predicted seesaw window. A fourth, light and sterile neutrino is extra structure without gate necessity: the same Bayesian class as the WIMPs and quintessence. Prediction: the remaining anomaly hints (reactor, gallium) evaporate; MicroBooNE has already deflated the MiniBooNE hint and JSNS² and IceCube will finish it. Failure condition: a confirmed light sterile neutrino. [motivated]

Structure 3: no Lorentz violation, at any scale. The (3,1) signature is a singleton under the observer gates in the chain; a preferred frame or energy-dependent speed of light is extra structure on top of that selection: it costs bits and buys no gate. Prediction: GRB time-delay measurements (CTA, Fermi successors) and clock anisotropy tests keep giving exactly zero. Failure condition: a confirmed energy-dependent photon arrival or frame anisotropy. [motivated]

Sharpening: "δ_CP is large" now has numbers. Under the anarchy measure (Haar on U(3)) the CP phase is uniformly distributed and independent of the angles. Frozen today, analytically plus Monte Carlo with fixed seed (script anarchie_delta_cp.py): P(|sin δ_CP| > 0.5) = 66.7%, P(|sin δ_CP| > 0.25) = 83.9%, median of the Jarlskog invariant |J| = 0.024 with the measured angles, P(|J| > 0.02) = 59%. Track-record agreement, fixed now: should DUNE and Hyper-K eventually measure |sin δ_CP| < 0.25, that counts here as a miss, even though it is formally a 16% outcome and not a hard falsification. The current experimental hint (δ near −π/2) sits in the fat part of the distribution. [derived within the anarchy reading]

Why this stake exists. The four failed predictions of this programme (F1*, F2′, f₀ and the Zipf demo) share one root: each time the a-local measure was put in a local chair, as boundary condition, as cutoff function, or as empirical law inside a member. The formal face of it: K counts bits, so e⁻ᴷ can rank countable structure alternatives, but singling out a point from a continuum costs infinitely many bits. Continuous constants are therefore address, not law. That boundary was formulated after the α falsifications and could thus be after-the-fact rationalisation. So it now becomes a frozen stake itself.

The meta-stake. No first-principles derivation of a dimensionless continuous constant of nature (α, mass ratios, mixing angles as continuous values) from pure structure selection will ever succeed, by anyone, inside or outside EEML. Does not count as a counterexample: a reduction to other continuous constants (that merely shifts the address) or a framework that itself introduces new continuous parameters. Does count: a parameter-free derivation accepted by the community as first-principles. Failure condition: one such accepted derivation falsifies the address reading of this framework, and with it half the paper. [meta-stake, frozen]

What this stake buys. The diagnosis of our own misses now cuts both ways: it explains why precisely the continuous routes broke, and it loses its credibility if anyone breaks through the boundary. The relation to history is strict: the many α derivations in the literature (from Eddington to today) have all fallen or proved numerological, and this framework has added three of its own. The stake says this pattern is not bad luck but structure.

The level-clean route that remains. The local/a-local distinction rules out the measure as deriver of continuous constants, but points to one untried deriver: the local dynamics itself. If the renormalisation-group flow of this member has a fixed point, that is a discrete, structural object, and whatever it fixes or bounds is law, not address. This programme exists (asymptotic safety) and holds one genuine pre-registration trophy: the 126 GeV Higgs mass, predicted in 2009 from exactly the structure our chain independently selected: Standard Model up to the Planck scale, no unification.

The stake, weak and precise. Continuous address coordinates are bounded by local fixed-point consistency, not fixed. For α concretely: the SM-only upper bound on the abelian coupling (Eichhorn-Versteegen, ~35% above the measured value) translates to α⁻¹ ≳ 75-90 (two frozen translations, script as_venster_krimp.py); the habitability window [59, 685] shrinks from below, a gain of 0.15-0.27 bits in the address ledger. Small, but principled: the first compression of the address by local physics rather than by conditioning. EEML's own contribution: e⁻ᴷ gives members with self-bounding dynamics more weight, so typicality expects us to live in such a member.

Edges and failure conditions. The strong form (unique fixing of α) is already empirically dead for SM-only: the interacting fixed point would put α ~35% too high, and the measurement sits below it. Via a GUT embedding unique fixing is possible in the literature, but that route is closed for this framework by its own selection chain (Theorem S′). Failure conditions of the weak form: (i) an accepted fixed-point computation pushing the upper bound below the measured value: then the bound reading is dead; (ii) should the fixed-point programme itself fall (truncation and BRST criticism is live), this stake lapses without damage to the address reading; (iii) an accepted unique fixing triggers the failure condition of the meta-stake above, not of this one. The two stakes cover each other's flanks: exactly one can win. [motivated, conditional on the fixed-point programme]

Where it comes from. A systematic top-10 of all routes from χ=24 to α (prompted by an external proposal to derive α via threshold integrals) singled out one route with an attainable goal: F-theory on elliptically fibred K3, where the 24 literally becomes 24 seven-branes on a sphere and the string coupling τ is itself geometry (Dasgupta-Mukhi 1996). Own enumeration, conventions frozen in the script header (vriespunt_24.py), with a calibration gate set in advance: the four known orbifold solutions (Z₂, Z₃, Z₄, Z₆) must roll out of the count, and all four do. Result: 56 constant-coupling configurations; 1 leaves the coupling free (Sen orientifold: a modulus, hence address), 9 freeze it at τ = i, 46 at τ = e^iπ/3. The menu of freezable values is {i, e^iπ/3}: two points, both fixed points of the duality group SL(2,ℤ), both order-1 coupling.

The stake. Discrete topological data never fixes a continuous coupling at a generic value. What it can do is two things: leave the coupling free (then it is a modulus: address) or lock it onto a fixed point of the duality group (then it is law). Anything in between, such as α⁻¹ = 137.036, is structurally not on the menu. Failure condition: one explicit, consistent construction in which discrete topological data, with no remaining continuous moduli, fixes a coupling at a non-symmetric point. That would falsify this stake and expose a flank of the meta-stake. [meta-stake family, frozen]

By-catch (route 2): the dilaton-free part is law. Differences between couplings at equal Kac-Moody level are dilaton-free and exactly computable: at the freeze point, Δ_a − Δ_b = +0.68·(b_a − b_b) (script drempel_verschillen.py; modular invariance gated numerically, η(i) calibrated to six decimals against the closed form). And the threshold kernel (the determinant of the torus Laplacian) is globally maximal at exactly ρ = e^iπ/3: the same point where 46 of the 56 freeze configurations sit. Two independent computations point at the same chair. Three flanks now covered: the meta-stake (nobody derives the value), the delegation stake (local consistency may bound), the freeze-point stake (discrete data may only fix symmetric points).

Test result (same day): the flux test. The stake was immediately pitted against its strongest known counter-candidates: flux vacua and the attractor mechanism, where free discrete numbers (flux quanta, black-hole charges) do fix moduli at generic values. Own enumeration (flux_toets.py, the rigid-CY standard model W = Aτ + B with Gaussian fluxes, tadpole ≤ 30, verdicts frozen in advance): the flux menu counts 266 distinct τ values, 99.6% of them outside every symmetric-point neighbourhood, and ρ is even unreachable in principle for this Gaussian-rational menu; the attractor witness gives 27 of 28 horizon values generic. The broad reading "discrete ⇒ symmetric point" is thereby refuted for free data, exactly as the frozen verdict foresaw. The refinement, booked explicitly (no silent rewording): forced discrete data (non-choosable topology, such as χ = 24) only fixes symmetric points; free discrete data (fluxes, charges: a choice from a menu) fixes nothing but quantises the address: every flux choice is a different member of the catalogue. By-catch for the ledger: a flux-chosen address costs finitely many bits (this menu: 8.1 bits; the IIB landscape anchor: log₂ 10⁵⁰⁰ ≈ 1661 bits): the continuum objection softens, in this class, to a countable price. [tested within frozen conventions]

The Sen-branch observation. The forced freeze-point menu is order 1 (τ = i or ρ); weak coupling, by contrast, is effortlessly reachable for the free menu (1,008 configurations with Im τ ≥ 10 in the count). We measure α⁻¹ = 137, far from order 1: within this picture our member demonstrably lives on a free branch, where the coupling is address. The smallness of α thereby itself testifies for the address reading.

The moduli bookkeeping of the menu (addendum, same night). For each configuration we counted how much continuous address remains: the positions of the branes on the sphere, modulo the sphere symmetry (vriespunt_moduli.py; only the constant-τ class counted, the Kähler sector honestly left outside). Range 0 to 9 moduli, median 3: even on the freeze branches address almost always remains. But exactly four configurations are fully rigid (2×III* + I₀*; IV* + II* + I₀*; IV + 2×II*; and 3×IV*, the Z₃ orbifold), and all four sit at a symmetric point. "No remaining continuous moduli" thus occurs in this class exclusively at a fixed point of SL(2,ℤ): maximal freezing exists, and where it exists, it is law. [counted within frozen conventions]

1. Primordial non-Gaussianity: f_NL consistent with zero. The twin of the tensor-mode stake: the smooth initial state is a measure preference, not field dynamics, hence no measurable non-Gaussianity either. Referee: SPHEREx (this decade, precision σ(f_NL) ~ 1). A confirmed detection hits the arrow-of-time reading as hard as B-modes would. [motivated]

2. The HL-LHC finds no new physics. No SUSY partners, no compositeness, no extended Higgs sector through ~2035: SM + ν_R is the final structure. Failure condition: any confirmed new coupled particle beyond ν_R. [derived from the chain]

3. The Hubble tension resolves without new physics. Early dark energy and extra relativistic species are structure without gate necessity (same Bayesian class as WIMPs and quintessence). Prediction: the tension ends in systematics and local effects. The boldest stake of this series: it is currently a genuine ~5σ tension. [motivated, bold]

4. The pulsar-timing background is astrophysical. The NANOGrav/IPTA gravitational-wave background resolves into supermassive black-hole binaries; no cosmic strings, no phase-transition component. [motivated]

5. The neutron EDM stays null. Follows from the axion-like sector driving θ̄ dynamically to zero: n2EDM and successors keep returning null results. [consequence of the axion stake]

6. The axion window, now with numbers. The bit accounting of the axion stake has been computed (script axion_venster.py, conventions frozen in the header): the axion as all of dark matter with a generic initial angle minimises the cost at a sharp edge, and the band within 2 bits above the minimum is m_a ≈ 2-20 µeV (core within 1 bit: 6-20 µeV): precisely the ADMX/MADMAX/DMRadio territory. Noted in advance: the post-inflationary scenario is the other branch and gives ~25-500 µeV; a detection there counts as a branch switch, not a miss. Miss condition: confirmed axion DM far outside 1-500 µeV, or dark matter found elsewhere while the bands turn up empty. [derived within the frozen conventions]

7. The price list is tool-independent. The same axiom measurement on a Rocq/Coq corpus must land within a few points of the Lean numbers (choice dependence ~70%, constructive core ~7%); a large deviation would mean the prices are properties of the tool rather than of mathematics. Outcome (measured the same night): miss. The measurement ran on Coq 8.20.1 with full Print Assumptions closure per lemma (method and script archived): the Coq standard library (16,664 lemmas, all non-compat modules) uses a logical axiom in 11.5% of its lemmas and is 88.5% constructive; Mathematical Components 2.3.0 (11,442 lemmas across ssreflect, finite groups and algebra) uses exactly 0% axioms. Against mathlib's 70.2% that is not "a few points": the failure condition is triggered and the miss is booked in the track record. The conclusion the stake itself prescribed: the measured prices are to a large degree properties of tool and proof style; mathlib's 70.2% is an upper bound on the price of choice within that proof style, not a property of mathematics itself. What does recur in both ecosystems, honestly labelled as a post-hoc hypothesis and not as a rescued stake: axiom use concentrates where the continuum enters. In Coq nearly all axiom load sits with the real numbers (LPO-like decidability axioms 11.0% plus functional extensionality 10.5%, strongly correlated; genuine choice: 0.03%), while the discrete mathematics is axiom-free: the same countable-cheap, continuum-expensive pattern as the law-address boundary. [miss, booked; hypothesis labelled separately]

Three retrodictions, booked as such (not predictions): (i) fifteen years of LHC desert is exactly what minimality retrodicts; (ii) the near-criticality of the Higgs vacuum (λ near zero at the Planck scale) is the data point the delegation reading expects; (iii) the existing f_NL and isotropy limits are already passed tests of the measure-preference reading. Retrodictions do not count toward the track record; they stand here because they carry the readings the stakes above rely on. [hindsight, honestly labelled]

Expectation, frozen before measuring. The miss of the tool-independence stake left a post-hoc hypothesis: axiom use concentrates where the continuum enters. The same day it was turned into a frozen stake (expectation document written in advance): measure the analysis extension of precisely the corpus that scored 0.0% (mathcomp-analysis + classical + reals: same community, same proof style), and logical axiom use jumps to a large majority: more than 50% counts as a hit, with the directional expectation above 90%, and a mix dominated by extensionality and choice rather than the stdlib's LPO family.

Outcome (full count, 5,887 lemmas). 77.2% uses a logical axiom: functional extensionality 77.1%, propositional extensionality 77.1%, choice-like indefinite description 69.2%. A hit on the frozen threshold. The directional expectation of > 90% was not reached (23% of the analysis corpus stays axiom-free: the discrete support lemmas), and that is stated honestly. The mix prediction was exactly right: extensionality plus choice, no LPO.

What this establishes. Within one ecosystem, same authors, same axiom-averse style, axiom use jumps from 0.0% (11,442 discrete lemmas) to 77.2% (5,887 analysis lemmas) the moment the real numbers enter. And choice goes from 0.03% (stdlib) and 0% (discrete mathcomp) to 69.2%. Countable is cheap, the continuum is expensive: no longer a post-hoc observation but a frozen, hit stake, visible in two independent proof-assistant traditions. The law-address boundary from physics now has its measured mirror image in mathematics. [hit, booked]

The question, and the frozen conventions. The quark masses and mixings are continuous parameters, so the meta-stake forbids deriving their values. What can be done is the θ̄ question: what does the observed hierarchy cost as bare address under the anarchy measure, against the structure cost of a mechanism? Conventions frozen in advance in the script header (yukawa_prijs.py, seed 20260813): the same Gaussian ensemble as the neutrino computation, windows of ±a quarter decade around the seven observed log ratios (four mass ratios, three CKM angles), independent-marginals estimate with power-law tail fits where the Monte Carlo runs dry.

Outcome. Total address price: ~95 bits (largest items: m_u/m_t 27 bits, m_c/m_t 16, θ₁₃ 14). The frozen Froggatt-Nielsen band (mechanism 10-20 bits, nine charges 25 bits, ε address 4 bits) is 39-49 bits: 46 to 56 bits in favour of a texture mechanism: over three times the margin that justified the axion stake at θ̄. Internal cross-check: the fitted tail exponents (2.00-2.08) are exactly the Ginibre value 2 from random matrix theory. Honest caveat: the marginals sum ignores correlations (under anarchy small angles and small mass ratios are linked), so 95 bits is an upper estimate within the convention; but even halved the margin stays wide.

Reading, and the diptych. The framework thus bets on a flavour mechanism in the quark sector (texture, symmetry or localisation), not on a bare address. And that completes a diptych now stated in one currency: the lepton sector is anarchic (the neutrino stakes run on it, P(NO) = 95.7% from the same ensemble), the quark sector demonstrably is not. K3-conditionally, in the style of the LISA fingerprint: at a symmetric address point the symmetry forces texture zeros via the trivial-representation selection rule, and the Fermat machinery for it stands ready; but the generation-bundle assignment is unspecified, so no concrete texture is claimed. Long-term referee: any confirmed flavour mechanism is a win for this reading; a proven mechanism-free hierarchy would leave 46-56 bits unexplained. [accounting complete; reading motivated]

Addendum (same day): the cheapest texture. The full finite search space of Froggatt-Nielsen charges has been traversed (textuur_zoektocht.py, criteria frozen in the header): 44 textures fit within one power of ε on all seven observables, and the CKM angles pin the quark-doublet charges uniquely at qQ = (3, 2, 0): across all passing candidates. The winner under the frozen ranking is qQ = (3,2,0), qu = (4,1,0), qd = (2,0,0): essentially the classic textbook texture, found here rather than chosen. Parameter-free testable order relations of the winner: V_ub ~ V_us·V_cb (holds within a factor 2.5, an O(1) coefficient) and a V_us ~ (m_d/m_s)-power relation; honestly noted: at the ε-power level the classic Gatto exponent ½ cannot be distinguished from the winner's ⅓. Cost band of the winner: 36-46 bits against ~95 bare: the margin stands. [search complete; identification motivated]

The question, with a verdict criterion frozen in advance. Is the habitability window of α a hall (then the value is address) or a needle (then stability forces the point, the honeycomb reading)? Fixed beforehand in the script header (venster_verscherping.py): the honeycomb reading wins if the window width drops below a factor 1.5 and 137 sits within 10% of an edge. Only α varies; everything else fixed.

The edges, computed ourselves. The lower edge comes from hydrogen stability: hydrogen decays once the electromagnetic part of the neutron-proton mass difference (linear in α, lattice band ±30% deliberately wide) overtakes the QCD part: α⁻¹ ≳ 77 central (band 65-86). That sharpens the literature value 59. The carbon-Coulomb candidate drops out (α⁻¹ ≳ 13, far weaker). The upper edge (galactic cooling, ∝ α³) is margin-sensitive: 640-1370, with the literature value 685 mid-band.

Outcome. Sharpened window: α⁻¹ ∈ [77, 685], width a factor 8.9 (was 11.6); adding the independent fixed-point edge gives [89.5, 685], factor 7.7. The log position of 137 stays comfortably interior (0.26; with the fixed-point edge 0.21). The needle criterion is missed by a wide margin: the window remains a hall. The honeycomb reading holds for the walls (those are structure), not for the point: the address reading stands, and the bits effect of the sharpening is honestly small (0.16 bits).

The surprise: two independent walls in the same place. The anthropic lower edge from hydrogen stability (77, band 65-86) and the fixed-point lower edge from the delegation stake (75-90) come from entirely different physics: lattice QED of the nucleon versus asymptotic safety of gravity: and land on top of each other. That proves nothing by itself, but it is exactly the kind of convergence the framework must record on the day it becomes visible. [computed within frozen conventions; edges soft where indicated]

Addendum: the map of the hall (same evening, zaalkaart.py). Inside the hall lies one certified sharp feature: the carbon island of the Hoyle resonance, with the published tolerance of ±4% in α: α⁻¹ ∈ [131.8, 142.7]. Three outcomes. One: α⁻¹ = 136 and 138 both sit well inside that island: the claim "136 burns the universe, 138 freezes it" is thereby refuted by the published sensitivities themselves. Two: the island address within the hall costs 4.8 bits: independently consistent with the ~5 bits of carbon conditioning already in the address ledger: the accounting checks itself. Three: within the island, 137.036 sits at log position 0.49: exactly mid-island, no boundary: whoever wants a "sweet spot on the edge" finds the opposite in the data: even at the finest known scale our position is generic. Honest limit of the map: whether other islands exist elsewhere in the hall (different resonance alignments at strongly different α) requires ab-initio nuclear physics at varied constants and remains an open computational question, with no claim; a first cluster-model answer is in the card below. [computed within frozen conventions]

What this is, and is not. The open question ("are there other habitable regions elsewhere in the hall?") has been tackled at the most honest level possible without a supercomputer: a transparent alpha-cluster model (zaalkaart_cluster.py, conventions frozen) with a validation gate fixed in advance: the model first had to reproduce the published ±4% Hoyle sensitivity with a physically plausible cluster distance, or stop. The gate passed (d = 6.9 fm, mid-range of the literature values for the dilute Hoyle state). This is explicitly not ab-initio nuclear physics; the real lattice-EFT computation remains the referee that can overturn this map.

The map, and the surprise. Next to the known Hoyle island [131.8, 142.7] a continent opens up: once α drops by about 6% (α⁻¹ ≳ 146), Be-8 becomes bound: the mass-8 bottleneck disappears and carbon forms without the Hoyle resonance. In the cluster model, 74% of the hall (log measure) is thereby habitable: [145.9, 685] as continent, plus the island. Uninhabitable remain [77, 131.8] and a narrow strip (142.7-145.9). The fine-tuning picture flips: the hall is mostly alive, not mostly dead: in line with the revisionist literature (Adams).

What it does to the accounting: the same number three times. We live on the island, which spans 4.9% of the habitable measure: 4.4 bits to be an islander rather than a continental. The ~5-bit figure has now appeared three times independently: as carbon conditioning (booked), as island address within the hall (4.8), and as island address within the habitable set (4.4). The interpretation shifts honestly: it is no longer a habitability price but an address choice within habitable territory, and the mild question "why the island rather than the continent" stands open at 4.4 bits: well within the normal address budget.

Frozen caveats. Continental habitability assumes stars with bound Be-8 burn livably (helium ignites earlier; qualitatively plausible, not computed), the O/C balance is not modelled, nor is state restructuring at large α shifts. Each of these can shrink the continent; the island and its triply consistent bits stand regardless. [cluster model, validation gate passed; ab-initio remains the referee]

Setup. The law-address triage already applied to α, θ̄ and the Yukawa sector has been run over the rest of the famous "unexplained" quantities (derivatie_audit.py, priors and windows frozen in the header). Each row receives one of five verdicts: law (derived, with receipts), address (price in bits), mechanism-favoured (structure cheaper than address), undecided, or grave.

New result one: the hierarchy "problem" costs 5 bits. The electroweak scale as an address under the ledger's log prior, with the ABDS habitability window (nuclei exist only for v within roughly 0.4-1.6 of the observed value): 5.0 bits. The notorious "one in 10³⁴ fine-tuning drama" is an artifact of a linear prior; in this register's currency the hierarchy is a cheap address. This also retroactively explains why the HL-LHC desert stake is comfortable: a 5-bit address needs no protection mechanism (SUSY, compositeness).

New result two: Koide, honestly deflated. The famous lepton relation Q = 2/3 (measured 0.6666605, off by 6×10⁻⁶) costs 14.0 bits as a coincidence under a uniform prior on the mathematically allowed range: while the mechanism band (discrete flavour symmetry) is 10-20 bits. The bands overlap: undecided, and the relation's pole-mass sensitivity counts against the structure reading. A riddle more famous than its evidential force. The Gatto receipt, by contrast, is striking: √(m_d/m_s) = 0.2247 against V_us = 0.2243 (0.2%): exactly what the found textbook texture implies: structure side.

New result three: "why now" is 1.7 bits. The coincidence Ω_Λ ≈ Ω_m today, weighed over the habitable epoch: less than two bits: not a riddle, a shrug.

The full triage table. LAW: g-2 (QED, twelve digits), the proton mass (lattice QCD), the hypercharges (anomaly conditions), the Born rule. ADDRESS: α (~5 bits island address), generations (10), Λ (3.5-9), the hierarchy (5), why-now (2). MECHANISM-FAVOURED: θ̄ (axion, ~14 bit margin), the quark hierarchy (texture, 46-56). UNDECIDED: Koide (14 against 10-20). GRAVE: every 137 formula from Eddington to today. The total address ledger of the observable grows to ~25-32 bits: still less information than a single line of text. [priors frozen; every row reproducible]

The book. All 23 known dimensionless address coordinates of this member now sit in one atlas (adres_atlas.py, conventions frozen in the header): nine Yukawas, three CKM elements, α, α_s, sin²θ_W, m_H/v, four neutrino quantities and three cosmological ones. Total size at current experimental precision: 266 bits. The complete address book of physics is thereby shorter than the ASCII text of the law itself (~600 bits), and the why-residue under window conventions remains the postcard's familiar ~25-32 bits.

The comparison, with numerology discipline. A systematic sweep over 1.38 million product relations among the addresses (frozen grammar: 2-3 quantities, exponents ±1/3 through ±2, constants 1-3), with two gates frozen in advance. Calibration gate: the Gatto relation must be found, and it is. Null gate: the same sweep over twenty synthetic address books with random values. Result: the real top hits (log deviations down to 10⁻⁵, impressive-looking formulas such as y_μ^1/2·V_cb⁻²·r_ν^1/2 = 2) are exactly as tight as what the random books produce (real best 9×10⁻⁶ against null median 5×10⁻⁶). Verdict A, frozen in advance: the real address book behaves like a random book.

What this means. The address book is incompressible under this grammar: comparing finds no law, and every individual tight hit is expected coincidence (the MDL accounting: grammar cost ~20 bits against at most ~21 bits saved for even the tightest hit). This is also the quantitative death certificate of the entire 137-numerology industry: search long enough and you find five-decimal formulas, and the null calibration shows that means nothing. The two genuine relations escape only via a mechanism pedigree: Gatto pays through the found textbook texture, Koide stands undecided as a sum relation in the derivation audit above. [grammar, gates and seed frozen; 2 seconds of compute, fully reproducible]

Setup, frozen in the script header (anarchie_joint_mc.py, seed 20260813). Two million draws from the seesaw anarchy ensemble (basis-invariant Gaussian Y and M_R, shape m ∝ Y M_R⁻¹ Yᵀ), conditioned on only the measured mass-squared ratio; angles and phase stay free and are reported as they fall.

Outcomes. P(normal ordering) = 95.7%: inside the previously frozen 95-99.9% band. Mass sum: median 60.6 meV, 68% [59.2, 64.3] meV: the "peak near 0.06 eV" comes out sharply, and P(sum < 120 meV) = 100%. |sin δ_CP|: P(> 0.5) = 66.9%, median 0.71: the ensemble reproduces the analytic Haar 66.7% to 0.2 points: independent cross-validation of the δ_CP sharpening. Octant: 49.9 versus 50.1: no preference, as registered.

And one honest revision, booked before the data. m_ββ: median 15.3 meV, 68% [6.1, 30.3] meV, 95% [2.1, 42.4] meV; P(1-10 meV) = 32%. The earlier "1-10 meV" band (adopted from the literature) put the weight too low: our own joint computation shifts most of it to 6-30 meV. That makes a detection at LEGEND-1000/nEXO more likely than the earlier text suggested, and the revised interval is the operative stake from now on. Both versions remain visible: this is a register change before the measurement, not an adjustment after it. [derived within the frozen ensemble]

Top mass toward the criticality value. The delegation reading says λ(M_Pl) ≈ 0 is no accident. Translated into a frozen stake: as the top-quark pole mass is measured more sharply (HL-LHC toward ~0.2-0.3 GeV; later e⁺e⁻ threshold scans), it converges to the criticality window ~171.0-171.5 GeV. Current world averages (~172.5 ± 0.3) sit well above it; that tension is known and not hidden: which is exactly why this is a real stake. Failure condition: m_t pinned at high precision at or above ~172 GeV; then the near-criticality data point of the delegation reading falls. Own computation (13 August, vastpunt_dossier.py, 1-loop, conventions frozen in the header): the shape is confirmed: λ(M_Pl) drops by ~0.006 per GeV of top mass, so a ±0.2 GeV measurement decides the stake sharply. But the criticality window itself is loop-sensitive: the own 1-loop run gives ~167-168 GeV where the 2/3-loop literature gives 171.0-171.3. The stake therefore stays formulated on the literature band, with the own run as shape and decidability check, not as replacement. [motivated, bold]

Four small stakes in the minimality class. (i) No echoes in black-hole ringdowns: exotic compact objects are structure without a gate (LVK O5). (ii) No fifth force: MICROSCOPE-class equivalence tests stay null. (iii) The cosmic-birefringence hint fades; note: our own axion stake (2-20 µeV) produces no birefringence, so double bookkeeping is impossible here. (iv) c_gw = c stays exact in every multimessenger event. [motivated]

Two retrodictions, booked as such: the muon g-2 episode ended without new physics (the lattice computation closed the gap: minimality), and GW170817 measured c_gw = c to 10⁻¹⁵. Outside the track record. [hindsight]

What stays closed and what opened. The absolute frequency remains unclaimed: it hangs on the stabilisation scenario, and any number would repeat the retracted 3.2 mHz mistake. But the ratio question turned out to be posed wrongly: at a generic moduli point nothing forces a ratio (the point is address), while the cheap-address weighting e⁻ᴷ favours symmetric points: a symmetric point is cheaper to specify, exactly the logic of the 10-bit generation ledger. And at a symmetric point the symmetry does force structure: potential-independently.

Protocol and computation (frozen in the script header, lisa_vingerafdruk.py, seed 20260813). Choice of point: the Fermat point, the maximally symmetric algebraic K3 already used by the a₆ computations (an existing frozen choice, not newly picked). Symmetry group: the 1536 linear automorphisms (Z₄)³⋊S₄, acting on the 22 axion and 20 Kähler directions via the exact Fermat eigenspace model. Measurement: generic invariant mass matrices via group averaging, with two independent random matrices as a genericity cross-check (passed).

Outcome. Axion sector (22): five lines with degeneracies [12, 6, 2, 1, 1]. Kähler sector (20): four lines with [12, 6, 1, 1]. Total: at most 9 distinct lines instead of 42, dominated by two twelvefold and two sixfold multiplets. That is the fingerprint: if the multi-peak pattern is ever observed, it comes with this multiplet structure (or that of another symmetric point, with different but equally computable multiplets).

Failure conditions and honest edges. An observed spectrum of dozens of separate lines without multiplets would make the cheap-address weighting lose. Conditional on: the pattern being seen at all (the scale stays open), the address point being symmetric (the core of the stake), the potential being invariant, and the linear automorphism group being the relevant symmetry. The sign convention in the eigenspace model has been checked to be irrelevant for the degeneracy structure. [derived within frozen choices]

The decision tree, frozen before nature speaks (13 August). Branch A, pattern with multiplets of 12 and 6: the first direct observation of the a-local weighing itself; the address becomes measurable (the multiplet sizes identify the symmetry group), the frequency scale becomes computable after the fact, the K3 layer confirmed. Branch B, multiplets of other sizes: the cheap-address weighting wins, only the Fermat sub-choice misses; the recipe is ready: compute the automorphism group back from the observed sizes. Branch C, dozens of separate lines: the cheap-address weighting loses, and the fallback is fixed now already: the address measure is then flat (K weighs structures only, not points), with mandatory recomputations: the generation bits become reinterpretation instead of price, and the axion window is recomputed under the flat measure (wider). What does not fall in branch C: the law-address split, P = e⁻ˢ, the Born derivation and the selection chain, which is carried by gates and measurements, not by address weighting. Branch D, nothing seen: undecided, condition unmet: honestly the most likely outcome, and the stake stands for every successor of LISA. There is no branch without information.

• L1 — Sector restriction. Existence of minimisers is only hard within the Ricci-flat K3 sector; on bare superspace boundedness demonstrably fails (conformal-factor problem, Caveat 5.1 mathematics).
• L2 — Factorisation. That the weight factorises as e^−S_dyn/ℏ · 2^−K (quantum-weighted dynamics × prior-weighted description) is a structural assumption, not a theorem. Sharpened (Aug 2026): across the finite geometries S_dyn is demonstrably selection-flat, so the assumption carries no load there; its remaining work sits in the continuous sector only.
• L3 — MDL language choice. K depends on the description language. Qualitative orderings (K3 ≪ CY3) are robust; quantitative values are not.
• L4 — Reading rule. Typicality within self-representing universes (Reading rule 10.8) is an explicit assumption. Without it the measure yields no observer predictions. Sharpened (Aug 2026): the observer gates have been reduced to one abstract gate (stable rewritable self-representation) which is performatively self-validating; what remains is the typicality reading itself plus one named loophole (see the gates card in section 2).
• L5 — Numerical access. No formula exists for the Ricci-flat K3 metric; all spectral predictions run through numerical geometry with corresponding error bars.
• L6 — Open mathematics. T_t/T_h orthogonality within the slice, the boundary λ → 0, and the zero-point dynamics (11.2) are unsolved.
• L7 — Underdetermination. Other meta-laws could yield partly the same predictions. Convergence of independent arguments reduces this risk but does not eliminate it.
• L8 — Scope. No claims about morality, meaning, consciousness or subjective experience. EEML is a natural-law hypothesis, not a worldview.

Philosophical foundation (HERE, three-layer, both-and):      8/10
Mathematical core (P = e⁻ˢ, Kraft, dichotomy):                9/10
Decomposition & reductions (α_K = ln 2, two Hessians):        7/10
Physical selection (K3 via Rokhlin, N = 3):                   7/10
Constant derivations (F2′ triggered Aug 2026: retracted):     1/10
Structure selection (SM + ν_R, within explored universe):     7/10
Reading rule / typicality (assumption, conjecture open):      5/10
Overall (consistent with physics page):                       ~7/10

The lowest scores mark the research agenda, not the weakness of the whole: precisely there (the α number, the convergence conjecture) lie the testable gains.

The framework has publicly retracted claims before: the α point formulas (v0.9, v0.10), the LISA frequency 3.2 mHz (dimensionally inconsistent), Ω_DM = 0.264 (inconsistent scalings), Σm_ν ≈ 0.06 eV (not derivable), the strong uniqueness claim "the only possible universe" (v0.13, replaced by the three-layer), and now also the full α derivation itself (F2′ triggered, v0.19, after three pre-registered falsifications by the framework's own computations). Every retraction is documented in the changelog. That is not a parade of weakness: a framework that never retracts anything is not testing itself. Addition (12 August, night): the tool-independence stake on the axiom price list was measured the same night and missed (Coq 11.5% and mathcomp 0% against mathlib's 70.2%); this is the second booked miss alongside the Zipf demo. The repaired continuum stake was frozen, measured and hit the same day (77.2% against a 50% threshold; see the card in section 2).