Category-Native Solomonoff Approximation: From Algorithmic Geometry to Kernels, Operators, and Induction † Authors: Boumediene Hamzi, Marcus Hutter Published: 2026 ABSTRACT Solomonoff induction mixes all computable explanations with description-length weights, but it is incomputable. This theory-and-position paper argues that practical approximation must be category-native: one should first declare the mathematical category in which a computable shadow will live, then use that category’s native comparison functional, complexity code, and inductive object. The proposal is not an omnibus theorem asserting that all categories are equivalent. It is a research architecture that separates comparison, representation, and prediction and makes the information lost by each projection explicit. The metric–measure branch supplies the developed realization. Compression data define an empirical Solomonoff space; Gromov–Wasserstein (GW) distance supplies relational distortion; minimum description length (MDL) controls candidate complexity; and distance-to-kernel embedding produces a positive-semidefinite predictor. For finite or countable coded classes, we prove existence and stability results, a held-out validation oracle inequality, a Kolmogorov–Solomonoff kernel unification, conditional empirical-GW consistency, and a coding-redundancy bound. Stronger learning-oracle statements remain conditional on marked/predictive selection, candidate-family adequacy, and kernel stability. Topological, Banach/Barron, graph, tree, and operator branches are presented as a constructional and testable research programme, with their maturity stated explicitly. 1. Introduction Solomonoff induction is a normative ideal for sequential prediction: it averages all computable explanations with weights that decrease exponentially in description length. Its universality and incomputability are inseparable; therefore, any practical approximation chooses what structure to retain, i.e., a distance, measure, topology, norm, graph, spectrum, operator, tree, or predictive law. The choice is made before choosing an estimator, and determines which invariances are respected and which information is discarded. This paper advances a position about that choice: computable shadows of Solomonoff induction should be category-native, that is, the target category should supply its comparison functional, complexity code, and downstream inductive object. The phrase is methodological, not a claim that one theorem already covers every mathematical category. Specifically, it says that Gromov–Wasserstein (GW) distortion is appropriate for relational metric–measure data, persistence distances for topological summaries, norm or variation discrepancies for Banach/Barron models, graph or diffusion distances for graph structure, and operator distances for spectral models. Forcing all of these objects through a single Hilbertian or metric functional would erase the very structure that the approximation is meant to preserve. The metric–measure (mm) branch is this paper’s theorem-bearing realization. Observations are encoded as an empirical mm-space in which distances are compressor-based surrogates for algorithmic distance and masses are empirical or Occam-weighted. Candidate mm-spaces are compared to this object by GW distance and selected byθ^∈arg minθdGW,pp(S^n,Mθ)+λCode(θ).For supervised tasks, the objective must also contain a marked or held-out predictive term. This distinction is substantive: unmarked GW is invariant to relabeling, and can preserve relational geometry while destroying prediction. The resulting design rule is that the native comparison must include the marks or predictive loss relevant to the task, and that the candidate family must retain the structure on which those marks depend. Empirical scope. The numerical section evaluates the finite surrogate through mechanism checks rather than a benchmark claim. Controlled string corpora test recovery of planted low-complexity block structure, rejection of structureless noise, the need to penalize fitted flexibility against memorizing candidates, and held-out relational selection. Separate finite studies probe spectral calibration under exact and approximate orthogonality, GW-to-kernel stability, the pairing dependence of metric repair, cross-compressor stability, concentration under Occam-reweighted mass, and the marked-coupling mechanism. Downstream prediction is then stress-tested in two harder settings. On a cellular-automaton classification task, geometry-only selection favors a one-block quotient and produces chance-level balanced accuracy; a nested stratified selector that is allowed to retain the raw compression geometry avoids this collapse but does not improve on the raw compression-kernel baseline. On a balanced subset of the UCI SMS Spam Collection, fused marked selection does not rescue average-linkage block quotients whose partitions are label-mixed, whereas the uncoarsened compression geometry retains strong predictive signal. Even a post-hoc oracle over the fitted block kernels remains at chance. Together these studies distinguish two requirements: task alignment can protect against an unsuitable geometry-only objective, but neither marks nor validation can recover predictive information already discarded by an inadequate candidate family. This paper makes five contributions at two deliberately separated levels:It formulates category-native Solomonoff approximation as a common problem architecture: choose a target category, a native distortion, a prefix-decodable model code, and a native inductive object. This is the paper’s positional and organizational contribution, not a cross-category equivalence theorem.It supplies a maturity map for topological, Hilbert/Banach, graph, tree, operator, predictive, and metric–measure branches. The map makes the proposed objects and the missing stability or prediction results explicit, so the vision is testable rather than merely rhetorical.It develops the metric–measure branch: the empirical Solomonoff mm-space, the GW + MDL selector, and a marked extension that aligns geometry and task loss through the same coupling.It proves the finite/countable technical core: existence and perturbation results, a held-out validation oracle inequality for coded candidates, a finite transfer statement, downstream kernel/operator recovery, conditional empirical-GW consistency, and a coding redundancy bound for mixtures over coded mm-models.It supplies numerical stress tests of structural recovery, negative controls, spectral and kernel stability, compressor sensitivity, marked coupling, and downstream prediction. The cellular-automaton and real SMS studies show that unmarked selection can collapse and that task marks cannot repair a quotient family that has already discarded predictive structure. A validation safeguard that can retain the raw geometry prevents collapse. These results correct an unconditional “geometry first” reading without abandoning the wider claim that different representational categories require different, task-aligned notions of fit. It formulates category-native Solomonoff approximation as a common problem architecture: choose a target category, a native distortion, a prefix-decodable model code, and a native inductive object. This is the paper’s positional and organizational contribution, not a cross-category equivalence theorem. It formulates category-native Solomonoff approximation as a common problem architecture: choose a target category, a native distortion, a prefix-decodable model code, and a native inductive object. This is the paper’s positional and organizational contribution, not a cross-category equivalence theorem. It supplies a maturity map for topological, Hilbert/Banach, graph, tree, operator, predictive, and metric–measure branches. The map makes the proposed objects and the missing stability or prediction results explicit, so the vision is testable rather than merely rhetorical. It supplies a maturity map for topological, Hilbert/Banach, graph, tree, operator, predictive, and metric–measure branches. The map makes the proposed objects and the missing stability or prediction results explicit, so the vision is testable rather than merely rhetorical. It develops the metric–measure branch: the empirical Solomonoff mm-space, the GW + MDL selector, and a marked extension that aligns geometry and task loss through the same coupling. It develops the metric–measure branch: the empirical Solomonoff mm-space, the GW + MDL selector, and a marked extension that aligns geometry and task loss through the same coupling. It proves the finite/countable technical core: existence and perturbation results, a held-out validation oracle inequality for coded candidates, a finite transfer statement, downstream kernel/operator recovery, conditional empirical-GW consistency, and a coding redundancy bound for mixtures over coded mm-models. It proves the finite/countable technical core: existence and perturbation results, a held-out validation oracle inequality for coded candidates, a finite transfer statement, downstream kernel/operator recovery, conditional empirical-GW consistency, and a coding redundancy bound for mixtures over coded mm-models. It supplies numerical stress tests of structural recovery, negative controls, spectral and kernel stability, compressor sensitivity, marked coupling, and downstream prediction. The cellular-automaton and real SMS studies show that unmarked selection can collapse and that task marks cannot repair a quotient family that has already discarded predictive structure. A validation safeguard that can retain the raw geometry prevents collapse. These results correct an unconditional “geometry first” reading without abandoning the wider claim that different representational categories require different, task-aligned notions of fit. It supplies numerical stress tests of structural recovery, negative controls, spectral and kernel stability, compressor sensitivity, marked coupling, and downstream prediction. The cellular-automaton and real SMS studies show that unmarked selection can collapse and that task marks cannot repair a quotient family that has already discarded predictive structure. A validation safeguard that can retain the raw geometry prevents collapse. These results correct an unconditional “geometry first” reading without abandoning the wider claim that different representational categories require different, task-aligned notions of fit. The logical status of the two levels is summarized in Table 1. A position can be substantive without necessarily being a theorem: it should organize previously separate constructions, state design commitments, expose failure conditions, and generate a programme of discriminating results. The metric–measure theorems then demonstrate that one branch can be carried from the position to a computable selector and conditional learning guarantees. In particular, the conceptual GW + MDL definition does not imply the later learning oracle inequality. That inequality additionally assumes a coded candidate class, independent validation or an equivalent generalization device, marked/predictive alignment, and stability of the induced kernel learner. What kind of paper this is. The claims have three statuses: (A) Position: Category-native approximation is an organizing proposal, defended by the separation of native comparison, representation, and induction and by the distinct questions it generates; it is not presented as a derived cross-category theorem. (B) Proved core: The metric–measure branch contains the finite/countable results stated with their hypotheses. (C) Research programme: The topological, Banach/Barron, graph, tree, and broad operator branches specify constructions and proof obligations that remain open. These levels are complementary, not interchangeable. How to read the paper. Readers interested primarily in the technical case may read Section 5, Section 7, Section 11, Section 15 and Section 17 and the predictive-collapse and real-data marked-selection analyses in Section 14. Readers evaluating the positional contribution should additionally follow the three category-by-category sweeps, covering native comparisons in Section 8, representation and kernelization in Section 11, and induction in Section 11.5. The maturity map in Section 2 states which claims are proved, constructional, or open. Parts I–VI of the Bridging Algorithmic Information Theory and Machine Learning series [1,2,3,4,5,6] studied compression-based kernels, spectra, and Gaussian/Hilbert constructions. The present paper supplies the metric–measure selection layer that precedes those constructions and shows precisely when the earlier kernels are recovered. A broader kernel-first synthesis of Parts I–XXI, including the category-native extensions, their comparison–representation–induction atlas, and a registry separating proved, conditional, empirical, and programme-level claims, is developed in the accompanying research monograph preprint [7]. The present paper remains responsible for the claims made here; the monograph supplies wider programme-level context rather than replacing the metric–measure arguments below. Neither work claims that a kernel, Gaussian process, or mm-space equals the universal semimeasure. Table 2 collects the acronyms used in the manuscript, and Table 3 fixes the notation that separates ideal, empirical, and selected objects. The manuscript keeps two threads visible: Section 2 states the category-native position and the commitments by which it should be judged, while the remainder develops metric–measure approximation, kernel recovery, error decomposition, and finite/countable guarantees as the most mature realization. This asymmetry is intentional; a position paper can propose the architecture of a field while also being explicit about which branch has reached theorem level. Table 4 records that maturity assessment branch by branch. 2. Category-Native Solomonoff Approximation: The Position The central position is that no category-independent scalar discrepancy can preserve every computable shadow of Solomonoff induction. A topology forgets metric scale; a metric without mass forgets posterior weight; a spectrum can forget eigenfunctions; a kernel commits to a Hilbertian representation; and an unmarked geometry forgets labels. These are not interchangeable defects. They are category-specific projection choices. The right question is not “which single distance approximates Solomonoff induction?” but “which invariant should be preserved for the representation and task at hand?” Let S denote a Solomonoff-side source: a prior prefix tree, a posterior space of explanations, or a computable empirical shadow. For a target category C, the organizing problem is(1)ΔC(S)=infA∈CDCS,FC(A)+λCode(A)+ηPredLossC(A).Here, FC exposes the part of a candidate that can be compared with the source, DC respects the invariances native to C, Code(A) prices the computable candidate, and the predictive term is included whenever structure alone is insufficient for the task. The commonality lies in the architecture of the objective; its first and last terms are deliberately category- and task-dependent. “Comparison functional” here is intentionally broader than “metric minimized at zero”. A multiplicative distortion with optimum (1) may be logged and a similarity maximized at (1) may be converted into a loss, but its orientation, normalization, and null equivalence must be declared. Persistence and spectral summaries may also be invariant without being faithful to the underlying object. Thus, (1) is a typed optimization schema, not a universal metric axiom; forcing every row to satisfy one metric definition would contradict the category-native premise rather than formalizing it. The accompanying research-monograph preprint develops this architecture across Parts I–XXI and gives a more extensive atlas of metric, metric–measure, Hilbert/RKHS, Banach/Barron, operator, tree/graph, topological, and predictive law targets [7]. The present paper supplies the detailed metric–measure realization within that wider programme. Three commitments make this more than a naming exercise. First, native invariance: the comparison must respect the equivalences of the target representation rather than importing an arbitrary Euclidean proxy. Second, compositional accountability: comparison, representation, and induction must be stated separately, so that a good structural match is not silently converted into a predictive guarantee. Third, empirical corrigibility: the cellular-automaton collapse and the real-data marked-selection results must be allowed to revise the objective. In supervised settings, permutation-invariant unmarked fit is inadequate; marks, validation loss, or another task-aware term must enter, and the candidate family must retain task-relevant structure. The programme would fail as a scientific position if category-specific choices could never change model rankings, stability statements, or predictions. It instead generates concrete questions: when does a persistence comparison retain information that GW discards? when does a Barron variation code outperform an RKHS code? when does a graph diffusion preserve task-relevant structure better than its shortest-path mm-space shadow? and when does a direct predictive mixture dominate every single structured projection? This paper answers the corresponding metric–measure question farthest. The other rows remain invitations to construct and falsify branch-specific theories, not decorative claims of completed unification. 3. Metric–Measure Realization This section identifies the Solomonoff-side source objects represented as metric–measure spaces. GW is not asked to discover the space where induction “really lives.” A source mm-space or finite shadow is fixed first; GW then measures the distortion incurred by representing it inside a restricted candidate class. 3.1. Variational, Not Ontological The operative distinction is:GWisvariational,notontological.For a source mm-space S and a restricted class G, define(2)ΔG(S)=infM∈GdGW,pp(S,M)+λCode(M).The code or declared MDL score prevents a vacuous preference for an increasingly flexible copy of the source. This restriction is essential. If G contains every mm-space and S itself is admissible at no complexity cost, the GW term can be zero. The research content lies in computability, code length, finite rank, ultrametric, graph-induced, manifold, spectral, or task-compatibility restrictions on G. 3.2. Three Solomonoff-Side Source Objects There are three source objects which should not be conflated. Prior-side output tree. Solomonoff induction is classically defined on finite and infinite binary strings. Let{0,1}≤∞={0,1}*∪{0,1}N and equip it with the prefix ultrametricdtree(ω,ω′)=0,ω=ω′,2−|lcp(ω,ω′)|,ω≠ω′.The monotone universal semimeasure is normalized at the empty string, with any remaining initial deficit assigned to that node. Assigning each prefix mass deficit to its finite-output node then gives a probability measure on this compact extension. This gives the prior-side mm-spaceStree={0,1}≤∞,dtree,M.This is canonical and useful for sequence-prediction questions, but captures output-prefix geometry rather than the posterior geometry of explanations after data have been observed. Posterior explanation geometry. For a dataset or history D, let HU be a computable hypothesis, program, grammar, dictionary, compressor state, or semimeasure class. In the measure–mixture presentation, writeπD(h)=2−ℓ(h)νh(D)∑g2−ℓ(g)νg(D).Assume the denominator is finite and positive. In the monotone program presentation, consistency replaces the likelihood. Choose a bounded measurable pseudometric dexp, quotient by its zero-distance relation ∼, push the posterior mass forward to that quotient, and take its metric completion (with the pushed-forward measure supported on the original quotient). Define(3)SD=HU/∼,dexp,πD.The predictive pseudometric specified below provides one such construction. Algorithmic or hybrid choices are admissible only after their metric and quotient properties have been established. This is the native object closest to Solomonoff induction as a data-conditioned learning procedure: its points are explanations, not observed data points. Empirical data-side shadow. The computable object used in Section 5 isS^n=(Xn,d^Sol,μ^n).This is a behavioral shadow of SD: observations inherit a finite relational geometry from compressor surrogates, finite program dictionaries, or other computable approximations to algorithmic distance. This is the object we can compare by GW in practice. Therefore, the intended conceptual order isStreeorSD⟶S^n⟶restrictedmm-space⟶inducedkernel/operator⟶predictor. 3.3. Triage of Approximation Problems The following table records which approximation cells are mathematical content and which are calibration or sanity checks. The paper’s main objective in Equation (5) should be read as a restricted approximation principle: the candidate family G is doing mathematical work; without that restriction, the problem is vacuous, while with it GW can measure how expensive it is to represent Solomonoff relational structure in the chosen geometry class. Table 5 triages the restricted problems by data regime and candidate class. 4. Materials and Methods: Mathematical and Computational Framework 4.1. Kolmogorov Complexity and Algorithmic Distances Fix a universal-prefix Turing machine U. We use standard notation from Kolmogorov complexity and universal induction [8,9,10,11,12,13,14,15,16,17]; see [18] for a treatment of universal induction in this journal, [19] for a philosophical critique, and [20] for general information-theoretic background. The prefix Kolmogorov complexity of a finite object x isK(x)=min{|p|:U(p)=x}.The conditional complexity isK(x∣y)=min{|p|:U(p,y)=x}.The algorithmic mutual information isIalg(x:y)=K(x)+K(y)−K(x,y).The normalized information distance is an ideal algorithmic dissimilarityNID(x,y)=max{K(x∣y),K(y∣x)}max{K(x),K(y)} whose metric properties hold only up to the usual precision terms. Exact metric–measure theorems below require a specified genuine metric (or a pseudometric quotient and completion); an unrepaired dissimilarity is used only in the finite network objective. For notation, we write dSol for an ideal algorithmic distance of this kind and d^Sol for its computable compressor surrogate; hatless occurrences always refer to the ideal object. Since K is incomputable, applications use computable surrogates such as normalized compression distance [21,22,23,24]NCDC(x,y)=C(x∘y)−min{C(x),C(y)}max{C(x),C(y)}, where C is a lossless compressor and x∘y denotes a self-delimiting concatenation of the two strings, following the compression distance viewpoint of [22,23]. For the purposes of this paper, d^Sol denotes any computable algorithmic dissimilarity intended to approximate an ideal Solomonoff or Kolmogorov distance. The framework does not require a single canonical empirical Solomonoff distance. Different compressors, program dictionaries, or KC-kernel surrogates give different empirical views of algorithmic structure. This is a feature, not a bug: cross-compressor stability becomes an empirical test of whether a discovered geometry is genuinely algorithmic rather than compressor-specific. 4.2. Metric–Measure Spaces The metric–measure viewpoint, its topology of isomorphism classes, and its convergence theory originated with Gromov and Sturm [25,26]; see also [27]. A metric–measure space, or mm-space, is a triple M=(X,dX,μX), where (X,dX) is a complete separable metric space and μX is a Borel probability measure on X. If X is finite, we identify μX with a probability vector. Isomorphism statements concern the supports of the measures; zero-mass points do not affect GW. Two mm-spaces are identified if there is a measure-preserving isometry between their supports. This is the correct equivalence relation for relational geometry; labels and coordinates are irrelevant. 4.3. Gromov–Wasserstein Distance Let M=(X,dX,μX) and N=(Y,dY,μY). A coupling between μX and μY is a probability measure π∈Π(μX,μY) on X×Y with marginals μX and μY. GW distances were introduced by Mémoli [28], and in the form used here, by Sturm [29]; computational aspects and entropic solvers are surveyed in [30,31,32,33,34], and structured-data variants include fused GW [35], graph matching and node embedding [36], network invariants [37], generative alignment across incomparable spaces [38], low-rank solvers [39], Gaussian restrictions [40,41], Procrustes metrics on covariance operators [42], generalized spectral clustering via GW learning [43], and GW-based structure learning of latent graphs and dictionaries [44,45,46]. For p≥1, the p-Gromov–Wasserstein distance is dGW,pp(M,N)=12infπ∈Π(μX,μY)∫dX(x,x′)−dY(y,y′)pdπ(x,y)dπ(x′,y′). For finite spaces X={xi}i=1n, Y={ya}a=1m, this becomesdGW,pp(M,N)=12minΠ∈U(α,β)∑i,i′,a,a′|Dii′X−Daa′Y|pΠiaΠi′a′, where DX,DY are distance matrices, αi=μX({xi}) and βa=μY({ya}) are the probability vectors of the two measures, and U(α,β) is the transportation polytope with those marginals. 4.4. D2KE and Geometry-Induced Kernels Given a metric–measure space M=(Z,d,μ), one can construct positive semidefinite kernels from d in several ways. If d2 is conditionally negative definite, thenkγ(z,z′)=exp(−γd(z,z′)2) is positive semidefinite by Schoenberg’s theorem [47,48]. More generally, D2KE (distance-to-kernel-and-embedding, due to Wu et al. [49]) constructs a positive semidefinite kernel by using distances to landmarks, in the spirit of the KC-kernel and Mercerization constructions of the earlier BAITML papers [1,2,3]; general RKHS background is standard [50,51,52]:(4)kD2KE,γM(z,z′)=∫Zexp(−γd(z,ω))exp(−γd(z′,ω))dμ(ω).This kernel is positive semidefinite because it is an inner product of feature mapsΦz(ω)=exp(−γd(z,ω))inL2(Z,μ). In the present paper, (4) is the default way of passing from learned geometry to RKHS/GP machinery. This step guarantees a positive-semidefinite representation; it does not guarantee that the induced kernel metric reproduces the input distance. For a finite weighted sample, a practical audit should report both a conditionally negative-definite defect obtained from the negative spectrum of the centered matrix −12JD∘2J and the mass-weighted distortion between D and the D2KE-induced kernel distance. A large defect means that Hilbertization can change neighborhoods or rankings even though the resulting kernel remains valid for prediction. That loss belongs to δker in the error budget, and is precisely why the category-native comparison is performed prior to D2KE. 5. The Empirical Solomonoff Metric–Measure Space 5.1. Definition Let Xn={x1,…,xn} be a dataset and let d^Sol:Xn×Xn→R+ be a symmetric computable algorithmic dissimilarity with zero diagonal. Let μ^n be either the uniform empirical measure or an Occam-reweighted empirical measure. The empirical Solomonoff mm-space is S^n=(Xn,d^Sol,μ^n). The distance d^Sol need not be a perfect metric. In practice, NCD-like quantities may have small violations of the triangle inequality or even mild asymmetries. There are three standard repairs:d^sym(x,y)=12(d^(x,y)+d^(y,x)),d^met(x,y)=shortest-pathmetricgeneratedbyd^sym, or direct use of a distortion functional for dissimilarity spaces rather than metric spaces. The clean theory below assumes a bounded metric, but the algorithmic formulation can be applied to repaired empirical distances as well. 5.2. Choice of Empirical Mass The simplest mass is uniform:μ^n=1n∑i=1nδxi.For a more Solomonoff-like empirical object, use an Occam-reweighted measure. Here, K^(xi) denotes a computable code length for xi, for example a compressed bit length plus any delimiter or model-header convention used by the compressor. In exact algebra, one may writeμ^nOccam(xi)∝2−K^(xi).For numerical work, this expression should be evaluated in log-space and with a declared effective code K^†,Kmin†:=min1≤j≤nK^†(xj),μ^n,τOccam(xi)=exp{−τ(K^†(xi)−Kmin†)}∑j=1nexp{−τ(K^†(xj)−Kmin†)},τ≥0.For objects of comparable lengths, one may take K^†=K^. When raw lengths vary materially, raw Solomonoff weighting intentionally favors physically shorter objects; if that is not the intended estimand, one must instead declare a conditional code K^†(x)=K^(x∣|x|), stratify by length, or use another explicitly justified nuisance-length adjustment. Dividing by |x| gives a compression rate, but does not automatically provide a Kraft-valid object code and should not be called Solomonoff mass without an additional coding argument. Subtracting Kmin† does not change the normalized weights but prevents floating-point underflow. The temperature τ controls the strength of the Occam bias (throughout, log denotes the natural logarithm and code lengths are in bits): τ=ln2 recovers the bit-length convention, smaller values give a softer Solomonoff tilt, and τ=0 gives the uniform empirical measure. To make clear when the Occam measure has collapsed toward a small number of atoms, in numerical experiments one should report K^†, τ, any clipping or flooring of weights, and the effective sample size (which is an entropic quantity; ESS(μ^)=exp(H2(μ^)), the exponential of the Rényi-2 entropy, so the collapse diagnostic below is a Rényi-entropy floor):ESS(μ^)=1∑iμ^i2.The uniform version treats the dataset as observed samples, while the Occam version treats shorter descriptions as more central in the empirical geometry. The correct choice depends on whether the goal is descriptive modeling of the observed distribution or approximation of a universal prior. 5.3. Marked Empirical Solomonoff Spaces For supervised learning, the empirical object should include labels. Let yi∈Y. A marked empirical Solomonoff space isS^nmark=(Xn,d^Sol,μ^n,ℓn), where ℓn(xi)=yi. The labels are not inserted by changing the input metric alone; rather, they enter through a marked GW objective. This avoids confusing algorithmic similarity of inputs with predictive relevance. 6. Behavioral Transfer from Explanation Geometry to Data Geometry The empirical space S^n is useful only if it preserves enough of the posterior explanation geometry SD. This section makes that bridge explicit. It is a load-bearing assumption for interpreting data-side GW as evidence about explanation-side GW. LetBD:HU/∼⟶Bn be the behavior map sending an explanation to its finite behavior on the observed sample, histories, labels, or query set. In the ideal case,μ^n≈(BD)#πD, and d^SolBDh,BDg approximates the explanation distance dexp(h,g) on the posterior mass that matters. The behavior map is (c,C,ε,δ)-faithful on posterior mass, with normalized constants 00. If (9) λΔC>B(D), then every minimizer of the unmarked distortion-plus-penalty objective over that menu has k=1. For every k≥2, non-negativity of the distortion and Lemma 2 giveJ(1)−J(k)≤B(D)−λ(Ck−C1)≤B(D)−λΔC<0.Thus, k=1 strictly dominates every other member of the menu. □ Four scope restrictions are essential. First, M1 is an n-point equilateral blow-up coded by one block parameter, not the honest one-point quotient. Against the one-point quotient, the distortion is121−1nVoff(D)+d¯2, which can be much larger. Second, Proposition 2 concerns the declared k-block menu. A tree, raw-geometry, or other candidate with a smaller penalty gap must be analyzed with its own ΔC. Third, the certificate is scale-dependent: D↦cD sends B(D) to c2B(D) while leaving the code unchanged. It is meaningful only after the distance normalization and λ calibration have been fixed. Fourth, the threshold λ★=B(D)/ΔC depends on both sample size and the empirical distance dispersion. For the implemented score, increasing n raises ΔCdecl(n)=2+log2n, while a representation change that shrinks Voff lowers B(D); either effect can make the sufficient condition easier to satisfy at fixed λ. This is not a universal monotonicity law, as the distribution of D may itself change. The controlled elementary cellular automaton (ECA) audit reported in Section 14 supports recalibrating λ whenever the sample size, object representation, or distance normalization changes. Most importantly, (9) diagnoses the objective and candidate family, not the information content of the data. Small Voff indicates that the one-block blow-up is cheap under the particular squared relational loss, not that local neighborhoods, marks, or other statistics of D lack predictive signal. The cellular-automaton and SMS prediction studies in Section 14 exhibit exactly this distinction. 7.4. Marked Objective For labeled data, let each candidate geometry carry a candidate-side mark map mθ:Zθ→Y, declared in advance or estimated using construction/training data before the marked coupling is solved. This map is distinct from the downstream predictor fitted after kernelization. Given a coupling π between μ^n and μθ, defineLmark(π,θ)=∫Xn×ZθL(y(x),mθ(z))dπ(x,z), where L is the pointwise prediction loss. This notation keeps the loss distinct from the label map ℓn used in the marked empirical space. The supervised or marked objective is (10) Jnmark(θ)=infπ∈Π(μ^n,μθ)[12∫|d^Sol(x,x′)−dθ(z,z′)|pdπ(x,z)dπ(x′,z′)+η∫L(y(x),mθ(z))dπ(x,z)]+λCode(θ). This objective matches both algorithmic geometry and task structure. When η=0, it reduces to the unsupervised objective. Structurally, it is the Solomonoff/MDL version of a Fused Gromov–Wasserstein objective: one term aligns pairwise relational geometry and the other aligns attached marks or labels through the same coupling. Existing FGW solvers can be used as implementation templates, with the code penalty and algorithmic distance supplying the Solomonoff-specific layer. If we want to learn mθ jointly with the coupling, then it is a different alternating or joint optimization problem and must be stated as such. The implemented labeled block experiments in Section 14 use training-derived block marks before solving marked GW. The later KRR predictor is then fitted separately on the selected kernel. 7.5. Interpretation The optimal coupling π★ is a soft correspondence between observed data and points in the explanatory geometry. If z is a latent program, then π★(x,z) says that program z explains data point x. If z is a node in a prefix tree, the coupling aligns data objects with branches of the algorithmic hierarchy. If z is a point on a manifold, then the coupling identifies the manifold coordinates whose internal distances best reproduce the empirical Solomonoff distances. Therefore, the GW term plays the role of an algorithmic structural likelihood that states how well the candidate model reproduces the algorithmic relational pattern in the data. 8. Category-Native Comparison Functionals: The Broad Tour This is the first systematic sweep of the category-native programme. The tour is repeated at two different levels by design: comparison functionals are specified here, kernelization or representation functors in Section 11, and induction rules later in that section. These are three columns of one architecture, not three interchangeable descriptions. Each paragraph below cross-references the corresponding representational and predictive constructions rather than treating them as consequences of the comparison alone. There is a hierarchy of ways to compare structured objects: isometry, bi-Lipschitz distortion, Gromov–Hausdorff (GH), Gromov–Wasserstein (GW), Wasserstein/MMD on a shared space, and operator-spectral comparison. They are often presented as competitors; in the present framework, however, they control different parts of the recovered object, so the principled construction may be layered rather than governed by a single scalar functional. Isometry is the ideal, not a separate computational tool. For mm-spaces, dGW=0 if and only if the spaces are measure-preserving isometric [28,29]. Isometry is the perfect score of the GW term. Isometry is the ideal, not a separate computational tool. For mm-spaces, dGW=0 if and only if the spaces are measure-preserving isometric [28,29]. Isometry is the perfect score of the GW term. Gromov–Hausdorff versus Gromov–Wasserstein. GH compares metric geometry without choosing a measure; GW compares metric probability spaces and is sensitive to mass. Neither criterion forces a particular weighting: uniform, empirical, and Occam masses are all possible in GW. A KC or Solomonoff kernel additionally requires the declared landmark weighting of Propositions 8 and 9. GH remains a measure-free geometric check, not a prescription for uniform mass. Gromov–Hausdorff versus Gromov–Wasserstein. GH compares metric geometry without choosing a measure; GW compares metric probability spaces and is sensitive to mass. Neither criterion forces a particular weighting: uniform, empirical, and Occam masses are all possible in GW. A KC or Solomonoff kernel additionally requires the declared landmark weighting of Propositions 8 and 9. GH remains a measure-free geometric check, not a prescription for uniform mass. Operator-spectral comparison, either dspec of Section 16 or a richer heat kernel/diffusion comparison, targets the eigenvalue profile, and hence the Occam-calibration layer. It complements relational GW. Operator-spectral comparison, either dspec of Section 16 or a richer heat kernel/diffusion comparison, targets the eigenvalue profile, and hence the Occam-calibration layer. It complements relational GW. Wasserstein and MMD require a shared ambient space, and as such do not directly compare the empirical Solomonoff space with an arbitrary candidate. After kernelization, however, candidates can be represented in feature spaces, and a basis-sensitive discrepancy or task loss becomes natural for the marking or predictive layer of (10). Wasserstein and MMD require a shared ambient space, and as such do not directly compare the empirical Solomonoff space with an arbitrary candidate. After kernelization, however, candidates can be represented in feature spaces, and a basis-sensitive discrepancy or task loss becomes natural for the marking or predictive layer of (10). Category-by-Category Comparison Functionals The practical rule is as follows:useGWwhenbothobjectsarerepresentedasmetric–measurespaces.Otherwise, the invariant that is native to the structure being preserved should be used, and should be attached to the common architecture (1). This does not demote GW to an arbitrary menu item; GW is the relational mm-space functional, and the topology, normed linear structure, spectra, and prediction all have different invariances. Figure 1 depicts the shared architecture. Topological targets. A bare topological space (X,τ) has no distances and no probability measure, so GW is not defined on it. Homeomorphism, homotopy equivalence, and shape equivalence are usually too rigid for approximation. Thus, a practical topological comparison passes through a signature:(X,τ)⟼topologicalsignature⟼distancebetweensignatures.For example, choose a filtration FX(r), compute the persistent homology PHk(FX) [53,54,55], and compare the persistence diagrams [56] byDTop(X,Y)=∑kwkdbottleDgmk(FX),Dgmk(FY); alternatively, use interleaving distance [57] when the filtration-level structure is the object of interest. Natural Solomonoff-side filtrations include algorithmic distance thresholds, prefix depth, description-length thresholds, posterior mass thresholds, and algorithmic mutual information thresholds. This tests preservation of connected components, holes, branches, persistent clusters, and multiscale topology. If a compatible metric and measure are imposed, the object re-enters the mm-space layer and GW applies to that representation. Metric targets without measure preservation. If only distances matter, use Gromov–Hausdorff or Lipschitz distortion rather than GW:dGH(X,Y),c(X,Y)=infϕLip(ϕ)Lip(ϕ−1).Thus, GH is metric-only calibration, whereas GW is metric-plus-mass calibration. In the BAITML setting, dGH is appropriate for KC-style relational geometry that does not commit to Occam mass, while dGW is appropriate when probability weights such as μ(x)∝2−K(x) are part of the model. Hilbert targets. For Hilbert approximations, a supplement to GW is Hilbert distortion. Given (X,d), seek ϕ:X→H satisfyingad(x,y)≤∥ϕ(x)−ϕ(y)∥H≤bd(x,y) and minimize c2(X,d)=infϕb/a. Equivalently, a metric is Hilbertian when d2 is conditionally negative definite, so one may measure the distance of d2 to the CND cone. This is the structural diagnostic behind the Hilbert gap function in Section 9. GW measures mass-weighted relational cost, while Hilbert distortion or CND defect measures what is lost by forcing the object into Hilbert form. Banach targets. For finite-dimensional normed spaces E,F, the classical linear comparison is the Banach–Mazur distance:dBM(E,F)=infT:E→F∥T∥∥T−1∥.For a metric object represented inside a Banach space B, the flexible analogue is the Lipschitz distortion:cB(X,d)=infϕ:X→BLip(ϕ)Lip(ϕ−1).This makes it meaningful to ask whether a Solomonoff geometry is better represented in L1, L2, a variation-norm/Barron space [58,59,60], or another Banach geometry. After a Banach candidate is represented as (Z,∥z−z′∥B,μZ), GW can score its relational mm-space fit; the Banach distortion states what is gained or lost by insisting on that normed-linear presentation. Spectral and operator targets. Once a geometry has been kernelized, operator distances become natural. For kernel or covariance operators T1,T2, one can use∥T1−T2∥HS,∥T1−T2∥tr,dspec2(T1,T2)=∑i|λi(T1)−λi(T2)|2.This is the appropriate calibration when the target is the complexity spectrum λi≈2−ℓi. GW checks relational structure before or at kernelization, while spectral distance checks whether the induced operator has the desired Occam/Solomonoff eigenvalue profile. Predictive targets. If the goal is to approximate Solomonoff as a predictor, then the native comparison is predictive divergence rather than GW. Solomonoff convergence theory is built around cumulative KL,∑t=1nEμDKLμ(·∣x0. The floor is necessary because couplings that split atomic mass encounter diagonal distances equal to zero. These choices define modified comparison problems and can change candidate rankings; the normalization and floor must therefore be declared before selection. For p=2, the finite GW objective factorizes [30] as ∑i,i′,a,a′(Dii′X−Daa′Y)2ΠiaΠi′a′=α⊤(DX∘DX)α+β⊤(DY∘DY)β︸κ(α,β),independentofΠ−2〈DX,ΠDYΠ⊤〉F. Thus minimizing finite dGW,22 maximizes the Frobenius alignment 〈DX,ΠDYΠ⊤〉F. The relational comparison and its optimization are two expressions of the same alignment problem. 9. Boundary Diagnostics for Metric–Measure Projections This section records boundary diagnostics for forcing algorithmic or predictive explanation geometry into restricted metric–measure subclasses. They establish the limits of the developed mm-space realization. While they do not prove the other branches of Table 4, they do illustrate the category-native premise that different projections discard different information. 9.1. Algorithmic, Predictive, and Hybrid Explanation Distances There is no single canonical distance on explanations. On syntactic descriptions, normalized algorithmic information gives the dissimilarityaalg(h,g)=max{K(h∣g),K(g∣h)}max{K(h),K(g)} for program-theoretic convertibility, restricting to descriptions with positive denominator. It is not automatically an exact metric and is not invariant under semantic equivalence. For a metric–measure construction, fix a common support Q and a bounded genuine metric dalg on it, for example a finite metric repair of this dissimilarity on a declared dictionary, followed by completion. Claims about this repaired metric do not assert an exact identification with normalized information distance. On the same support definedpred2(h,g)=∑t≥1βtEx0, the hybrid on Q isdα=αdalg+(1−α)dpred,α∈[0,1].Throughout this display, (βt)t≥1 is a fixed deterministic discount sequence with βt≥0 and ∑t≥1βt<∞; for example, one may take βt=(1−ρ)ρt−1 with 0<ρ<1. This summability condition makes dpred finite whenever the one-step Hellinger losses are uniformly bounded, as one may normalize ∑tβt=1. The predictive distance is Hilbertian through the square-root embedding of probability laws. The chosen exact metric dalg need not be Hilbertian. Its Schoenberg defect must be checked for that metric; an obstruction for unnormalized information distance is not a proof for every normalized or compressor-based surrogate. 9.2. The Hilbert Gap Function Define the Hilbert-restricted approximation gap(11)g(α)=infM∈CHilbdGW,p(Q,dα,πD),M,For every α, take the metric completion; at α=0, first take the zero-distance quotient, always pushing forward the same posterior probability. Here CHilb is the class of metric–measure spaces realized as subsets of Hilbert space with their Hilbert distance and a probability measure. Assume that the predictive metric dpred is represented exactly in CHilb; then, g(0)=0. Moreover, for all α,α′∈[0,1], |g(α)−g(α′)|≤2−1/p|α−α′|∥dalg−dpred∥∞. If the source has finite support and the infimum is restricted to Hilbert realizations on that fixed support with the same mass vector, then a positive posterior-mass obstruction to Hilbert embeddability, for example, a positive mass-weighted CND defect, implies g(1)>0 for this fixed-support/mass restriction. The equality g(0)=0 follows by choosing the exact Hilbert realization of the predictive-Hellinger geometry. The Lipschitz bound is the GW distance perturbation estimate used in Section 10, together with∥dα−dα′∥∞≤|α−α′|∥dalg−dpred∥∞.For the final statement discard zero-mass points and write mi>0 for the fixed masses. On an objective sublevel, Minkowski’s inequality under any coupling gives∥DY∥Lp(m⊗m)≤∥DX∥Lp(m⊗m)+21/pdGW,p(X,Y).Thus every entry of DY is bounded. The Hilbertian pseudometric matrices form a closed set: their squared matrices are conditionally negative definite. A minimizing sequence therefore has a convergent subsequence in this bounded finite-dimensional set. If the infimum were zero, the limiting pseudometric quotient would be GW-isomorphic to the source. That quotient is Hilbertian, contradicting the assumed non-Hilbertian source. This proves positivity; attainment is asserted in the pseudometric closure, not necessarily among strictly positive metrics on the original labeled support. □ The function g is a compact diagnostic for how much of the posterior geometry is genuinely algorithmic rather than merely predictive. It also prevents the reader from mistaking an exact Hilbert embedding of an ultrametric or Hellinger geometry for evidence that all Solomonoff-side geometry is Hilbertian. 9.3. Gaussian Atom Gap A Gaussian or SGP approximation retains covariance geometry but discards atomic and higher-order posterior structure. This mismatch can be measured before kernelization. Letα(μ)=∑x∈Atoms(μ)μ({x})2.This is the mass at zero in the distance distribution d#(μ⊗μ). Fix p≥1, a dimension r≥1, constants 00, the D2KE kernel k(z,z′)=∫Ze−γd(z,ω)e−γd(z′,ω)dμ(ω) is positive semidefinite. For finite coefficients c1,…,cm and points z1,…,zm,∑i,jcicjk(zi,zj)=∫Z∑icie−γd(zi,ω)2dμ(ω)≥0. □ Thus, the selected geometry induces an RKHS, kernel ridge regression estimators, GP priors, and kernel mean embeddings without requiring the original algorithmic distance to be positive definite. 11.2. Metric–Measure Payoff: Recovering the Earlier Kernels The geometry-first construction recovers the earlier kernel papers by making two independent choices:distancedsetstherelationalfeatures,measureμsetstheirweights.Thus, the recovery is not a new derivation of a different object, it is a factorization of the old constructions through the selected geometry:uniformmassonalgorithmiclandmarks⟹KC/D2KEkernel,Occammassonthesamelandmarks⟹SolomonoffFeatureMixture,near-orthogonalweightedfeatures⟹Solomonoffspectrallawλi≍2−ℓi. The induced kernel(12)k(d,μ)(z,z′)=∫Zϕω(z)ϕω(z′)dμ(ω),ϕω(z)=e−γd(z,ω) is a master template that is parametrized by a distance d and landmark measure μ. The two classical objects of the programme are exactly the two canonical choices of the pair (d,μ), and the geometry-first pipeline returns each as a special case. Take the selected geometry to be the empirical Solomonoff space itself, M^=S^n=(Xn,d^Sol,μ^n), with μ^n the uniform empirical measure. Then (12) is the uniform-landmark D2KE kernel of d^Sol. It recovers the KC construction when its features and normalization match those in Parts II–III. In the Solomonoff–Gromov objective, this is the degenerate selection Mθ=S^n (zero relational distortion, dGW=0) with uniform mass. Zero distortion does not force its selection when a positive complexity penalty is present. On a nonempty countable dictionary, let wω=2−ℓ(ω) with 00; set the weight-scaled features vω=wω1/2uω and the operator T=AA*, (Ac)=∑ωcωvω. The nonzero eigenvalues of T coincide with those of the Gram matrix S, Sωω′=〈vω,vω′〉L2(ρ). Assume:(H1) bounded features: 0 < b ≤ ∥uω∥ ≤B<∞ for all ω;(H2δ) δ-near-orthogonality: for every ω,∑ω′≠ω|〈vω,vω′〉|≤δ〈vω,vω〉.Then, with w(1)≥w(2)≥⋯ as the sorted weights and λ1≥λ2≥⋯ as the sorted nonzero eigenvalues:1. (Exact case). If {vω} are orthogonal, define aω=wω∥uω∥2 and let a[1]≥a[2]≥⋯ be these products in decreasing order. Then, λi=a[i]. In particular, when ∥uω∥ ≡ 1, the product ordering is the weight ordering: uniform weights wω≡1 give λi≡1 (the finite KC-HS spectrum is flat/degenerate), and Solomonoff weights wω=2−ℓ(ω) give λi=2−ℓi exactly.2. (Perturbation—finite global band). If Ω is finite, then under (H1) and (H2δ), every nonzero eigenvalue λ of T lies in (1−δ)b2wmin≤λ≤(1+δ)B2wmax,wmin=minωwω,wmax=maxωwω.For Solomonoff weights, this reads (1−δ)b22−ℓmax≤λ≤(1+δ)B22−ℓmin. For countably infinite Ω with ∑ωwω<∞, there is generally no positive wmin. The correct statement is local; every nonzero eigenvalue belongs to at least one interval [(1−δ)wω∥uω∥2,(1+δ)wω∥uω∥2],and the only possible nonzero spectral points are discrete eigenvalues of finite multiplicity accumulating, if at all, at zero.3. (Perturbation—finite per-index band). If Ω is finite and the weight dynamic range is bounded as wmax≤κwmin with κ≥1, then for every i, b2−δκB2w(i)≤λi≤B2+δκB2w(i).In particular, for normalized Solomonoff weights ∥uω∥ ≡ 1, wω=2−ℓ(ω), with δκ<1, (1−δκ)2−ℓi≤λi≤(1+δκ)2−ℓi,so λi≍2−ℓi per index. (H1) bounded features: 0 < b ≤ ∥uω∥ ≤B<∞ for all ω; bounded features: 0 < b ≤ ∥uω∥ ≤B<∞ for all ω; (H2δ) δ-near-orthogonality: for every ω,∑ω′≠ω|〈vω,vω′〉|≤δ〈vω,vω〉. δ-near-orthogonality: for every ω,∑ω′≠ω|〈vω,vω′〉|≤δ〈vω,vω〉. 1. (Exact case). If {vω} are orthogonal, define aω=wω∥uω∥2 and let a[1]≥a[2]≥⋯ be these products in decreasing order. Then, λi=a[i]. In particular, when ∥uω∥ ≡ 1, the product ordering is the weight ordering: uniform weights wω≡1 give λi≡1 (the finite KC-HS spectrum is flat/degenerate), and Solomonoff weights wω=2−ℓ(ω) give λi=2−ℓi exactly. (Exact case). If {vω} are orthogonal, define aω=wω∥uω∥2 and let a[1]≥a[2]≥⋯ be these products in decreasing order. Then, λi=a[i]. In particular, when ∥uω∥ ≡ 1, the product ordering is the weight ordering: uniform weights wω≡1 give λi≡1 (the finite KC-HS spectrum is flat/degenerate), and Solomonoff weights wω=2−ℓ(ω) give λi=2−ℓi exactly. 2. (Perturbation—finite global band). If Ω is finite, then under (H1) and (H2δ), every nonzero eigenvalue λ of T lies in (1−δ)b2wmin≤λ≤(1+δ)B2wmax,wmin=minωwω,wmax=maxωwω.For Solomonoff weights, this reads (1−δ)b22−ℓmax≤λ≤(1+δ)B22−ℓmin. For countably infinite Ω with ∑ωwω<∞, there is generally no positive wmin. The correct statement is local; every nonzero eigenvalue belongs to at least one interval [(1−δ)wω∥uω∥2,(1+δ)wω∥uω∥2],and the only possible nonzero spectral points are discrete eigenvalues of finite multiplicity accumulating, if at all, at zero. (Perturbation—finite global band). If Ω is finite, then under (H1) and (H2δ), every nonzero eigenvalue λ of T lies in (1−δ)b2wmin≤λ≤(1+δ)B2wmax,wmin=minωwω,wmax=maxωwω.For Solomonoff weights, this reads (1−δ)b22−ℓmax≤λ≤(1+δ)B22−ℓmin. For countably infinite Ω with ∑ωwω<∞, there is generally no positive wmin. The correct statement is local; every nonzero eigenvalue belongs to at least one interval [(1−δ)wω∥uω∥2,(1+δ)wω∥uω∥2],and the only possible nonzero spectral points are discrete eigenvalues of finite multiplicity accumulating, if at all, at zero. 3. (Perturbation—finite per-index band). If Ω is finite and the weight dynamic range is bounded as wmax≤κwmin with κ≥1, then for every i, b2−δκB2w(i)≤λi≤B2+δκB2w(i).In particular, for normalized Solomonoff weights ∥uω∥ ≡ 1, wω=2−ℓ(ω), with δκ<1, (1−δκ)2−ℓi≤λi≤(1+δκ)2−ℓi,so λi≍2−ℓi per index. (Perturbation—finite per-index band). If Ω is finite and the weight dynamic range is bounded as wmax≤κwmin with κ≥1, then for every i, b2−δκB2w(i)≤λi≤B2+δκB2w(i).In particular, for normalized Solomonoff weights ∥uω∥ ≡ 1, wω=2−ℓ(ω), with δκ<1, (1−δκ)2−ℓi≤λi≤(1+δκ)2−ℓi,so λi≍2−ℓi per index. Throughout, either Ω is finite or ∑ωwω<∞ (automatic for Solomonoff weights by Kraft) together with (H1), in which case T is trace class and its nonzero spectrum is discrete. In the infinite case, the Gershgorin localization applies to nonzero eigenvalues by the maximal-coordinate argument for ℓ2 eigenvectors. The spectral point zero may be an eigenvalue when the operator has a kernel, or it may lie in the continuous spectrum; no stronger classification is used here. For uniform weights on an infinite dictionary, the operator need not be compact, so the flat-spectrum statement is restricted to finite Ω. The nonzero spectra of AA* and A*A=S coincide (standard), and S is real symmetric PSD with diagonal Sωω=wω∥uω∥2∈[b2wω,B2wω] by (H1). (1) Exact case. If vω are orthogonal, then S=diag(wω∥uω∥2), for which the eigenvalues are the sorted diagonal products a[i]; only when the feature norms are constant does this sorting coincide with the weight order. (2) Global localization. By Gershgorin’s circle theorem in the finite case, every nonzero eigenvalue λ satisfies |λ−Sωω| ≤∑ω′≠ω|Sωω′|≤δSωω for some index ω. Thus, λ∈[(1−δ)Sωω,(1+δ)Sωω]. Taking the extrema gives the finite global band exactly as stated. For countably infinite trace-class S, a nonzero eigenvector has a coordinate of maximal absolute value (its coordinates tend to zero); applying the eigenvalue equation at that coordinate gives the same local interval. Because positive summable weights need not have a positive infimum, no uniform positive lower band is asserted in the infinite case. (3) Per-index band. Split S=D+E with D=diag(Sωω) and E the off-diagonal part (symmetric, zero diagonal). The eigenvalues of D are its diagonal entries; let S(1)≥S(2)≥⋯ be their decreasing order statistics, so λi(D)=S(i). For every i, Weyl’s perturbation inequality for symmetric matrices gives|λi−S(i)|≤∥E∥op≤maxω∑ω′≠ω|Sωω′|≤δmaxωSωω=δS(1)≤δB2wmax, where ∥E∥op≤maxω∑ω′≠ω|Sωω′| is the row-sum (Schur) bound for a symmetric matrix and the third inequality is (H2δ). Next, the pointwise bracketing b2wω≤Sωω≤B2wω of (H1) passes to the order statistics: using the max–min representation S(i)=max|T|=iminω∈TSωω, which is monotone in each entry, and the fact that scaling by b2,B2>0 preserves order,b2w(i)≤S(i)≤B2w(i)(alli).Combining the two displays with wmax≤κwmin≤κw(i),λi≤S(i)+δB2wmax≤B2w(i)+δκB2w(i)=B2+δκB2w(i),λi≥S(i)−δB2wmax≥b2w(i)−δκB2w(i)=b2−δκB2w(i), which is the stated per-index band; the normalized Solomonoff specialization sets b=B=1. Without the dynamic-range control wmax≤κwmin, the additive error δS(1) need not be small relative to w(i) at large i; the global band of part 2 remains valid in the finite case, while the stated infinite-dimensional conclusion is local. There is also a sharper consequence of the row condition when δ<1: with D=diag(S),|x⊤(S−D)x|≤∑ixi2∑j≠i|Sij|≤δx⊤Dx.Hence (1−δ)D⪯S⪯(1+δ)D, and the min–max principle gives relative eigenvalue bounds without a dynamic-range factor. For a trace-class dictionary, extend this quadratic-form inequality from finitely supported vectors by continuity. □ Two limitations concerning the informativeness of (H2δ) require attention. (i) Across length classes, the weight dynamic range can be enormous (w ranging over 2−1,…,2−L); since the radius/diagonal ratio for a deep feature against a shallow one scales like 2(ℓ−ℓ′)/2〈uω,uω′〉, (H2δ) with small δ requires the cross-class coherence to decay with the length gap or a bounded effective dynamic range (multiplicity within length classes). When this fails, the Gershgorin band is loose, indeed vacuous, once δ≥1. (ii) The synthetic spectral-calibration study of Section 14 confirms the sharp statement, i.e., for the generalized synthetic orthogonal dictionary, the ratio λi/2−ℓi is 1 to machine precision, and under increasing coherence it degrades gracefully, staying in [0.95,1.02] at coherence 0.05 and [0.75,1.02] at coherence 0.1, so that the true spectrum tracks 2−ℓi far more tightly than predicted by the worst-case Gershgorin envelope. Strictly positive exponential D2KE features at finite distances cannot be exactly orthogonal under a nonzero reference probability. The exact synthetic check therefore concerns a generalized feature dictionary. The theorem states that the correspondence is exact under orthogonality and robust under perturbation, with (H2δ) being the same near-orthogonality condition under which Part IV’s analogous result holds. Propositions 8 and 9 show that the two classical kernels differ only in the landmark measure: on the same GW-selected geometry, uniform μ yields the KC-kernel and μ∝2−ℓ (Occam) yields the Solomonoff kernel. The correct division of labor is that the distance d determines the raw feature dictionary {ϕω}, the measure μ weights the feature covariance contributions, and the eigenvalues and eigenfunctions of the resulting operator are determined jointly by the weights and the feature Gram matrix. It is false in general that d fixes the eigenfunctions while μ fixes only the eigenvalues; for coherent (non-orthogonal) features, changing μ rotates the eigenvectors as well, as shown by an immediate two-feature counterexample. Only under explicit orthogonality of the features, simultaneous diagonalization, or a perturbation estimate with appropriate spectral gaps may one keep the basis fixed or approximately fixed. The row condition (H2δ) by itself controls eigenvalues, not individual eigenvectors: GWselectsd︸featuredictionary/relationalOccamreweightingsetsμ︸spectrumand,ingeneral,theeigenbasis. GW model selection alone, being isometry-invariant, fixes only the relational layer; recovering the Solomonoff kernel (not merely a KC-kernel) requires the additional Occam choice of μ, and reading 2−ℓi as the eigenvalue law requires the conditioning hypotheses. Predictive (induction-level) fidelity additionally requires that the marking/labeling be matched, which is the role of the marked objective (10). 11.3. Category-Specific Kernelization and Induction Functors The selected space is not automatically a kernel, and a kernel is not automatically an induction rule. The full pipeline isSolomonoffobject⟶bestapproximatingspace⟶kernel/operator/predictor⟶induction.For a structured class C, letAC★∈arg minA∈CDC(SSol,A)+λCode(A)+ηPredLossC(A;D).To turn AC★ into a learning object, one must specify a representation or prediction functorKC:C⟶{positive-semidefinitekernels,operators,regularizers,summaries,orpredictors}.The Kolmogorov/Solomonoff distinction is often not a different geometry but a different weighting convention on the same family of features:Kolmogorovprojection=algorithmicfeaturesplusempiricaloruniformmass,Solomonoffprojection=algorithmicfeaturesplusOccammass2−Code.For every feature sum or integral below, require measurable features and finite diagonal second moments, ∑ωwω|ϕω(x)|2<∞ for every input (or its integral analogue). Cauchy–Schwarz then ensures that cross terms converge. Kraft summability alone does not control unbounded features. Accordingly, a category may admit two related functors KKCC and KSolC. The following paragraphs state what those functors mean in each branch. Metric–measure spaces. This is the developed case. For A=(Z,d,μ),KC(A)(z,z′)=∫Ze−γd(z,ω)e−γd(z′,ω)dμ(ω).With algorithmic distance and empirical or uniform landmark mass, this is the KC/D2KE kernel. With the same features but Occam mass dμSol(ω)∝2−ℓ(ω), it is the Solomonoff feature mixture. The distance carries algorithmic similarity; the measure distinguishes empirical/Kolmogorov from Solomonoff weighting. Hilbert spaces. If the selected approximation is already Hilbertian,A=(Z⊂H,∥·∥H,μ), then the kernel may be the inner product kH(z,z′)=〈z,z′〉H. If only Hilbert distances are available, choose a base point z0 and recoverkH(z,z′)=12dH(z,z0)2+dH(z′,z0)2−dH(z,z′)2.When the embedding preserves NCD, CKC, AMI, or another algorithmic distance with empirical mass, this is a Kolmogorov-type Hilbert kernel. With posterior/Occam weighting, the natural Solomonoff version is a covariance kernel such askSol,H(x,y)=Covh∼πDϕh(x),ϕh(y), which is a centered variant of the SFM construction. It equals the uncentered second-moment kernel only when the feature means vanish. Banach spaces. A Banach space has a norm but no canonical inner product, so it need not yield an RKHS directly. Three routes remain natural. First, choose a measure ρ on dual functionals ℓ∈B* and definekB(x,y)=∫B*σ(ℓ(x))σ(ℓ(y))dρ(ℓ).An empirical dictionary produces the Kolmogorov projection, whereas ρ(ℓ)∝2−Code(ℓ) yieldskSol,B(x,y)=∑ℓ2−Code(ℓ)σ(ℓ(x))σ(ℓ(y)).Second, one may retain an RKBS or duality pairing rather than force a Hilbert representation; this is appropriate for L1, bounded-variation, Besov, Barron, and variation-norm geometries. Third, one may apply D2KE after Banach selection:kB(x,y)=∫Be−γ∥x−ω∥Be−γ∥y−ω∥Bdμ(ω), which is positive semidefinite even when the Banach norm is not Hilbertian. Topological spaces. A bare topology gives no kernel or predictor. It must be represented by features. Choose a filtration, compute persistent homology, and map each object to a persistence landscape [61], image [62], diagram feature, or related vector [63],Φtop(x)=persistencefeatureofx.Then, ktop(x,y)=〈Φtop(x),Φtop(y)〉 is a topological kernel. A Kolmogorov version uses filtrations induced by algorithmic distance, prefix depth, AMI, or compressor geometry. A Solomonoff version additionally weights classes, covers, branches, or filtration events by code:kSol,top(x,y)=∑η∈T2−Code(η)Φη(x)Φη(y). Graphs, trees, and ultrametrics. Graphs and trees have native kernels: diffusion kernels (e−tL)(x,y) [64], reversible random-walk resolvents (I−βS)−1=∑m≥0βmSm, where S is the symmetric realization of the transition operator and 0≤β<1, subtree kernels [65], and prefix kernels. For a prefix tree,kpref(x,y)=∑u⪯x,yw(u).Uniform or empirical w(u) gives a Kolmogorov/tree kernel; w(u)=2−|u|, 2−K(u), or a depth-summable modification of M(u) gives the tree-native Solomonoff kernel of Equation (16). Operator and spectral models. If the selected object is a positive operator T with measurable eigenfunction representatives and ∑iλi|ei(x)|2<∞ at every input, its pointwise kernel iskT(x,y)=∑iλiei(x)ei(y).The Kolmogorov version lets the eigenvalues come from empirical geometry or learned covariance. The Solomonoff version requires the spectral calibration λi≈2−ℓi developed in Section 16. This is the operator-level form of the same rule: empirical or geometry-derived weights give KC-type kernels, while description-length weights give Solomonoff kernels. Induction requires a predictive rule. The category-specific functor supplies a kernel, operator, feature map, Banach/RKBS structure, topological summary, or graph diffusion. Induction begins only after one supplies a predictive mechanism: kernel ridge regression, a Gaussian-process prior, a Bayesian posterior over explanations, a Markov law, a regularized empirical risk problem, a context-tree rule, or a mixture over candidate spaces. Therefore, a complete inductive candidate is not merely a space A but a tupleA=(A,DA,KA,PA,Code(A)), where DA is the native distortion, KA is the representation functor, and PA is the induced predictive law. The strongest Solomonoff-like construction is not a single selected object but a mixture over such candidates. 11.4. Category-Native Analogues of Solomonoff Gaussian Processes The maturity map in Table 4 should be read constructively. The Solomonoff Gaussian process is the Gaussian/Hilbert realization of a more general rule: approximate the Solomonoff source, calibrate the representation by description length or Occam mass, and use the native predictor of the target category. This becomes Solomonoff-inspired TDA in topological spaces, Occam-weighted variation-norm learning in Banach spaces, algorithmic diffusion on graphs, prefix/context-tree prediction on trees, and a mixture over coded spaces at the universal level. These are projection rules rather than assertions of equivalence: each row indicates which data structure, comparison, representation, and predictor replace the Hilbert/Gaussian construction. 11.5. How Induction Looks in Each Approximation Category WritePC(·∣D)=IndC(AC★), where IndC is the native learning rule. Different categories preserve different projections of Solomonoff induction; none is the full universal semimeasure unless the category itself is the universal class of computable semimeasures; thus, this subsection provides a constructional dictionary. Universal dominance is available only at an appropriate mixture level, and a single Hilbert, Banach, topological, graph, spectral, or Gaussian candidate carries only the projection described below. Metric–measure induction. For A=(Z,d,μ),(Z,d,μ)⟼kA⟼HkA⟼KRR,GP,orBayesianprediction.With the D2KE kernel, one may perform kernel ridge regression,f^=arg minf∈HkA∑i=1nℓ(f(zi),yi)+ρ∥f∥HkA2, or Gaussian-process prediction, f∼GP(0,kA). The Kolmogorov version uses algorithmic/compression distance with empirical mass. The Solomonoff version keeps the same geometry but uses μ(ω)∝2−ℓ(ω), yielding the SFM/SKCO/SGP route. Metric induction without a native measure. For A=(Z,d), one must add empirical, diffusion, or Occam weights before obtaining a probabilistic model. When d2 is conditionally negative definite, one can use kd(z,z′)=exp(−γd(z,z′)2); otherwise, a finite-landmark D2KE construction remains available. A purely metric alternative is the algorithmic Nadaraya–Watson ruley^(x)=∑iwie−γd(x,xi)yi∑iwie−γd(x,xi).The Kolmogorov version uses NCD/CKC/AMI distance and empirical weights; the Solomonoff version uses wi∝2−K(xi) or explanation weights wh∝2−ℓ(h). Metric induction approximates Solomonoff by proximity in algorithmic distance. Hilbert induction. For Z⊂H, induction is ordinary Hilbert space learning:f^=arg minf∈H∑iℓ(〈f,zi〉H,yi)+ρ∥f∥H2.Equivalently, after kernelization, one uses RKHS regression or a Gaussian expansion:f=∑iλiξiei,ξi∼N(0,1).The Hilbert-valued Gaussian expansion requires ∑iλi<∞; pointwise GP evaluation also requires the diagonal convergence stated above. The Solomonoff-calibrated version imposes λi≍2−ℓi and kSol(x,y)=∑i2−ℓiei(x)ei(y). This is a second-order covariance approximation, useful for regression and uncertainty quantification but lossy relative to the discrete universal mixture. Banach induction. For A=(B,∥·∥B,μB), prediction is typically the regularized empirical riskf^=arg minf∈B∑iℓ(f(xi),yi)+ρ∥f∥Bp,p≥1.This formula requires bounded evaluation functionals on the chosen function space. For L1 or general BV equivalence classes, replace f(xi) by specified bounded observation functionals; point values are not intrinsically defined. Existence of a minimizer additionally requires the usual coercivity and lower-semicontinuity/compactness hypotheses. A Solomonoff–Banach construction uses the Occam-weighted atomic gauge∥f∥Sol-Var=inf∑ω2ℓ(ω)|aω|:f=∑ωaωϕω.The series is required to converge in the declared ambient space. This gauge is a norm on its finite-valued span only under separation/nondegeneracy hypotheses; no norm claim is made without them. Banach induction thus approximates Solomonoff by sparse or variation-norm aggregation of simple explanations rather than by Gaussian smoothing. Topological induction. For A=(Z,τ), prediction passes through summaries(Z,τ)⟼filtration⟼persistentfeatures⟼predictor.An algorithmic-distance filtration can be mapped to a persistence feature Φtop(x), after which ktop(x,y)=〈Φtop(x),Φtop(y)〉 supports KRR or classification; statistical foundations for such diagram-valued features are available [66,67]. The Solomonoff-topological version weights persistent classes, births, deaths, branches, covers, or filtration events by 2−Code. It preserves multiscale shape and connectivity rather than all pairwise distances. Graph induction. For A=(G=(V,E),w,μ), use graph diffusion, Laplacian regularization [68,69], graph kernels, or random walks. A semi-supervised learner can solvef^=arg minf:V→R∑i∈Lℓ(f(i),yi)+ρf⊤LGf or use the diffusion kernel kt=e−tLG [70]. A Kolmogorov graph has edge weights wij=e−γdK(xi,xj). A Solomonoff graph may use node masses μi∝2−K(xi) or joint-explanation weightswij=∑ω:ωexplainsi,j2−ℓ(ω).Graph induction approximates Solomonoff by diffusion over algorithmic-similarity networks. Tree and ultrametric induction. For A=(T,dtree,μT), especially a prefix tree, induction is naturally context-tree or prefix-mixture prediction [71,72],P(a∣x0. Write Kij=kSGP(zi,zj) and kZ(z)=(kSGP(z1,z),…,kSGP(zN,z))⊤. Then the posterior mean is (13) mN(z)=E[f(z)∣y]=kZ(z)⊤(K+σ2I)−1y=∑i=1NαikSGP(zi,z),α=(K+σ2I)−1y. The same function is the unique minimizer (14) mN=arg ming∈HSGP1N∑i=1Nyi−g(zi)2+σ2N∥g∥HSGP2. If a reference measure provides the Mercer identification of the RKHS with the range of the covariance square root (quotienting any zero-observation directions), and the operator has eigenpairs (er,λr) with λr>0, then for g=∑rarer, ∥g∥HSGP2=∑rar2λr. Consequently, under Solomonoff calibration λr≍2−ℓr, ∥g∥HSGP2≍∑r2ℓrar2: Occam weights in the SGP covariance become inverse-Occam penalties in the SGHS norm. The joint Gaussian law of (f(z1),…,f(zN),f(z)), together with the independent observation noise, gives the first two equalities in (13) by Gaussian conditioning. For the variational statement, decompose any g∈HSGP as g=g‖+g⊥, where g‖ belongs to the span of the kernel sections kSGP(zi,·). The reproducing property gives g⊥(zi)=0 for every i, while ∥g∥2=∥g‖∥2+∥g⊥∥2. Hence the minimizer of (14) has g⊥=0. Substitution of g=∑iαikSGP(zi,·) and the normal equations give (K+σ2I)α=y, which is exactly the conditional-mean coefficient vector. The spectral norm identity is the standard Mercer-coordinate description of the RKHS and yields the final equivalence when λr≍2−ℓr. □ The posterior mean and the representer theorem are two descriptions of the same estimator, not two successive lossy projections. Finiteness follows from the N observations once the kernel is fixed. Computability additionally requires that kSGP be supplied by a declared finite or effectively computable code, dictionary, or truncation. Exact Kolmogorov weights, an unrestricted universal mixture, or an untruncated exact Solomonoff covariance are not made computable merely by Gaussian conditioning. For a fixed computable kernel and Gaussian likelihood, however, prediction reduces to the finite linear system in (13). Mixture over coded spaces. The strongest category-native construction is not to select one category or one space. Let A=⋃CC be a countable coded collection of trees, graphs, Hilbert spaces, Banach models, kernels, topological models, and predictive laws, each carrying PA. Programmatically,PSpaceMix(·∣D)=∑A∈Aqβ(A∣D)PA(·∣D), with an energy combining native distortion, description length, and evidence. The metric–measure specialization is proved in Section 12. A cross-category mixture requires a fixed data-independent distortion calibration or another sequentially valid weighting rule; it is not obtained by substituting arbitrary empirical scores into the proved redundancy bound. Table 7 summarizes the per-category induction dictionary. Tree induction preserves prefix/program hierarchy; metric induction preserves algorithmic similarity; metric–measure induction preserves similarity plus Occam mass; Hilbert and Gaussian induction preserve second-order covariance geometry; Banach induction preserves sparse or variation structure; topological induction preserves multiscale shape; graph induction preserves diffusion and adjacency structure; spectral induction preserves complexity decay; predictive induction preserves conditional laws; and mixture-over-spaces induction preserves the Solomonoff mixture principle itself. 11.6. The Parent Object: A Schoenberg Metric–Measure Space A basepointed Hilbertian metric and a reference measure determine a centered kernel, its RKHS, and its covariance operator. Applied to a kernel metric, this construction recovers a basepoint-centered version of the original kernel. Exact recovery of an arbitrary original kernel requires its feature-origin data as well. A triple (X,d,ρ) is a Schoenberg mm-space if (X,d) is a complete separable metric space, ρ is a Borel probability measure on X, and d is of negative type, i.e., −d2 is conditionally positive definite; equivalently, e−γd2 is positive semidefinite for every γ>0 (Schoenberg); equivalently again, d is the kernel metric dk(x,y)=k(x,x)+k(y,y)−2k(x,y) of some PSD kernel k (for the canonical algorithmic distance, this gate is known to fail; [76] states exactly this Schoenberg equivalence (his Lemma 32 (iii): “d is Euclidean iff d is a metric and d2 is CND”) and proves dK does not satisfy it (Thm. 38); D2KE in Section 11.2 supplies one valid PSD representation, although it is not the only possible kernelization and need not preserve the original distances). Let (X,d) be separable with d of negative type (d2 conditionally negative definite), let x0∈X be a basepoint, let ρ be a Borel probability measure on X with d(·,x0)∈L2(ρ) and d(·,x0)2 measurable, and define the basepointed kernel kx0(x,y)=12d(x,x0)2+d(y,x0)2−d(x,y)2. Then: 1. kx0 is positive semidefinite with dkx0=d and kx0(x0,·)≡0 (Schoenberg); different basepoints give different kernels with the same kernel metric, and the covariance origin depends on x0. 2. The RKHS Hkx0 is determined by d and x0; as a function space modulo constants, it is basepoint-independent. 3. Since ∫Xkx0(x,x)dρ(x)=∫d(x,x0)2dρ(x)<∞ by hypothesis, the integral operator Tkx0f=∫Xkx0(·,y)f(y)dρ(y) on L2(X,ρ) is self-adjoint, positive, and trace-class, with trTkx0=∫d(·,x0)2dρ; without this integrability hypothesis, an arbitrary Borel measure need not yield a trace-class operator.4. The centered Gaussian measure N(0,Tkx0) on L2(ρ) is well defined, and its Cameron–Martin space is Ran(Tkx01/2) with the range norm. If I:Hkx0→L2(ρ) is the canonical inclusion, then Tkx0=II*, and this Cameron–Martin space is isometrically the image of Hkx0/kerI under I, completed in the induced quotient/range norm. 1. kx0 is positive semidefinite with dkx0=d and kx0(x0,·)≡0 (Schoenberg); different basepoints give different kernels with the same kernel metric, and the covariance origin depends on x0. kx0 is positive semidefinite with dkx0=d and kx0(x0,·)≡0 (Schoenberg); different basepoints give different kernels with the same kernel metric, and the covariance origin depends on x0. 2. The RKHS Hkx0 is determined by d and x0; as a function space modulo constants, it is basepoint-independent. The RKHS Hkx0 is determined by d and x0; as a function space modulo constants, it is basepoint-independent. 3. Since ∫Xkx0(x,x)dρ(x)=∫d(x,x0)2dρ(x)<∞ by hypothesis, the integral operator Tkx0f=∫Xkx0(·,y)f(y)dρ(y) on L2(X,ρ) is self-adjoint, positive, and trace-class, with trTkx0=∫d(·,x0)2dρ; without this integrability hypothesis, an arbitrary Borel measure need not yield a trace-class operator. Since ∫Xkx0(x,x)dρ(x)=∫d(x,x0)2dρ(x)<∞ by hypothesis, the integral operator Tkx0f=∫Xkx0(·,y)f(y)dρ(y) on L2(X,ρ) is self-adjoint, positive, and trace-class, with trTkx0=∫d(·,x0)2dρ; without this integrability hypothesis, an arbitrary Borel measure need not yield a trace-class operator. 4. The centered Gaussian measure N(0,Tkx0) on L2(ρ) is well defined, and its Cameron–Martin space is Ran(Tkx01/2) with the range norm. If I:Hkx0→L2(ρ) is the canonical inclusion, then Tkx0=II*, and this Cameron–Martin space is isometrically the image of Hkx0/kerI under I, completed in the induced quotient/range norm. The centered Gaussian measure N(0,Tkx0) on L2(ρ) is well defined, and its Cameron–Martin space is Ran(Tkx01/2) with the range norm. If I:Hkx0→L2(ρ) is the canonical inclusion, then Tkx0=II*, and this Cameron–Martin space is isometrically the image of Hkx0/kerI under I, completed in the induced quotient/range norm. The frequently quoted normalization k=1−12d2 is not generic; it is the special case of a constant-norm (spherical) embedding ∥Φ(x)∥≡1 and may be used only when that normalization is separately justified. (1) By Schoenberg’s theorem, negative type of d is equivalent to the existence of a Hilbert embedding Φ0:X→H with ∥Φ0(x)−Φ0(y)∥=d(x,y). Center at the basepoint: Φ(x)=Φ0(x)−Φ0(x0). Then, 〈Φ(x),Φ(y)〉=12[∥Φ(x)∥2+∥Φ(y)∥2−∥Φ(x)−Φ(y)∥2]=kx0(x,y), which is a Gram matrix; hence, PSD and dkx0(x,y)2=kx0(x,x)+kx0(y,y)−2kx0(x,y)=∥Φ(x)−Φ(y)∥2=d(x,y)2. Changing x0 translates Φ, which changes kx0 but not dkx0. (2) is Moore–Aronszajn plus the observation that re-basing changes functions by additive constants. (3) kx0 is a measurable PSD kernel with ∫kx0(x,x)dρ<∞; therefore, the associated integral operator is positive self-adjoint with trTkx0=∫kx0(x,x)dρ(x)<∞ (monotone convergence over an orthonormal basis), and hence trace class. (4) A positive trace-class covariance on a separable Hilbert space determines a centered Gaussian measure; its Cameron–Martin space is Ran(T1/2) with ∥h∥CM=∥T−1/2h∥L2(ρ). The inclusion I is Hilbert–Schmidt under the diagonal-integrability hypothesis and satisfies T=II*. Polar decomposition identifies Ran(T1/2), with its range norm, with the included RKHS after quotienting by kerI and completing. □ For k∈{kKC,kSol}, retain a common reference probability ρ and choose a basepoint x0. Quotient any zero-distance pairs and impose the integrability hypotheses of Proposition 11. Then the construction F(X,dk,x0,ρ) produces the kernel kx0(x,y)=k(x,y)−k(x,x0)−k(x0,y)+k(x0,x0) and its RKHS, integral operator, and centered Gaussian measure. Substitute dk(x,y)2=k(x,x)+k(y,y)−2k(x,y) into the basepoint formula. The remaining conclusions follow from Proposition 11. □ The metric and basepoint determine kx0, not an arbitrary original k. For example, k and k+c, c>0, have the same kernel metric and can have different RKHS dimensions. The reference measure remains necessary for the integral operator and Gaussian measure. Occam feature weights yield the stated eigenvalue calibration only under the conditioning hypotheses of Theorem 2. The Cameron–Martin space has Gaussian measure zero for infinite-rank covariance and measure one for finite-rank covariance. A PSD kernelization such as D2KE gives a Hilbertian kernel pseudometric; its zero-distance quotient is a metric space. Suppose a feature map Φ:X→H into a real Hilbert space is retained. Then k(x,y)=〈Φ(x),Φ(y)〉H is PSD and determines Hk. Its kernel pseudometric is dk(x,y)=∥Φ(x)−Φ(y)∥H. Applying Proposition 11 to the metric quotient and a basepoint gives the translated-feature kernel 〈Φ(x)−Φ(x0),Φ(y)−Φ(x0)〉H. Thus KC-HS and SGHS are recovered exactly from their respective retained kernels or feature maps; their metrics alone recover only the basepoint-centered constructions. If the feature norms are separately known to equal one, then k=1−12dk2. The PSD property is the Gram identity; expanding the squared difference proves the metric formula and the final normalization. □ The unnormalized complexity metric studied in [76] is not Hilbertian, and that work proves a non-Euclidean power obstruction for exponents greater than 0.3. Its extension to normalized information or compression distances is not established there. A finite surrogate must therefore be checked separately for metricity and for conditional negative definiteness of its squared distances. D2KE supplies a PSD kernel for any measurable nonnegative dissimilarity and finite landmark measure, but it can change distances. It is one kernelization option, not the unique route to a Hilbert representation. A specified kernel determines its RKHS; its eigenfunctions also depend on the reference measure, and its Solomonoff spectral calibration still requires the feature assumptions of Theorem 2. 11.7. Geometry-First Versus Kernel-First The kernel-first route, including Parts I–VI, searches over kθ (equivalently, over its RKHS Hkθ) and reads off dkθ and Tkθ. The geometry-first route proposed here searches over (Zθ,dθ,μθ) and then obtains kθ, Hkθ, and Tkθ. Figure 3 contrasts the two routes. This ordering is the main conceptual contribution of the paper. 11.8. Stability of Kernelization The next result states that if a candidate geometry is close to the empirical Solomonoff geometry in GW, then the induced D2KE kernels are close after transport by the optimal coupling. Let M=(X,dX,μX) and N=(Y,dY,μY) be finite mm-spaces with diameters bounded by D (the diameter bound is not needed for the constant below; it is retained for the transfer statements that reuse this setting). Let kX,kY be their D2KE kernels with the same γ>0. For any coupling π∈Π(μX,μY), define the transported kernel discrepancy Δπ2=∫|kX(x,x′)−kY(y,y′)|2dπ(x,y)dπ(x′,y′). Then, there exists a constant Cγ,D such that Δπ≤Cγ,D∫|dX(x,x′)−dY(y,y′)|2dπ(x,y)dπ(x′,y′)1/2. In particular, for an optimal GW coupling and the normalization of Definition 2, Δπ★≤2Cγ,DdGW,2(M,N). Write the D2KE features ψωX(x)=e−γdX(x,ω)∈[0,1] so thatkX(x,x′)=∫ψωX(x)ψωX(x′)dμX(ω), and likewise for Y. Couple the landmark integration variable by the same π. For fixed coupled landmark pair (ω,ϖ), use r↦e−γr being γ-Lipschitz and all factors lying in [0,1]:|ψωX(x)ψωX(x′)−ψϖY(y)ψϖY(y′)|≤|ψωX(x)−ψϖY(y)|+|ψωX(x′)−ψϖY(y′)| since |ab−cd|≤|a||b−d|+|d||a−c|≤|b−d|+|a−c| when |a|,|d|≤1. Each term is ≤γ|dX(x,ω)−dY(y,ϖ)| (resp. with x′,y′,ϖ). Integrating the kernels’ difference against π⊗π on the points and using the same coupling to compare landmarks then applying Cauchy–Schwarz over the pair-coupling givesΔπ≤2γ∫∫|dX(x,x′)−dY(y,y′)|2dπ(x,y)dπ(x′,y′)1/2.With the convention dGW,22=12infπ∫∫|dX−dY|2dπdπ, an optimal coupling yieldsΔπ★≤22γdGW,2(M,N).Thus, one may take Cγ,D=2γ in the integral-discrepancy display and 22γ in the final GW-normalized display; diameter-dependent variants only change this constant. □ The constant 22γ is rigorous but pessimistic. The finite GW-to-kernel stability study of Section 14 measures Δπ★/dGW,2≤0.69 over random mm-space pairs at γ=1, inside the bound. This experiment is only a small finite sanity check, not a statistical estimate of a worst-case constant. This proposition is the formal bridge from geometry selection to kernel approximation: small Solomonoff–Gromov distortion implies small discrepancy between the resulting transported kernels. It is the kernel-level analogue of saying that the candidate geometry preserves the empirical algorithmic relational structure. 12. Mixtures over Coded Metric–Measure Models The selector above returns one metric–measure geometry. A mixture retains uncertainty over a finite or countable coded family of mm-models, each carrying an induced predictor. This section proves the mixture result within that family. A mixture spanning the other rows of Table 4 is a programmatic extension of Equation (1), not a theorem that is claimed here. The estimator (5) returns a single geometry θ^, and hence a single RKHS/GP. With a genuine prior code, it is a MAP choice; with the experiment’s BIC/MDL score, it is a penalized-model-selection analogue. By contrast, Solomonoff induction is not a single model; it is the Bayesian mixtureM(x)=∑ν2−K(ν)ν(x) over lower-semicomputable semimeasures with a universal description weighting. This universal-mixture representation is understood up to multiplicative constants relative to a universal monotone semimeasure. Selecting one geometry replaces a mixture by its mode, a second collapse on top of Gaussianization. The level of abstraction that moves closer to M than the RKHS route is to keep the mixture over spaces. Independently of this Solomonoff mixture argument, one might ask whether a single kernel over geometries, e−γdGW2, would already suffice. On the candidate families of Section 14, this Gram is positive semidefinite to solver tolerance across instances, so a kernel over geometries is empirically admissible; the mixture is preferred for the reason given here, not because the Schoenberg gate fails one level up. 12.1. The Universal-Form Space-Mixture Posterior Let A be a countable coded collection of complete metric–measure candidates A=(MA,DA,KA,PA,Code(A)), where DA=dGW,pp(SSol,MA) is a population-level mm-space distortion fixed independently of the realized prediction sequence. The map KA kernelizes MA, the law PA(·∣D) is its induced predictor, and LA(D) is its evidence or likelihood on the observed data D. For an inverse temperature β≥0 satisfying 00. For β≥β0, each nonminimal weight e−β(EA−E*) tends to zero and is dominated by the summable sequence e−β0(EA−E*). Dominated convergence shows that their total tends to zero. Each minimum has unnormalized relative weight one, so its limiting probability is 1/m. □ 12.2. Why the Mixture Is Closer: Dominance and Regret Fix β≥0. Assume all component laws are probability laws on the same sequential sample space (with termination included when necessary), and all conditionals below have positive denominators along the evaluated history. Assume: (i) the candidate class A is finite or countable with Kraft-summable codes, ∑A∈A2−Code(A)≤1; (ii) each GW distortion DA≥0 is a data-independent constant of the candidate (a property of the population source and the candidate, not of x1:n); and (iii) the sequential predictor is the Bayes conditional of the space-weighted joint mixture evidencePβ(x1:n)=∑A∈A2−Code(A)e−βDAPA(x1:n),PSpaceMix(xt∣x0 distortion tilt is not exercised in the numerical validation programme of Section 14. Let AN⊂A be a finite coded subset and define the truncated joint mixture Pβ,N(x1:n)=∑A∈AN2−Code(A)e−βDAPA(x1:n). For every included candidate A★∈AN, −logPβ,N(x1:n)+logPA★(x1:n)≤Code(A★)ln2+βDA★. If Pβ,A denotes the full countable mixture and rN(x1:n)=1−Pβ,N(x1:n)Pβ,A(x1:n)∈[0,1), then the additional unnormalized negative log-evidence paid by truncating to AN is exactly −logPβ,N(x1:n)+logPβ,A(x1:n)=log11−rN(x1:n). These identities assume positive full and truncated evidence. For normalized sequential predictors, let WN=Pβ,N(ϵ) and W=Pβ,A(ϵ). The cumulative log-loss difference is instead LossN−Lossfull=log11−rN+logWNW. If truncated evidence vanishes, its loss is infinite. Finite SpaceMix remains competitive with each included candidate by the first inequality (normalizing its subprobability prior can only reduce its negative log-evidence). The first display is the same one-term lower bound as Proposition 15, applied to the finite sum. For the truncation identity, write Pβ,N=(1−rN)Pβ,A and take the negative logarithms. □ Proposition 16 is the version that can actually be implemented. The mm-model list may come from a finite dictionary, a prefix-code enumeration cut off by code length, beam search, reversible-jump exploration, or a hand-built collection of metric, graph-induced, tree, and manifold mm-spaces. The theorem gives no protection against a good candidate that was never enumerated. singlekernel,fixedRKHS︸PartsI–VI∩oneGW-selectedgeometry︸Section 7∩universal-formmixtureoverspaces︸thissection↓(A→alllower-semicomputablesemimeasures,Code→K)M. The GP step discards everything beyond second order when the chosen category is Gaussian, while the arg min step discards the mixture (MAP) in every category. This section undoes the second collapse by mixing over spaces. If the components are Gaussian, then PSpaceMix is a mixture of Gaussians and can therefore represent non-Gaussian and multimodal distributions; if the components include trees, graphs, Banach models, topological predictors, and direct predictive laws, then the mixture becomes broader. Full fidelity would additionally let A range over all lower-semicomputable semimeasures with true Kolmogorov weights, which this construction does not claim. PSpaceMix remains a computable surrogate: the sum runs over a coded candidate class, not all programs, and Code is a description-length proxy for K. It is closer to M than a single RKHS or a single selected space in two precise senses: first, it has the mixture form, and second, it inherits the code-plus-distortion regret bound of Proposition 15; however, it is not equal to M. For a finite candidate list on a common observation space, mixing probabilities is affine in the probability coordinates. No smooth-manifold or global information-geometric structure is assumed for the universal semimeasure class. The dominance bound gives the precise operational comparison used here. 13. The Universal Predictor and the Metric–Measure Projection Tower The mixture of Section 12 is a mixture over coded spaces and their induced predictors. Lifting it one step further, to a mixture over programs or all lower-semicomputable semimeasures, gives the most primitive object of the theory and reorganizes the whole construction as a tower of projections, with the RKHS route at the bottom. Two things are easy to get wrong here and we are explicit about both: first the form of the posterior, and second, which transformations lose information and which retain their input kernel. 13.1. Level 1: The Universal Algorithmic Experiment Fix a universal prefix machine for description complexity and a universal monotone machine Um for output prediction. The monotone form isM(x)=∑p∈Px2−|p|, where Px consists of minimal input prefixes that make Um output a string beginning with x; it is prefix-free. Consistency carries the likelihood in this presentation. A universal measure-mixture presentation uses an effective enumeration of lower-semicomputable semimeasures νp, with prefix description weights:M˜(x)=∑p2−|p|νp(x),πD(p)=2−|p|νp(D)M˜(D).Assume positive evidence. With suitable universal description weights, M˜ and M dominate each other up to constants; they need not coincide. The posterior mixture identity is exact for M˜ and its component conditionals. For probability-valued prediction, include termination as an outcome, including any initial mass deficit; this convention is also used in the Hellinger and KL comparisons below. Conditioning the monotone program experiment likewise gives a posterior on consistent inputs; one must not insert a second likelihood factor into its consistency sum. 13.2. Level 2: The Posterior Explanation Geometry Use the metric quotient and completion specified in Equation (3), with pushed-forward posterior mass:SD=(Q,dexp,πD).A predictive zero-distance quotient is canonical relative to the declared history law. A syntactic algorithmic construction instead requires a separately specified exact metric; raw conditional program complexities do not descend automatically to semantic classes. The behavioral transfer results compare a declared measurable behavior map with its image under their faithfulness assumptions. They do not assert that an arbitrary data corpus reconstructs the universal explanation space. 13.3. Level 3: The Covariance Projection Project the posterior-weighted feature family onto its second moment:Πcov:(P,πD,ϕp)⟼kD(x,y)=∑pπD(p)ϕp(x)−mD(x)ϕp(y)−mD(y), with mD=∑pπD(p)ϕp, assuming finite pointwise second moments. This centered covariance equals the uncentered SFM second moment minus mD(x)mD(y); the two agree only for zero mean. Level 4 is the existing RKHS/GP layer HkD, f∼GP(0,kD), TkD. For Gaussian regression at Level 4, Proposition 10 shows that GP conditioning and KRR give the same finite representer expansion. This is a computational realization of the Level-4 object, not another projection of the universal experiment. 13.4. The Tower and What Each Projection Discards The map from general posterior feature laws to their covariance is not injective. Consequently, a covariance kernel alone does not in general determine the posterior feature law or its program identities. Unmarked GW comparison also does not identify external task labels. In contrast, a PSD kernel and its RKHS as an evaluated function space determine one another, and a centered Gaussian process determines its covariance kernel. The scalar law assigning probability 1/2 to each of −1,1, and the law assigning probabilities 1/4,1/2,1/4 to −2,0,2, both have mean zero and variance one but are different. Applying these random scalars to a fixed deterministic feature gives identical covariance kernels and different feature laws. Relabeling marks leaves an unmarked mm-space unchanged. Finally, the reproducing identity recovers k(x,y) from its RKHS, while k(x,y)=E[f(x)f(y)] recovers covariance from the centered GP. □ A covariance representation is generally insufficient to reconstruct the universal experiment, although subsequent RKHS and centered-GP representations retain the specified kernel. Mixing over models restores a mixture representation; it does not invert covariance reduction or recover omitted programs. Figure 4 distinguishes these operations. 13.5. The Comparison Principle at the Top At Levels 3–4, one compares kernels (norms, alignment, spectra); at Level 2, geometries (GW); and at Level 1, one should compare predictive experiments. For a candidate predictor PA(·∣D), define the Solomonoff deficiencyΔ(A;D)=−logPA(D)︸LA(D)−−logM(D)︸LU(D)+K(A)ln2 and sequentially define the cumulative predictive divergenceRT(A)=∑t=1TKLM(·∣xλ★, the selected model has k=1. It is deliberately not a characterization. On the ordinary ECA corpus it fires in only 4/15 runs although collapse occurs in 15/15. In a 61-point λ sweep on three ECA and three periodic-versus-noise seeds, the largest tested value still selecting k>1 lies below λ★ by factors 1.22–3.14; no implication is violated. The exact identity in Lemma 2 is also checked numerically: the solver-to-B(D) ratio is one to numerical precision on every split. The supplementary task-aware experiment uses a nested stratified validation split to rank already constructed block kernels and the raw geometry by predictive loss plus code length. It reaches 0.785±0.139 and usually retains the raw geometry, while raw D2KE–NCD reaches 0.838±0.067 and NCD-1NN reaches 0.755±0.043 on the same 20 splits. Every raw and block geometry considered by the nested selector is evaluated with the same KRR learner; the NCD-1NN row is a separate baseline, not the predictor used for the raw candidate. These results show that predictive alignment can prevent unmarked collapse, but the geometry-selection layer has not improved on the raw D2KE compression baseline. This selector is label-aware at the model-selection level; unlike the real-data marked-selection study below, it does not optimize the fused marked-GW coupling. Accordingly, “geometric priority” is not a valid unconditional proposition. The defensible statement is that geometry supplies a structural prior, while supervised model selection requires a marked or held-out predictive term. The real-data marked-selection study below tests this requirement; no real-data superiority claim is made. Real-data marked selection on SMS spam (E11). To evaluate label-aware geometric selection on real data, we conduct a fused/marked-GW experiment on a deterministic balanced subset of the public UCI SMS Spam Collection [77]: 100 ham and 100 spam messages. Ten repeated outer splits are stratified 70/30. Within each outer training set, a second stratified split chooses η∈{0,0.25,1,4}; the candidate menu is k∈{1,2,3,4}, λ=2×10−4, the compressor is bz2, and only training labels and training landmarks form the block marks and D2KE predictor. The outer test labels are excluded from fitting and practical model selection; they are used for final scoring and the explicitly declared post-hoc oracle diagnostic. Balanced accuracy is primary. Because the ten splits reuse one fixed corpus, the displayed standard deviations are descriptive rather than independent-sample confidence intervals. Table 10 reports the results. Adding marks to the coupling is insufficient for the tested block family: the unsupervised average-linkage block partitions remain label-mixed (the best training adjusted Rand index over k is 0.00045±0.00047), so their learned marks differ too little to overcome the distortion and complexity preference for the one-block model. The post-hoc test-label oracle confirms that changing only the rule that selects among the four implemented block kernels cannot repair the pipeline: every k obtains 0.500±0.000, against 0.910±0.048 for raw D2KE–NCD, and raw geometry is strictly better on all ten outer splits. This is evidence of candidate-family or quotient inadequacy at the tested sample size, not absence of predictive information in the raw NCD matrix. Its scope is the fixed 200-message subset, the 140 outer-training observations per split, average-linkage partitions, D2KE construction, and KRR learner; it does not distinguish inherent family inadequacy from sample starvation and does not rule out a larger sample, different partition generator, supervised quotient family, kernel, or learner. A sample-size curve is required before making a population-level inadequacy claim. The raw-geometry safeguard prevents collapse but exactly matches rather than improves upon raw D2KE–NCD; the character baseline remains better. The supported finite-experiment design rule is conjunctive:task-alignedcomparison+adequatecandidateclass,includingasafesourcerepresentationwhenadequacyisunprovedNeither unmarked geometry nor marking an impoverished quotient family is sufficient. Mechanism Checks for Parts III–VI The second suite asks whether the numerical pipeline reproduces the main mechanisms of Parts III–VI where the ground truth is known. Table 11 reports the corresponding checks. They are deliberately synthetic falsifiable regression tests of mechanisms, not benchmarks. What these experiments do and do not establish. They establish, at toy scale and under fixed seeds, that BIC/MDL-penalized selection recovers planted structure on most but not all corpora; that the relational term alone never does; that the memorizer control fails without a fitted-dimension penalty; that PSD repair, spectral calibration, effective-dimension ordering, and pullback preservation behave as the theory predicts; and that the marked mechanism runs. They also establish two negative results: unmarked selection can be far worse than the raw compression kernel for prediction, and conventional character features beat all compression methods on the one natural-data task. They do not estimate distance to the true Solomonoff semimeasure, establish real-data superiority, or provide an unsupervised real-data rule for calibrating the latent-structure penalty, and are not certified at the exact GW solver level. 15. Error Decomposition The ideal object is an incomputable Solomonoff mm-space S. The learned object is a computable finite geometry M^θ^. The total error can be decomposed into conceptually distinct terms. Let S(L) be a depth-L truncation of the ideal Solomonoff space, S^n,C the empirical space built from compressor C, Mθ★ the best candidate in the chosen model class, and θ^ the output of a numerical algorithm. For any p≥1, (15) dGW,p(S,Mθ^)≤dGW,p(S,S(L))︸Solomonofftruncation+dGW,p(S(L),S^n,C)︸compressor+sampling+dGW,p(S^n,C,Mθ★)︸model-classapproximation+dGW,p(Mθ★,Mθ^)︸selectiondisplacement. This is the triangle inequality for dGW,p, applied through the chainS→S(L)→S^n,C→Mθ★→Mθ^. □ The last term is a distance between candidate geometries, not an objective optimization gap. Two exact minimizers can have positive mutual GW distance. For example, a uniform two-point source with distance 2 and equal-code candidates with distances 1 and 3 has two minimizers with squared GW distortion 1/4, while the candidates have mutual GW distance 1. Converting objective excess to selection displacement requires an additional identifiability or error-bound hypothesis. Downstream errors must be compared on a common scale. If a kernelization satisfies ∥TMπ−TN∥≤LκdGW,p(M,N), apply Lκ to the geometric triangle bound before adding an operator approximation error. Predictor risk then requires its own stability estimate. Entropic objective bias and solver objective gaps are recorded separately; they are not distances between selected geometries. Each term has a distinct interpretation. Truncation error. This is the error from replacing the universal infinite program prior by descriptions of length at most L. When the prior is prefix-free, tail mass is controlled by Kraft-type bounds, although the exact rate depends on how the truncation is defined and on whether one truncates programs, outputs, or equivalence classes. Compressor and sampling error. This term combines finite sample error with the discrepancy between K and the computable compressor C. Comparing compressors measures robustness within that family; it does not measure discrepancy from incomputable K. Sampling error can be controlled separately when a population surrogate and sampling law are specified. Approximation error. This is the price of the chosen geometry class. A tree class may approximate prefix structure well but continuous manifolds poorly. A manifold class may approximate smooth physical data well but fail on hierarchical program-like data. The decomposition is a geometric decomposition before kernelization. If the selected geometry is subsequently passed through D2KE, a heat kernel, a negative-type repair, or an SFM construction, there is an additional kernelization defect measured at the kernel/operator level, as in Proposition 11. This structural loss is not part of the GW triangle inequality itself. Optimization error. Exact GW optimization is generally difficult. Entropic GW, low-rank couplings, sliced variants, and landmark approximations introduce computational error. This term records that cost explicitly rather than hiding it inside the definition of the model. In computations it should be split, when possible, into an optimization gap and a regularization/discretization bias, for example the difference between the entropically regularized GW value and the unregularized GW value. The finite centered entropic bias is bounded in Proposition 19; geometry-dependent refinements and certification of a particular numerical run remain separate questions. Consistency of Empirical GW Let S=(X,d,μ) be a compact metric–measure space with diam(X)≤D<∞. Let X1,…,Xn∼μ be i.i.d. and Sn=({Xi}i=1n,d|Xn,n−1∑iδXi). Then, dGW,p(Sn,S)→n→∞0almostsurely. Moreover, if rn is any rate satisfying EW1(μ^n,μ)≤rn, then EdGW,p(Sn,S)≤C(D,p)rn1/p. In particular, if (X,d) has upper box-counting (Minkowski) dimension dbox, then for p=2 and every s′>max(dbox,2) there is a space-dependent constant Cs′(X,d,μ)<∞ with EdGW,2(Sn,S)≤Cs′(X,d,μ)n−1/(2s′). Here, the constant necessarily depends on the space’s covering profile, not only on its diameter and dimension exponent. More explicitly, under a covering bound N(X,ε)≤Aε−dbox, it may be taken to depend on A,D,s′. The theorem asserts the strict-exponent bound only. Critical exponents require a separate entropy or matching analysis; Euclidean matching results such as [78] do not by themselves establish a critical logarithmic rate on every compact metric space with the same box dimension. Almost-sure convergence. The identity map embeds Sn and S into the common space (X,d); using the (suboptimal) coupling induced by the empirical-to-population optimal transport plan πn∈Π(μ^n,μ),dGW,pp(Sn,S)≤12∫∫|d(x,x′)−d(y,y′)|pdπn(x,y)dπn(x′,y′)≤(2D)p−1W1(μ^n,μ).The last inequality uses|d(x,x′)−d(y,y′)|≤d(x,y)+d(x′,y′),|d(x,x′)−d(y,y′)|≤2D.Under πn, each of the two triangle terms integrates to W1(μ^n,μ); the resulting factor 2 cancels the 12 of Definition 2, so the constant is (2D)p−1 with no residual 12. By the Varadarajan/Glivenko–Cantelli theorem, μ^n→μ weakly a.s.; on a bounded metric space, this implies W1(μ^n,μ)→0 a.s., giving dGW,p(Sn,S)→0 a.s. Taking expectations in the display gives EdGW,pp(Sn,S)≤(2D)p−1rn, and Jensen’s inequality (concavity of t↦t1/p) yields EdGW,p(Sn,S)≤(2D)p−1rn1/p=C(D,p)rn1/p. The stated p=2 rate follows from the empirical W1 rates of [79,80]: for every s′>max(dbox,2), one has EW1(μ^n,μ)≤Cs′′n−1/s′ (with a space-dependent constant). Sharper and entropic-GW-specific sample-complexity results are available [81,82]. A scope caveat for dynamics is that the theorem’s sampling model is i.i.d. draws from a fixed population mm-space; data generated by a dynamical system—a single trajectory with dependent samples—are not covered as stated, and an ergodic-sampling variant (Birkhoff-type, with mixing controlling the empirical-measure rate) is an open problem recorded in Section 23. □ This isolates the sampling term of the error decomposition; the compressor term (d^Sol vs. the ideal algorithmic distance) and the truncation term are separate and are not removed by sampling alone (Section 23). The square root in this p=2 upper bound comes from the displayed W1 comparison and Jensen inequality. It is not asserted to be an intrinsic lower-bound obstruction for GW. The diameter dependence in (2D)p−1 is explicit: it is order one for normalized compression dissimilarities, which are designed to lie near [0,1] (the E7 corpus reaches only 1.1), whereas an unnormalized cost must be rescaled or retain its actual diameter in the bound. The theorem does not solve the incomputability of the true Solomonoff space. It says that if a population algorithmic mm-space is fixed and sampleable, then its empirical relational geometry is statistically consistent. The additional compressor approximation error must be handled separately. 16. Spectral Solomonoff Approximation 16.1. The Ideal Spectrum Earlier parts of the BAITML programme emphasize the complexity–spectrum correspondence [3,4,5]:λi≍2−ℓi, where ℓi is a description length associated with the i-th feature, program, or canonical equivalence class. In the simplest ideal ordering ℓi=i, this becomesλi≍2−i.This is the clean toy case: every additional bit halves the spectral mass. It serves as the ideal calibration because it expresses the exact Occam/Solomonoff intuition in spectral form. The ordering ℓi=i assumes one description per length. With realistic multiplicity—a program-counting function N(L)=#{ω:ℓ(ω)≤L}≍2αL for some α∈(0,1) (an additional counting hypothesis, not a consequence of Kraft)—the i-th sorted weight sits at length ℓi≈α−1log2i, soλi≍2−ℓi≍i−1/α,a polynomial envelope (exponent 1/α>1), not exponential-in-i. Kraft also permits counting exponent one: for sufficiently small c>0, the counts aL=⌊c2L/L2⌋ satisfy ∑LaL2−L<1 but N(L)≍2L/L2. Hence strict inequality α<1 is not forced by code summability. Under the stated counting ansatz, the exponent α is an entropy rate: α=limLL−1log2N(L) is the exponential growth rate of the description ensemble (the topological entropy of the program tree), so the sorted Solomonoff spectrum is governed by an entropy. The exponential law λi≍2−i is the zero-entropy case with one effective description per length, more precisely, ℓi=i+O(1) gives this exact comparability, whereas N(L)=Θ(L) gives only 2−Θ(i); N(L)=O(1) would instead allow only finitely many descriptions. This matters for the trichotomy below: whether the globally sorted Solomonoff spectrum looks exponential, polynomial, or plateau is a property of N(L), i.e., of the domain’s algorithmic content, not a universal constant (on exponent conventions across the series: Part III writes the polynomial regime of a fixed kernel operator as λi≍i−2s/d (Sobolev smoothness s, input dimension d), which is the generic operator exponent a=2s/d of Section 17. The program-counting exponent α here is a different object: 1/α governs the sorted Solomonoff envelope across length classes, and is unrelated to the decay exponent a of any single kernelization). Changing the reference universal machine shifts description lengths by additive constants on stable code families, and changing compressors perturbs finite distance and spectrum estimates. Raw eigenvalues and individual mode identities are therefore not invariant claims. The robust statements are regime-level: exponential toy decay when there is one effective description per length, polynomial sorted envelopes when program-counting multiplicity grows, finite-shell plateaux under truncation, and stability of these conclusions under reasonable compressor swaps. 16.2. Spectral mm-Spaces A compact positive trace-class covariance operator T with eigenvalues (λi)i≥1 defines a Gaussian measure on a Hilbert space and hence a probabilistic metric object. One may compare two such objects directly through their covariance spectra. For the purposes of this paper, define the spectral discrepancydspec2(T,TSol)=∑i≥1(λi−2−ℓi)2 whenever the sum is finite after a common ordering or matching. This quantity is not a replacement for GW on data spaces; it is an operator-level proxy. It measures whether the learned geometry, after kernelization, produces the correct Solomonoff-style spectral mass distribution. 16.3. GW Interpretation of the Spectral Trichotomy The three classical regimes become three different kinds of GW/spectral mismatch. Polynomial spectra. If λi∼i−a, then high-index modes receive much more mass than an exponential Solomonoff spectrum. Such kernels tend to over-weight high-complexity hypotheses relative to Solomonoff calibration. Here, a is the decay exponent of a fixed kernel operator, not the program-counting exponent α above. Plateau spectra. A finite-depth Solomonoff truncation has length classes. Within a fixed length class, many programs have comparable weight. Thus, a plateau followed by a cutoff is a natural finite-depth approximation within a single length class. This does not weaken the plateau results of Parts III and V: those results are capacity and finite-resolution learning statements, where the learner only sees a finite shell, truncation, or effective rank. The globally sorted Solomonoff envelope answers a different question, namely, how mass is distributed across all length classes, and by the remark above, may be polynomial with exponent 1/α. Hence, the right reading is additive: plateaux are the correct local/finite-capacity regime, while the global envelope records the growth of the program-counting function. Super-smooth spectra. Gaussian/RBF-type spectra may decay too fast, collapsing mass onto very low-frequency modes and under-weighting moderately complex but short-program structure. Such kernels look Occam-like but are not necessarily Solomonoff-calibrated. A kernel k with covariance eigenvalues λi(k) is spectrally Solomonoff-calibrated to depth L if there is a matching of its first NL modes to description classes of length at most L such that ∑i=1NL|λi(k)−cL2−ℓi|2≤εL, where cL is the normalizing constant for the truncated description prior. 16.4. Relation to Solomonoff–Gromov Selection The GW objective selects a geometry by matching pairwise distances. The spectral calibration objective tests the kernel/operator induced by that geometry. A convincing Solomonoff approximation should pass both tests:relationalcalibration:dGW(S^n,Mθ)small, andspectralcalibration:dspec(Tkθ,TSol(L))small.This two-level criterion separates generic simplicity priors from genuinely algorithmic geometry. The two levels address distinct but interacting factors of the master kernel (12), as identified in Remark 7: relational calibration tests the distance-defined feature dictionary together with the mass-weighted coupling, while spectral calibration tests the eigenvalue multiset of the resulting operator. Neither alone suffices. GW is invariant under measure-preserving isometry and does not identify an operator basis; the eigenvalue-only discrepancy dspec is blind to eigenfunctions; and changing the reference measure can rotate eigenfunctions when the features are coherent. For supervised use, a third condition is needed that matches the marking/labeling, which is supplied by the marked objective (10). Relational-plus-spectral calibration constrains the prior geometry and spectrum, but not which response attaches to which input. 17. The Metric–Measure Learning Core We now spell out how the spectral regimes of Parts III–VI reappear from the metric–measure viewpoint. The key point is simple but important: an mm-space does not have a unique spectrum by itself; it has spectra after choosing a canonical operator generated by the geometry. This is not a defect, and is exactly the same choice already made in the earlier papers when passing from an algorithmic distance to D2KE, SFM, SKCO, an SGP covariance, or a deep pullback kernel. 17.1. Operator Functors from mm-Spaces Let M=(Z,d,μ) be a bounded mm-space. A kernelization functor assigns to M a PSD kernelκM(z,z′)=Ψ(d,μ)(z,z′), and hence a compact integral operatorTMf(z)=∫ZκM(z,z′)f(z′)dμ(z′).Examples include the D2KE kernel (4), heat kernels, Schoenberg kernels, diffusion kernels on graphs, and the SFM/Occam kernel (12). The mm-space spectrum is the spectrum of TM for the chosen functor:λ1(M)≥λ2(M)≥⋯≥0.Different functors answer different questions: D2KE asks for a PSD Hilbertian lift of the distance, SFM asks for an Occam-weighted feature mixture, and heat kernels ask for diffusion regularity. The regime classification below applies to any such functor once fixed. Fix a kernelization functor M↦TM. We say that M is: 1. Polynomial with exponent a>1 if λj(M)≍j−a (three polynomial exponents appear across the series and should not be conflated. The operator decay a here is Part III’s λi≍i−2s/d, i.e., a=2s/d (with a>1 in the RKHS-embedding regime s>d/2). It is distinct from the sorted-envelope exponent 1/α of the ideal-spectrum remark in Section 16, which comes from the program-counting function N(L)≍2αL and governs how Solomonoff mass distributes across length classes rather than the decay of any one fixed operator).2. Plateau/truncated at level m if λ1,…,λm are comparable and λm+1 is separated by a gap or cutoff. 3. Supersmooth if λj(M)≲exp(−cjρ) for some c,ρ>0. 1. Polynomial with exponent a>1 if λj(M)≍j−a (three polynomial exponents appear across the series and should not be conflated. The operator decay a here is Part III’s λi≍i−2s/d, i.e., a=2s/d (with a>1 in the RKHS-embedding regime s>d/2). It is distinct from the sorted-envelope exponent 1/α of the ideal-spectrum remark in Section 16, which comes from the program-counting function N(L)≍2αL and governs how Solomonoff mass distributes across length classes rather than the decay of any one fixed operator). Polynomial with exponent a>1 if λj(M)≍j−a (three polynomial exponents appear across the series and should not be conflated. The operator decay a here is Part III’s λi≍i−2s/d, i.e., a=2s/d (with a>1 in the RKHS-embedding regime s>d/2). It is distinct from the sorted-envelope exponent 1/α of the ideal-spectrum remark in Section 16, which comes from the program-counting function N(L)≍2αL and governs how Solomonoff mass distributes across length classes rather than the decay of any one fixed operator). 2. Plateau/truncated at level m if λ1,…,λm are comparable and λm+1 is separated by a gap or cutoff. Plateau/truncated at level m if λ1,…,λm are comparable and λm+1 is separated by a gap or cutoff. 3. Supersmooth if λj(M)≲exp(−cjρ) for some c,ρ>0. Supersmooth if λj(M)≲exp(−cjρ) for some c,ρ>0. Thus the three spectral regimes of Parts III–V are not lost in the mm-space language. They become regimes of the geometry-induced operator TM. The advantage is that we can now ask why a regime appears: from program-counting growth, from a finite length shell, from diffusion geometry, from redundancy in the feature dictionary, or from a deep pullback. Let M and N be finite mm-spaces and let κ be a kernelization functor satisfying a transported Hilbert–Schmidt stability bound∥TM−TNπ∥HS≤CdGW,2(M,N)for an optimal or near-optimal coupling π, and assume the transport preserves the spectrum λj(TNπ)=λj(TN) for all j (automatic when π is induced by a measure-preserving bijection and an additional hypothesis on the kernelization functor for soft couplings, otherwise the conclusion holds with λj(TNπ) in place of λj(N)). Then,∑j≥1|λj(M)−λj(N)|21/2≤CdGW,2(M,N).Consequently, every finite-resolution spectral feature separated from its competitors by more than 2CdGW,2(M,N) is stable under the perturbation: eigenvalue clusters, finite plateaux, cutoffs, and effective ranks at thresholds away from the spectrum. This is not an asymptotic regime-inheritance statement. Asymptotic regime labels require a tail condition in addition to Hilbert–Schmidt closeness. For example, a polynomial law λj(M)≍j−a transfers to N with the same exponent if|λj(M)−λj(N)|=o(j−a),j→∞,and analogous relative-tail bounds transfer super-smooth decay. GW-close kernelizations preserve only the spectral information visible above the perturbation scale; full tail regimes transfer only under tail-stable kernelization. After transporting TN to the support of M by π, the claimed inequality is the Hoffman–Wielandt/Weyl bound for compact self-adjoint Hilbert–Schmidt operators. The D2KE finite stability result of Section 11 is one instance of the assumed transported stability bound. For finite spaces a common realization is explicit: on L2(π), lift the first kernel to k˜M((x,y),(x′,y′))=kM(x,x′), and the second similarly using y,y′. The coordinate pullbacks are isometries because π has the prescribed marginals; each lifted operator has the original nonzero spectrum, with zeros added. Their Hilbert–Schmidt difference is the coupled kernel discrepancy. Arbitrary conditional averaging along a soft coupling need not preserve spectra, and this construction alone does not establish the task transport in the regression theorem. The finite-resolution statement follows because a perturbation of size ε cannot close a spectral gap larger than 2ε. The final assertion is separate: an ℓ2 perturbation of the eigenvalue sequence alone does not determine the tail exponent, so asymptotic transfer must be assumed or proved as a relative tail bound for the chosen kernelization functor. □ 17.2. Recovering Parts III–VI The kernel constructions in Parts III–VI reappear in a fixed order once the three maps are separated:algorithmicdatageometry⟶selectedmm-space⟶kernel/operator⟶learningtheory.Parts I–II live at the first two arrows: MDL/code length chooses among computable descriptions, while KC/NCD/CKC/AMI provide the empirical algorithmic distances and the KC/D2KE kernel recovered in Proposition 8. Parts III–VI live after kernelization. Table 12 gives the recovery map before the details. Part III: D2KE and the CSZ-style trichotomy. Part III begins with algorithmic distances and asks how to obtain a PSD kernel with a spectrum that falls into one of the complexity regimes. In the present notation, the route isS^n=(Xn,d^Sol,μ^n)⟶M^⟶TD2KE,M^.The Part III regimes are precisely the regimes of TD2KE,M^. Polynomial spectra correspond to geometries for which metric entropy or program-counting function grows polynomially across resolution; plateau spectra correspond to finite length shells or finite effective ranks; and supersmooth spectra correspond to analytic/diffusive geometries for which the kernelization kills high modes exponentially. Thus, D2KE is not merely a repair of non-PSD distances; it is the functor that turns algorithmic mm-geometry into the Hilbertian object on which the CSZ-style trichotomy lives. Part IV: SFM, SKCO, SGP, and SGHS. Part IV is recovered by changing the measure in the same master kernel. Uniform landmark mass gives the KC/D2KE Hilbert space, while Occam mass dμ(ω)∝2−ℓ(ω) giveskSFM(z,z′)=∫ϕω(z)ϕω(z′)dμSol(ω), that is, the probability-normalized uncentered Solomonoff feature mixture and its integral operator. Under orthogonality or the near-orthogonality/Riesz hypotheses of Theorem 2, the eigenvalues of this operator track 2−ℓi. This is a second-moment construction from Occam-weighted features; identifying it with the posterior-centered covariance projection requires zero feature mean and matching weights. Part V: learning rates and spectral collapse. Part V studies what happens to learning after the projection to TM. In the mm-space language, the central quantity is the effective dimensionNM(λ)=trTM(TM+λI)−1=∑jλj(M)λj(M)+λ.Redundancy inflation is the statement that the chosen geometry/feature dictionary increases NM(λ) without adding predictive structure; degeneracy concentration may instead decrease it. The plateau results of Part V remain exactly the finite-capacity case NM(λ)≈m below the shelf. The global Solomonoff envelope and the Part V plateau regime are compatible, as the former calibrates mass across length classes and the latter controls finite-resolution learning inside an effective shell. Part VI: deep KC and pullback geometries. Part VI is recovered by allowing the candidate geometry itself to be a learned pullback. A representation Φθ:X→Zθ inducesdθ(x,x′)=dZ(Φθ(x),Φθ(x′)),Mθ=(Φθ(X),dZ,Φθ#μ^n).Solomonoff–Gromov selection chooses the pullback geometry with relational distances matching d^Sol, while the marked objective chooses the simplest pullback that is also predictive. Applying D2KE or SFM after this selection recovers the deep KC/Solomonoff kernels of Part VI. 17.3. A Learning Theory Directly on mm-Spaces The preceding discussion suggests a direct extension of Parts III–V. Let G be a class of candidate mm-spaces, let κ be a kernelization functor, and let HM be the RKHS induced by κM. For observations (xi,yi), defineM^∈arg minM∈GdGW,pp(S^n,M)+λCode(M)+ηTaskLoss(M), then learn f∈HM^. The complexity is no longer assigned to a kernel alone; it is now assigned to the geometry plus its kernelization. Under squared loss and Gaussian observation noise, Proposition 10 identifies this downstream KRR estimator with the posterior mean of the SGP having covariance κM^. Thus the learning bounds below apply to the same finite predictor under its deterministic SGHS and probabilistic SGP readings. For a kernelized mm-space M, define NM(λ)=∑jλj(M)λj(M)+λ. Define the mm spectral dimension by βM=inf{β≥0:NM(λ)=O(λ−β)asλ↓0}. The infimum need not be attained. Logarithmic/supersmooth growth has dimension zero but can have unbounded effective dimension; its logarithmic factors must remain in rate calculations. If λj≍j−a, a>1, then βM=1/a. Capacity is now two-layered. The inner capacity is the effective dimension NM(λ) of the selected geometry; this is the usual CSZ/Caponnetto–De Vito quantity after kernelization [83,84]. The outer capacity is the complexity of the geometry class G itself, measured either by a Kraft code Code(M) in a countable class or by metric entropy of (G,dGW) in a continuous class. The MDL term in the Solomonoff–Gromov objective is not cosmetic, it is the outer-capacity regularizer. Let A1,…,AN be a finite list of complete computable mm-model candidates. Candidate Aj includes its mm-space, a fixed GW distortion Dj≥0, its fitted predictor fj, and a prefix code length Lj satisfying ∑j2−Lj≤1. Assume the candidates, codes, and predictors are fixed independently of an iid validation sample V of size nval, drawn from the distribution defining Rj. For a bounded loss ℓ∈[0,B], write Rj=E[ℓ(fj(X),Y)],R^V,j=1nval∑(x,y)∈Vℓ(fj(x),y), and define qj(δ)=BLjln2+log(2/δ)2nval. Let j^∈arg min1≤j≤NR^V,j+γDj+λLj+qj(δ). Then, with probability at least 1−δ, Rj^+γDj^+λLj^≤inf1≤j≤NRj+γDj+λLj+2qj(δ). The same statement holds for a countable coded class by replacing the finite list with any prefix-code family and interpreting the infimum over candidates with finite code length; since the penalized argmin need not be attained over an infinite list, j^ is then taken to be any ϵ-minimizer and ϵ is added to the right-hand side. For sub-Gaussian losses, replace B by the corresponding sub-Gaussian scale up to the usual absolute constant. Hoeffding’s inequality [85] with failure probability δj=δ2−Lj gives|R^V,j−Rj|≤qj(δ) for all candidates simultaneously, since ∑jδj≤δ. On this event,Rj^+γDj^+λLj^≤R^V,j^+qj^+γDj^+λLj^.The definition of j^ bounds the right-hand side by the same validation objective for any j, and the reverse concentration inequality converts R^V,j back to Rj, producing the displayed oracle inequality. □ This theorem is intentionally modest. It does not prove that a compressor has found the true Solomonoff posterior geometry, nor does it give a uniform theorem over arbitrary continuous geometry classes. Once a coded mm-model list and an independent validation protocol are fixed, it supplies the standard finite-sample Occam bound. The distortion Dj is only a non-negative score fixed before validation; all metric–measure content lies in how that score and the induced predictor are constructed. The conditional results below identify the additional transport, stability, and concentration assumptions needed for a geometry-level oracle statement. Let M★ be an oracle population geometry and let M^=Mθ^ be selected using data independent of the n-sample subsequently used for regression (equivalently, condition on a selected geometry obtained from an independent split). Suppose the kernels are bounded by 1, the regression noise is sub-Gaussian with scale σ, and the oracle covariance is a positive contraction. Assume the source condition f★=TM★sg,01 the bias saturates at the s=1 rate unless a higher-qualification regularizer is used. To compare operators on different L2 spaces without assuming equal Hilbert-space dimension, assume that the transport construction used in this statement supplies an isometric embedding Uπ:L2(μ^)⟶L2(μ★),Uπ*Uπ=I, and write Pπ=UπUπ* for the projection onto its range. Assume that the population problem is supported on that transported range, Pπf★=f★,TM★=PπTM★Pπ, and define TM^π:=UπTM^Uπ*. A measure-preserving bijection gives the special surjective (unitary) case. A general soft GW coupling does not canonically supply even this isometric transport; using one is an additional construction, and is not inferred from small GW alone. If the two range-support conditions fail, then the same argument controls the projected population problem and the unprojected risk acquires an additional projection/approximation residual. Assume that the kernelization is operator-Lipschitz for this transport: ∥TM^π−TM★∥op≤LκdGW,2(M^,M★). Also, impose the following additional conditional KRR sample-error estimate around the population ridge solution, E∥f^λ−Aλ(TM^π)f★∥L2(μ★)2|M^≤Csamσ2NM^(λ)n,Aλ(T)=T(T+λI)−1, as an explicit hypothesis on the regression experiment. It does not follow merely from bounded kernels, finite rank, or independent splitting: random-design fluctuations can remain even with noiseless labels. For the trace-stability conclusion below, assume the ridge-scale trace stability trTM^π(TM^π+λI)−1−trTM★(TM★+λI)−1≤LN,λdGW,2(M^,M★). This trace/effective-dimension stability is a real assumption: operator-norm closeness alone does not prevent many small eigenvalues from inflating the variance. Then, for squared-loss prediction risk and expectation over the independent regression sample and its noise, EE(f^λ)−E(f★)∣M^≤CR2λ2s+σ2NM★(λ)n+σ2LN,λndGW,2(M^,M★)+R2Lκ2λ2dGW,22(M^,M★), Alternatively, if only rankTM^≤r is assumed in place of trace stability, the conclusion is E[E(f^λ)−E(f★)∣M^]≤CR2λ2s+σ2rn+R2Lκ2λ2dGW,22(M^,M★). Both branches retain the explicit sample-error hypothesis. Here C depends only on the constant in the stated sample-error estimate and not on n, λ, or the selected geometry. Thus, the geometry error must be optimized jointly with the ridge parameter. By Equation (15), one may substitute δgeom:=δtrunc(L)+δcomp(C,L)+δsamp(n)+δapp(G)+δopt for the GW discrepancy only if these terms bound a chain between the selected and oracle geometries; here δopt must bound selection displacement, not merely objective excess. If the geometry is random and an unconditional expectation is desired, average its discrepancy and squared discrepancy separately. A bound on Eδgeom does not imply a bound on Eδgeom2 by its square. Under the p=2 box-dimension hypothesis of Theorem 4, Eδsamp(n)≲n−1/(2s′) for every s′>max(dbox,2), with a space-dependent constant. If NM★(λ)≲λ−β, the fixed-geometry choice λn≍n−1/(2s+β) gives n−2s/(2s+β) provided that both geometry contributions are of that order; in particular, δgeomλn2=On−2s/(2s+β). The trace contribution additionally requires LN,λnδgeom/n to be no larger than the target rate. Otherwise, λ must be chosen by optimizing the full displayed bound. The empirical-GW rate in Theorem 4 is an upper bound, not a matching lower bound. It can verify the displayed requirement when it is sufficiently fast, but a slower upper bound cannot be reversed to prove that the fixed-geometry rate is impossible; that conclusion would require a lower bound or a sharper geometry estimate. For 00. This bounds the effective-dimension factor conditional on a valid sample-error estimate. It does not establish that estimate. For example, let X be uniform on {0,1}, k(x,x′)=xx′, f★(x)=x, n=1, λ=1, and noise variance zero. The population ridge solution is x/3, while the empirical solution is either zero or x/2; their expected squared L2 difference is 5/144>0. Thus finite rank alone cannot justify a variance-only sample bound. Let Gr be a class of landmark/D2KE kernelized metric–measure spaces with kernels that all factor through a common r-dimensional feature space with ∥ϕM(x)∥≤1. Let M★∈Gr be an oracle geometry satisfying the source condition in Theorem 6, with sub-Gaussian noise of scale σ. Assume that the D2KE kernelization is operator-Lipschitz along GW couplings on this class, ∥TM^π−TM★∥op≤LκdGW,2(M^,M★), a condition supported at the finite kernel level by Proposition 13. Additionally, assume the independent split, isometric transport, range support, and conditional sample error conditions of Theorem 6 and suppose that dGW,2(M^,M★)≤ε. Then, kernel ridge regression with the induced kernel κM^ and ridge parameter λ satisfies EE(f^λ)−E(f★)∣M^≤CR2λ2s+σ2rn+R2Lκ2ε2λ2. Thus, on this finite-rank D2KE class, GW distortion enters the prediction guarantee through the induced kernel operator, not merely as an additive validation score, and the bound reduces to the usual finite-rank KRR rate as ε→0 when λ is chosen in the fixed-geometry regime. The finite-rank assumption gives NM^(λ),NM★(λ)≤r, so no separate trace-stability term is needed. The operator-Lipschitz assumption and dGW,2(M^,M★)≤ε give ∥TM^π−TM★∥op≤Lκε. Substitution in Theorem 6 gives the result. □ This theorem is the mm-space analogue of the CSZ/Parts III–V learning theory. The three regimes give the familiar consequences:polynomial:N(λ)≍λ−1/a⇒n−2s/(2s+1/a),plateau:N(λ)≈m⇒parametricm/nvariancebelowthecutoff,super-smooth:N(λ)≲log(1/λ)q⇒near-parametricratesuptologarithms.What this paper adds is the selection layer: the learner may choose the geometry for which the induced NM, code length, and task loss best match the empirical Solomonoff structure. 17.4. How the Unification and Oracle Results Fit Together Theorem 1 and the error/oracle results answer different questions. The unification theorem begins with a fixed distance, landmark measure, bandwidth, and D2KE feature family. It shows how empirical/uniform and Occam landmark measures recover the two earlier kernel constructions; its spectral clause retains the orthogonality or perturbative hypotheses of Theorem 2. It is selection-free once those ingredients are fixed, not a claim that D2KE is the unique category-independent kernelization. The error decomposition and the finite and countable oracle statements instead concern how a geometry or induced predictor is chosen from data. They do not make the chosen representation canonical. Their conclusions require additional bridges: a declared candidate class and code, independent validation or another generalization device, a marked or predictive selection criterion for supervised tasks, and stability of the induced kernel learner. These assumptions are not consequences of Theorem 1. The two layers compose only in the forward directionselectedgeometry⟶fixedD2KE/SFMrepresentation⟶predictorandconditionalriskcontrolThe unification theorem supplies the middle map; the selection theorems control the first and last arrows under their separate hypotheses. Neither layer implies the other. The cellular-automaton and SMS prediction studies (E10–E11) expose a further prerequisite that an oracle inequality cannot supply: competitiveness is only relative to the included candidate family. Even if every quotient in that family destroys task-relevant structure, an internally valid oracle comparison can still select a predictively poor model. 17.5. Oracle Inequalities over Geometries The preceding theorem is the fixed-oracle slice of the mm-space theory. The full selection statement has one more term: the price of searching over geometries. For countable geometry classes, this term is controlled by the same Kraft/MDL mechanism as the mixture dominance bound of Proposition 15; for continuous classes, it requires uniform predictive generalization control for the induced data-dependent learners. The assumptions form the explicit ladder in Table 13. The final oracle theorem is not a consequence of the definition of Solomonoff–Gromov selection alone. Each row adds a separate bridge; if that bridge is unavailable, the valid conclusion stops at the preceding row. Let G be a countable class of candidate geometries and let Λ be a fixed singleton or a countable regularization grid. For a=(M,λ), assign nonnegative code lengthsCode(a)=Code(M)+CodeΛ(λ),∑a2−Code(a)≤1,with CodeΛ=0 for a fixed singleton. Condition on construction data that fix the candidates and their codes. Fit each predictor f^a on regression data, independently of an iid validation sample of size nval. All predictors must be evaluated on the same observed input space through declared measurable maps. Suppose the validation loss lies in [0,B]. Define qa=BCode(a)ln2+log(4/δ)2nval,0<δ<1.Let Ga≥0 be a penalty fixed independently of validation, and let a^ be an εopt-minimizer of E^V(f^a)+qa+Ga.Then, with probability at least 1−δ/2, E(f^a^)≤infa{E(f^a)+2qa+Ga}+εopt.If, additionally, candidate bounds satisfy simultaneously with probability at least 1−δ/2 E(f^a)−E(f★)≤Ha,then with probability at least 1−δ, E(f^a^)−E(f★)≤infa{Ha+2qa+Ga}+εopt.For the KRR interpretation, a permissible additional candidate-bound assumption isHa=CRM2λ2sM+σ2NM(λ)n+Gageom+qaKRR,where ΔM is a valid bound on the relevant transported geometry discrepancy and Gageom=σ2LN,λnΔM+RM2Lκ2λ2ΔM2.This high-probability assumption must be proved for the specified regression experiment, including the transport, source, sample-error, and trace hypotheses; it does not follow from the expected-risk template alone. In its finite-rank alternative, replace the effective-dimension and trace terms by σ2r/n. Choosing Ga as a computable upper bound on Gageom gives a risk-scaled geometry penalty. Any extra MDL penalty included in the selection objective remains in Ga and hence in the comparator bound. For squared loss, bounded validation loss requires bounded responses and suitably bounded or clipped predictors; a uniform sub-Gaussian loss assumption permits the corresponding concentration radius instead. Condition on all construction and regression data. For each candidate, Hoeffding’s inequality [85] givesPr{|E^V(f^a)−E(f^a)|>qa}≤(δ/2)2−Code(a).Summing over the coded class bounds the failure probability by δ/2. On the complementary event, approximate minimization and Ga^≥0 give, for every a,E(f^a^)≤E^V(f^a^)+qa^+Ga^≤E^V(f^a)+qa+Ga+εopt≤E(f^a)+2qa+Ga+εopt.Taking the infimum proves the first assertion. Intersect with the assumed uniform candidate-risk event and substitute Ha to prove the second. A code-weighted union bound on separately proved candidate tails is one way to obtain that event. This is a validation-selector result; the sequential mixture redundancy bound in Proposition 15 concerns a different predictor and loss guarantee. Related concentration and coding approaches include U-statistics, bounded differences, and PAC-Bayes [86,87,88]. □ The countability and validation split assumptions are not cosmetic. The theorem does not assert a uniform oracle inequality over arbitrary continuous geometry classes. A continuous parameter family requires either discretization by a coded net, a PAC-Bayes prior over parameters, or a uniform predictive generalization theorem. Establishing such a theorem for broad continuous geometry classes is part of the open statistical theory of Solomonoff–Gromov approximation. The theorem deliberately assumes a marked or predictive selection objective. Pure unmarked Solomonoff–Gromov selection can certify that the prior geometry or covariance geometry is relationally close to the empirical Solomonoff space, but cannot by itself identify which labels or function values attach to which points. This is the same loss recorded in Proposition 17: GW is isometry-invariant and does not identify external task marks. Predictive oracle inequalities require either marked GW, an empirical risk term, or, for a mixture predictor rather than a selected KRR model, the separate sequential log-loss bound of Proposition 15. For every class G, even an uncountable one, the reverse triangle inequality gives supM∈G|dGW,p(S^,M)−dGW,p(S,M)|≤dGW,p(S^,S). For powered distances uniformly bounded by R, multiply the right side by pRp−1. Thus candidate-class entropy is not needed for this geometric comparison. The unresolved outer-capacity problem is uniform predictive-loss control when continuous geometry selection also changes the learned function class and its task alignment. Entropy or Rademacher complexity can enter that distinct generalization problem. 18. Ultrametric Solomonoff Geometry 18.1. Prefix Trees The natural geometry of prefix-free programs is a tree, and trees correspond to ultrametric spaces in a precise sense [89]; ultrametric GW and Wasserstein geometries provide useful comparison points [90,91]. Let Ω⊂{0,1}* be a prefix-free code set. Define the prefix distancedpref(p,q)=0,p=q,2−|p∧q|,p≠q, where p∧q is the longest common prefix of p and q. This is an ultrametric:dpref(p,r)≤max{dpref(p,q),dpref(q,r)}.For a nonempty code set, normalize the weights by W=∑p∈Ω2−|p|∈(0,1]. The metric completion, carrying the push-forward of these probability masses, is an ultrametric mm-space. Completion matters for infinite code sets, whose Cauchy sequences can converge to infinite strings. The same tree carries a native positive semidefinite prefix kernel. For strings x,y, set(16)kSol,pref(x,y)=∑u⪯x,y2−|u|, where u⪯x,y means that u is a common prefix of both strings. This is the inner product of shared-prefix indicator features weighted by Occam mass. More generally, one may use 2−K(u), whose total sum is finite by choosing shortest prefix descriptions. For M(u), either restrict depth or insert a summable depth factor, for example 2−|u|M(u), to ensure finite diagonal sums on infinite strings. Thus, the ultrametric presentation has its own Solomonoff kernel before any Euclidean or Gaussian projection is introduced. 18.2. Why Ultrametrics Matter Many algorithmic data sets are hierarchical: files have reusable subroutines, strings share prefixes, programs share modules, grammars share production trees. While a Euclidean geometry often hides this hierarchy, an ultrametric geometry exposes it. This preference is sharper than interpretability. Among finite metrics the Schoenberg gate of Definition 7—d2 conditionally negative definite—is exactly what decides L2-embeddability, and ultrametrics always pass it (every separable ultrametric space embeds isometrically into ℓ2 [92]; see also [93]), whereas generic graph and even tree metrics do not. The Hamming-distance star dH on {000,001,010,100} is already non-Euclidean; it is the very witness in [76] (Thm. 33). Both ultrametric and Euclidean candidates admit isometric Hilbert embeddings, whereas general graph shortest-path metrics need not. A separate kernelization can still change the original distances. The Solomonoff–Gromov framework makes this testable. Fit both a Euclidean candidate and an ultrametric tree candidate to S^n. If the ultrametric candidate has much smaller GW + MDL score, then the data are better described by hierarchical program structure than by smooth Euclidean variation. 18.3. Tree-Aligned Couplings For a truncated prefix tree, an optimal or near-optimal GW coupling has a natural interpretation: it aligns clusters of data points with subtrees. This gives a geometric form of algorithmic clustering, in the spirit of the ultrametric characterization of hierarchical clustering [94]. Instead of clustering by Euclidean proximity or kernel similarity, one clusters by common explanatory prefixes. An exact isometry between finite ultrametric probability spaces induces a zero-distortion GW coupling. Its named branches are identifiable only modulo measure-preserving tree automorphisms. Correct depth alone is insufficient: repeated subtree shapes or equal masses can produce indistinguishable alignments. Recovering specific prefix clusters from perturbed data requires a separation condition and a task-relevant correspondence in addition to compressor and sampling control. Quantitative recovery under such hypotheses remains open. 19. Algorithms 19.1. Basic Finite Algorithm Algorithm 1 states the basic finite selection procedure. Algorithm 1 Finite Solomonoff–Gromov model selectionInput: data x1,…,xn, and labels if a marked/predictive objective is used; compressor or algorithmic distance estimator; candidate geometry class G; penalties λ,η. 1.For predictive selection, reserve an iid validation split Ival independent of all construction and fitting data Itr. To invoke the conditional regression template as well, further split Itr into independent geometry-construction and regression samples. Fix candidate codes and any regularization grid without validation data. Pure geometric selection uses its separately declared GW + MDL objective. 2.Compute the empirical algorithmic distance matrix on the construction set,D^ijSol=d^Sol(xi,xj),i,j∈Itr. 3.Repair or symmetrize the matrix if necessary. 4.Choose an empirical measure μ^tr, uniform or Occam-weighted, on the construction set. 5.For each candidate θ, compute or sample its distance matrix Dθ and mass vector μθ. If marked GW is used, declare and fit its candidate-side mark map mθ using Itr only; the marks must exist before the coupling is optimized. In supervised applications, include an uncoarsened source-geometry representation or another declared safe candidate whenever the quotient family has not been shown adequate; the real-data SMS study (E11) shows that adding marks cannot by itself repair a menu that has already discarded the predictive structure. 6.Solve the finite GW or marked GW problem between (D^Sol,μ^tr) and (Dθ,μθ), using the already declared mθ in the marked case. A joint or alternating mark–coupling fit is a different procedure and needs its own training-only objective and convergence analysis. 7.Construct each induced kernel using D2KE, heat kernelization, or SFM, and fit its predictor on the regression data. If only the finite validation comparison is sought, construction and regression may share Itr; invoking Theorem 6 requires their independence. 8.For Theorem 7, score each fitted pair a by E^V(f^a)+qa+Ga, with its stated confidence radius and a nonnegative penalty fixed before validation. A geometry-to-risk conclusion requires the stated candidate-risk bounds and a valid risk-scaled geometry penalty. Retain any extra code or raw-GW score in Ga and in the comparator bound. Pure geometric selection instead uses its declared GW + MDL score. Compute validation-to-training-landmark distances only for evaluating the fitted predictor; validation labels do not enter construction or fitting. 9.Select θ^ minimizing the resulting objective.10.Use the resulting RKHS/GP/SGHS for regression, classification, clustering, density estimation, or posterior inference. Input: data x1,…,xn, and labels if a marked/predictive objective is used; compressor or algorithmic distance estimator; candidate geometry class G; penalties λ,η. Input: data x1,…,xn, and labels if a marked/predictive objective is used; compressor or algorithmic distance estimator; candidate geometry class G; penalties λ,η. 1.For predictive selection, reserve an iid validation split Ival independent of all construction and fitting data Itr. To invoke the conditional regression template as well, further split Itr into independent geometry-construction and regression samples. Fix candidate codes and any regularization grid without validation data. Pure geometric selection uses its separately declared GW + MDL objective. For predictive selection, reserve an iid validation split Ival independent of all construction and fitting data Itr. To invoke the conditional regression template as well, further split Itr into independent geometry-construction and regression samples. Fix candidate codes and any regularization grid without validation data. Pure geometric selection uses its separately declared GW + MDL objective. 2.Compute the empirical algorithmic distance matrix on the construction set,D^ijSol=d^Sol(xi,xj),i,j∈Itr. Compute the empirical algorithmic distance matrix on the construction set,D^ijSol=d^Sol(xi,xj),i,j∈Itr. 3.Repair or symmetrize the matrix if necessary. Repair or symmetrize the matrix if necessary. 4.Choose an empirical measure μ^tr, uniform or Occam-weighted, on the construction set. Choose an empirical measure μ^tr, uniform or Occam-weighted, on the construction set. 5.For each candidate θ, compute or sample its distance matrix Dθ and mass vector μθ. If marked GW is used, declare and fit its candidate-side mark map mθ using Itr only; the marks must exist before the coupling is optimized. In supervised applications, include an uncoarsened source-geometry representation or another declared safe candidate whenever the quotient family has not been shown adequate; the real-data SMS study (E11) shows that adding marks cannot by itself repair a menu that has already discarded the predictive structure. For each candidate θ, compute or sample its distance matrix Dθ and mass vector μθ. If marked GW is used, declare and fit its candidate-side mark map mθ using Itr only; the marks must exist before the coupling is optimized. In supervised applications, include an uncoarsened source-geometry representation or another declared safe candidate whenever the quotient family has not been shown adequate; the real-data SMS study (E11) shows that adding marks cannot by itself repair a menu that has already discarded the predictive structure. 6.Solve the finite GW or marked GW problem between (D^Sol,μ^tr) and (Dθ,μθ), using the already declared mθ in the marked case. A joint or alternating mark–coupling fit is a different procedure and needs its own training-only objective and convergence analysis. Solve the finite GW or marked GW problem between (D^Sol,μ^tr) and (Dθ,μθ), using the already declared mθ in the marked case. A joint or alternating mark–coupling fit is a different procedure and needs its own training-only objective and convergence analysis. 7.Construct each induced kernel using D2KE, heat kernelization, or SFM, and fit its predictor on the regression data. If only the finite validation comparison is sought, construction and regression may share Itr; invoking Theorem 6 requires their independence. Construct each induced kernel using D2KE, heat kernelization, or SFM, and fit its predictor on the regression data. If only the finite validation comparison is sought, construction and regression may share Itr; invoking Theorem 6 requires their independence. 8.For Theorem 7, score each fitted pair a by E^V(f^a)+qa+Ga, with its stated confidence radius and a nonnegative penalty fixed before validation. A geometry-to-risk conclusion requires the stated candidate-risk bounds and a valid risk-scaled geometry penalty. Retain any extra code or raw-GW score in Ga and in the comparator bound. Pure geometric selection instead uses its declared GW + MDL score. Compute validation-to-training-landmark distances only for evaluating the fitted predictor; validation labels do not enter construction or fitting. For Theorem 7, score each fitted pair a by E^V(f^a)+qa+Ga, with its stated confidence radius and a nonnegative penalty fixed before validation. A geometry-to-risk conclusion requires the stated candidate-risk bounds and a valid risk-scaled geometry penalty. Retain any extra code or raw-GW score in Ga and in the comparator bound. Pure geometric selection instead uses its declared GW + MDL score. Compute validation-to-training-landmark distances only for evaluating the fitted predictor; validation labels do not enter construction or fitting. 9.Select θ^ minimizing the resulting objective. Select θ^ minimizing the resulting objective. 10.Use the resulting RKHS/GP/SGHS for regression, classification, clustering, density estimation, or posterior inference. Use the resulting RKHS/GP/SGHS for regression, classification, clustering, density estimation, or posterior inference. Cost scale. For a finite empirical space with n data points and a candidate with m support points, storing dense distance matrices costs O(n2+m2) memory and the coupling costs O(nm). A naive GW objective evaluation uses O(n2m2) pair interactions. Entropic/conditional gradient implementations exploit matrix products and typically scale per major iteration, such as O(n2m+nm2) plus O(nm) coupling updates, with constants depending on the solver. Landmark approximations with r landmarks reduce distance construction/storage toward O((n+m)r) plus the cost of the reduced GW or core-set problem. These are implementation scalings, not new statistical guarantees. 19.2. Entropic Approximation Exact GW is often computationally hard [95]. Entropic regularization has an information-theoretic reading worth stating here: up to an additive constant, the regularizer is the Kullback–Leibler divergence of the coupling from the product of its marginals, i.e., the mutual information of the alignment, so entropic GW is distortion minimization under a mutual-information budget, a rate–distortion form of the alignment problem [32,33,81,96]. It replaces the finite problem byminΠ∈U(α,β)12∑i,i′,a,a′|Dii′X−Daa′Y|pΠiaΠi′a′+ε∑i,aΠia(logΠia−1).For fixed marginals, the displayed entropy term equals KL(Π∥α⊗β)−H(α)−H(β)−1, so it has the same minimizers as the centered mutual-information penalty, while its raw objective value differs by a known constant. Here, KL is measured in nats because natural logarithms are used, so ε has the same units as C(Π), namely, the chosen distance raised to the p-th power. Let α∈Δn and β∈Δm, let C(Π) denote the unregularized finite GW distortion including the 1/2 convention of Definition 2, and define C0=minΠ∈U(α,β)C(Π),Cε=minΠ∈U(α,β)C(Π)+εKL(Π∥α⊗β). With the usual conventions on zero-mass entries, 0≤Cε−C0≤εmin{H(α),H(β)}≤εlogmin{n,m}. Consequently, if d0=C01/p and dε=Cε1/p, then 0≤dε−d0≤εmin{H(α),H(β)}1/p. Non-negativity of relative entropy gives the lower bound. If Π0 minimizes C, thenCε≤C(Π0)+εKL(Π0∥α⊗β).The relative entropy is the mutual information of a pair with marginals α,β, and hence is at most min{H(α),H(β)}, which is at most logmin{n,m}. The final display follows from (a+b)1/p≤a1/p+b1/p for p≥1. □ The proposition bounds the regularization bias of the centered finite objective; it does not certify convergence of a particular non-convex numerical run. Sinkhorn-like or inexact gradient algorithms and their convergence conditions are studied in [96]. Accordingly, the error budget keeps the numerical solver gap in δopt and the regularization bias in δent, rather than identifying an entropic value with exact GW. The bound also gives a transparent tuning rule. Writing Hmin=min{H(α),H(β)}, a desired tolerance ρ on the powered objective is guaranteed by ε≤ρ/Hmin, while a tolerance r on the unpowered GW distance is guaranteed by ε≤rp/Hmin (with zero bias when Hmin=0). In practice, continuation from larger to smaller ε, followed on small instances by an unregularized multi-start check, separates numerical convenience from the regularization bias budget. 19.3. Landmarks and Scalable D2KE For large n, compute distances to landmarks rather than all pairwise distances. Let L={ω1,…,ωm} be landmark programs or representative data points. DefineΦ(x)=μ1e−γd^Sol(x,ω1),…,μme−γd^Sol(x,ωm).Here μi≥0, ∑iμi=1, and uniform landmarks use μi=1/m. Inner products give the D2KE kernel for this landmark probability measure; approximation to another reference measure requires a quadrature or sampling argument. For GW itself, the feature map is used in one of two concrete ways: either to form the low-rank surrogate distanceD^ij(L)=∥Φ(xi)−Φ(xj)∥2 and pass this smaller/structured distance representation to a GW solver, or to compute a core-set/landmark coupling first and extend it to the full sample. Thus, the landmark step is not a replacement of GW by ordinary MMD; it is a low-rank or core-set approximation to the distance matrices entering the GW problem. The resulting low-rank error is part of δopt or the empirical distance approximation term, depending on how the landmarks are chosen. The Euclidean feature distance branch is an early Hilbertization, making it a potentially lossy accelerator, not an exact preservation of the category-native dissimilarity. Its distortion should be audited, for example by δL=∥D^(L)−D^Sol∥∞, which feeds directly into the distance perturbation stability bound of Proposition 6. When preservation of non-Hilbert relational structure is essential, a core-set that retains the original dissimilarities is the safer approximation branch. 19.4. Barycenters Given several datasets or tasks, one can define their empirical Solomonoff spacesS^n(1),…,S^n(T).A Solomonoff–Gromov barycenter isS¯∈arg minM∑t=1TwtdGW,pp(M,S^n(t))+λCode(M).This barycenter is a shared algorithmic geometry across tasks, and could serve as a meta-learned prior. 20. Empirical Certification A kernel with decaying eigenvalues is not automatically Solomonoff-like; many ordinary priors penalize complexity. The Solomonoff–Gromov framework adds an empirical certification criterion. 20.1. Relational Solomonoff Preservation A candidate geometry Mθ preserves empirical Solomonoff structure at tolerance ε if dGW,p(S^n,Mθ)≤ε. A derived kernel kθ preserves empirical Solomonoff structure if its underlying geometry or kernel-induced metric satisfies the same condition. This test is stronger than spectral decay alone; it requires the pairwise algorithmic relationships among the data to be reproducible by the model geometry. 20.2. Certification Protocol Four requirements make the protocol falsifiable rather than self-certifying (a relational-preservation check on the same data and compressor is selection-biased, and does not ensure passage of an independently fixed tolerance): (a) held-out relations—the preservation inequality must be evaluated on relations not used for fitting or selection (the inductive held-out protocol, E3); (b) disjoint reference—the reference dissimilarity must come from a compressor (or ensemble member) not used in construction; (c) null calibration—the tolerance ε must be calibrated against the structureless-data negative control (E2), since a pass threshold that structureless data also pass certifies nothing; and (d) explicit pass/fail—each item below must state its threshold before the data are seen. A candidate Solomonoff-type model should be evaluated by the following tests:Semimeasure legitimacy: Weights should derive from prefix-free or approximately prefix-free code lengths.Aggregation: The model should aggregate mass over multiple explanations, not only over the shortest description.Algorithmic information capture: Pairs with high algorithmic mutual information should be close even when ordinary statistical correlation is low.Spectral calibration: Induced operator spectra should match 2−ℓi or a depth-truncated variant.Relational GW preservation: The learned geometry should have small GW distance to the empirical Solomonoff mm-space.Cross-compressor stability: The selected geometry should be stable across reasonable compressors or complexity surrogates. Semimeasure legitimacy: Weights should derive from prefix-free or approximately prefix-free code lengths. Semimeasure legitimacy: Weights should derive from prefix-free or approximately prefix-free code lengths. Aggregation: The model should aggregate mass over multiple explanations, not only over the shortest description. Aggregation: The model should aggregate mass over multiple explanations, not only over the shortest description. Algorithmic information capture: Pairs with high algorithmic mutual information should be close even when ordinary statistical correlation is low. Algorithmic information capture: Pairs with high algorithmic mutual information should be close even when ordinary statistical correlation is low. Spectral calibration: Induced operator spectra should match 2−ℓi or a depth-truncated variant. Spectral calibration: Induced operator spectra should match 2−ℓi or a depth-truncated variant. Relational GW preservation: The learned geometry should have small GW distance to the empirical Solomonoff mm-space. Relational GW preservation: The learned geometry should have small GW distance to the empirical Solomonoff mm-space. Cross-compressor stability: The selected geometry should be stable across reasonable compressors or complexity surrogates. Cross-compressor stability: The selected geometry should be stable across reasonable compressors or complexity surrogates. The fifth item is the new criterion contributed by this paper. 21. Relation to Existing BAITML Components 21.1. Relation to MDL Kernel Learning MDL kernel learning [1,97,98], in the tradition of MDL model selection [99,100,101,102,103], says to choose the kernel that compresses the data. Solomonoff–Gromov model selection says to choose the geometry with a relational structure that explains the algorithmic distances of the data via the shortest description. Thus, the MDL object shifts from a kernel to an mm-space. 21.2. Relation to KC-Kernels and D2KE KC-kernels start from algorithmic similarities or distances. D2KE converts non-PSD distances into PSD kernels. The present paper inserts a geometry-learning layer before D2KE:d^Sol⟶S^n⟶M^⟶kM^.Thus D2KE remains essential, but is applied to a selected explanatory geometry rather than directly to the raw empirical distance matrix. 21.3. Relation to SFM and SGP The Solomonoff feature mixture constructs kernels from weighted features or explanations. In the geometry-first view, those features live on a selected mm-space. A Solomonoff Gaussian process is then a GP for which the covariance is induced by the learned algorithmic geometry. This makes explicit that an SGP is a second-order probabilistic surrogate, not an exact replacement for the discrete Solomonoff prior. For Gaussian regression, Proposition 10 closes the theory-to-prediction chain: the SGP posterior mean is exactly the KRR solution in the induced SGHS and has one kernel section per observation. The covariance weights 2−ℓr therefore survive computationally as the inverse spectral penalties 2ℓr, subject to the explicit computability qualification of Remark 8. 21.4. Relation to the Countability–Gaussianity Mismatch A limitation of Gaussianizing Solomonoff induction is that a non-degenerate Gaussian measure assigns probability zero to any fixed countable set of computable functions. The present paper avoids overstating the relationship. The SGP is not claimed to equal Solomonoff induction as a measure on hypotheses; it is a covariance-level and geometry-level surrogate, and the Solomonoff–Gromov criterion measures how well this surrogate preserves algorithmic relational geometry, not whether it reproduces the exact universal semimeasure. 21.5. Relation to Learning Theory Learning rates depend on the geometry and spectrum induced by the selected model. If GW selection chooses a geometry for which the spectrum is Solomonoff-calibrated, then the resulting learning theory can be interpreted as learning under algorithmic compressibility. If it chooses a polynomial, plateau, or super-smooth geometry, then the mismatch can be diagnosed through the spectral and GW criteria described above. 21.6. Mutual Support with Parts VII–XII The recovery results above make the link to Parts III–VI explicit [3,4,5,6]. The later parts are Part VII (kernel discrepancies and the bidirectional bridge) [104], Part VIII (kernel flows as random dynamical systems, with algorithmic randomness [105]) [106], Part IX (Solomonoff–Barron spaces) [107], Part X (hardness and tractability, building on [108]) [109], Part XI (Rademacher complexity through Kolmogorov complexity) [110], and Part XII (operator-adapted multiresolution, building on gamblets [111]) [112]. The programme survey is in [113]. The later parts of the programme fit the same tower but at different levels: they do not compete with Solomonoff–Gromov selection, but rather supply the diagnostics, dynamics, approximation spaces, hardness limits, capacity bounds, and multiresolution bases needed after an explanatory geometry has been selected. Table 14 summarizes these relationships. This provides a two-way reading. The selected algorithmic mm-geometry supplies an input to the later diagnostics and learning constructions, while Parts VII–XII provide the statistical tests, dynamics, approximation spaces, hardness limits, capacity bounds, and multiresolution numerics needed after that selection step. 22. Examples and Experiments We pair each quantitative example with the experiment that tests it. Every numerical value reported in this section comes from the numerical-validation programme reported in Section 14. All empirical numbers should be read as finite sanity checks unless accompanied by dispersion over seeds and the full protocol. 22.1. Worked Examples and Status Strings generated by grammars. Strings from a small grammar; a Euclidean embedding of character counts may miss the grammar, while an NCD-based empirical Solomonoff space places strings with shared rules close. Result (latent-structure recovery, E1; Section 14): On data from K=3 low-complexity cores, the computed relational GW values in these runs decrease with k and does not identify the truth; the BIC/MDL-penalized objective with a cross-validated λ recovers K=3 on four of five exploratory data seeds and fails on one. The larger 20-plus-40 protocol separately measures transfer after supervised synthetic calibration. Program outputs. xi=U(pi) for short programs; the ideal geometry is the prefix tree of the pi. This is the latent-structure recovery setting (E1) instantiated with program-defined cores; the GW coupling aligns outputs with program structure and D2KE yields a PSD kernel on program-output similarity. Algorithmically related pairs and dependence testing. For y=g(x) with a simple computable transformation, algorithmic similarity offers a representation-free description even when the objects do not share a convenient vector space. This does not imply that characteristic-kernel tests fail: a deterministic relation is detectable by a properly calibrated characteristic-kernel independence test. Any comparison of NCD-based and conventional HSIC must use the same permutation calibration, sample size, and power criterion. Such a powered comparison is not part of the present numerical validation and remains future work. Physical fields with hidden low-complexity laws. For PDE-generated data, the best candidate geometry may be operator-adapted/multiresolution rather than a prefix tree; GW selection would compare a manifold, a graph, and an operator-adapted hierarchy against the empirical Solomonoff space. This case is a motivating direction, not an executed result; the operator-adapted instantiation is future work (Section 23). 22.2. Executed Experimental Programme The following summarizes the completed experiments. Prefix-tree/latent-structure recovery and its failure mode. The latent-structure and inductive held-out studies (E1 and E3; Section 14) show that the BIC/MDL-penalized objective with cross-validated λ recovers the true number of latent components on 4/5 exploratory seeds; the relational term alone never does, and one seed admits no recovering λ. In the separate frozen confirmation protocol, 34/40 structured corpora are recovered after a ground truth-assisted calibration stage. Spectral calibration. The synthetic spectral-calibration study (E4; Section 14) finds that the recovered kernel’s spectrum equals 2−ℓi exactly in the orthogonal control and degrades gracefully under coherence; the weights are inserted by hand on synthetic features, so this validates the mechanism of Theorem 2, not a recovery from compression data. Family-relative cross-compressor stability. The cross-compressor study (E7; Section 14) finds that the top candidate is stable across zlib, bz2, and lzma, but distance-level ensemble instability is large (0.28), the planted-ambiguity detector caught 0/2 planted items, and, because all three compressors are finite-window statistical coders, the agreement certifies family-relative robustness only, not compressor-independent algorithmic structure. Task performance and its limitations. The held-out prediction and real-data marked-selection studies (E10–E11; Section 14) provide the quantitative task-performance evidence. On periodic-versus-noise data, the selected geometry is competitive with raw compression baselines. On the cellular automaton task, unmarked relational selection collapses to chance-level balanced accuracy. On real SMS data, even fused marking does not rescue the inadequate block family; the validation safeguard recovers only by retaining the raw geometry, and the character 3–5-g SVM remains strongest. The benefit of the geometry selection layer over a raw compression kernel is not established on any tested task. Negative controls. The structureless-data, held-out-prediction, and real-data marked-selection studies (E2, E10, and E11; Section 14) show that the frozen latent-structure protocol selects the trivial geometry on 39/40 fresh noise corpora; unmarked cellular-automaton selection reaches chance-level balanced accuracy after one-block collapse; and, on the natural-language task, pure marked block selection also collapses, while the conventional character baseline wins. These controls constrain the interpretation without establishing broad external validity. 23. Discussion, Limitations, and Open Problems 23.1. Computability The true Solomonoff distance is incomputable. The framework is therefore a surrogate framework. Every practical result depends on the chosen compressor, program dictionary, or finite approximation. 23.2. GW Complexity Exact GW optimization is computationally hard in general. Practical use requires entropic approximations, low-rank couplings, slicing, landmarks, or restrictions to special geometry classes such as trees or ultrametrics. 23.3. Metric Repair Compression distances may violate metric axioms. The theory assumes metric inputs; practical algorithms need robust repair or dissimilarity-GW variants. If a raw dissimilarity D is repaired to a metric Drep, then the error decomposition should include a repair distortion term such asδrepair=∥D−Drep∥∞ or the corresponding distortion in the chosen GW cost. One may absorb this into the compressor/sampling term only if the repair procedure and its distortion are reported; otherwise, it is a separate approximation layer. The following bound shows that this layer is controlled rather than uncontrolled. Let D be a symmetric non-negative dissimilarity on Xn with zero diagonal, and let Drep be any repair with ∥D−Drep∥∞≤δrepair. Write S˜=(Xn,D,μ^n) and Srep=(Xn,Drep,μ^n) for the two (dissimilarity-) mm-spaces on the same support and mass. Here, dGW,p denotes the finite GW functional evaluated on the dissimilarity matrices, and metricity of D is not used in the perturbation argument. Then, for every candidate mm-space M, dGW,p(S˜,M)−dGW,p(Srep,M)≤2−1/pδrepair. This is the bound for the unpowered GW distance. If the two distances in the absolute value are at most R, then the powered objective used in Equation (5) obeys dGW,pp(S˜,M)−dGW,pp(Srep,M)≤pRp−12−1/pδrepair. Thus, metric repair is controlled by the sup-norm repair distortion, with distinct constants and units for the unpowered distance and the powered Solomonoff–Gromov objective. The two spaces share support and mass, so any coupling admissible for one is admissible for the other, and the two GW integrands differ only through the source cost, uniformly by at most δrepair. This is Proposition 6 with ε=δrepair; the 12 normalization of Definition 2 supplies the 2−1/p factor, and exchanging the two spaces gives the absolute value. Only perturbation of the cost matrix is used, so the bound holds for the dissimilarity-GW value even when D is not a metric. The powered bound follows from |ap−bp|≤pRp−1|a−b|. □ 23.4. Infinite-Dimensional SGP GW A full theory of GW between infinite-dimensional Gaussian process measures requires care. Finite-dimensional Gaussian and covariance-operator proxies are tractable, but the infinite-dimensional non-entropic GW problem remains delicate. 23.5. Universal Dominance Solomonoff’s universal dominance is a property of the discrete universal semimeasure. A Gaussian or kernelized surrogate does not automatically inherit it. The correct claim is weaker: the surrogate preserves selected relational, spectral, and algorithmic-information features of Solomonoff induction. The mixture of Section 12 does inherit a dominance/regret bound (Proposition 15), but at the level of the candidate geometry class, with the computable code length in place of K, not the unrestricted universal dominance. 23.6. Space-Mixture Truncation The space mixture PSpaceMix is only computable after truncating the mm-model collection A to a finite subset with computable weights and component laws, or supplying an effective tail bound for an infinite sum. Evaluating each weight requires a fixed GW distortion, evidence, and a code length. Thus, practical use requires finite dictionaries, beam search, or another declared enumeration scheme. The unnormalized log-evidence price of truncation is the omitted-mass term in Proposition 16; normalized sequential log-loss includes the initial-mass correction given there; designing enumerations for which that mass is estimable or non-vacuously bounded remains open. 23.7. Scope of the Proved Results We restate here, in one place, the limitations of the results. (i) Incomputability. The deficiency Δ and regret RT (Section 13) reference the uncomputable M; they are the correct targets, but are not evaluable against M itself, only against computable surrogates. The experiments check finite consequences and numerical selection behavior, not theorems in full generality or distance to M. (ii) Spectrum hypothesis. Theorem 2’s near-orthogonality condition (H2δ) is restrictive across length classes with large weight dynamic range. The exact product-spectrum case assumes orthogonality; the global two-sided and per-index bands are finite-dictionary statements; and part 2 also states a local localization for countably infinite dictionaries. The relative quadratic-form comparison in the proof gives an additional trace-class extension when δ<1. The stated conservative per-index band uses normalized features and δκ<1; the relative form bound alternatively gives this calibration for δ<1. Neither bound fixes eigenvectors without spectral gaps. The favorable synthetic spectral-calibration behavior (E4) is empirical, not a consequence of a tight worst-case band. (iii) Compressor circularity. The empirical reference S^n is built from a compressor, so the selection compares against a compressor surrogate; the cross-compressor study (E7, Section 14) is the strongest available hardening; it certifies family-relative robustness only, and its planted-ambiguity control failed, both of which are reported as such. (iv) GW vs. prediction. dGW is isometry-invariant, and hence certifies relational covariance geometry rather than the predictor; predictive validity additionally requires task alignment and an adequate candidate class, as the cellular-automaton and SMS studies (E10–E11) show, and the loss of information is made precise by Proposition 17. The SMS post-hoc oracle audit (E11) is limited to the fixed average-linkage partitions, D2KE kernel, and KRR learner; it is not an impossibility theorem for all supervised quotients. (v) No real-data task win. We provide synthetic recovery with an honest failure rate (latent-structure and inductive studies, E1 and E3), controls (structureless data, compressor variation, and held-out prediction; E2, E7, and E10), and a nested real-data marked-selection study (E11). Pure fused marking still collapses on SMS; the safe selector matches raw D2KE only by retaining the source geometry, and character features remain strongest. A larger pre-registered and adversarially fair real-data evaluation and an estimable surrogate for RT remain open work; neither is claimed here. (vi) Status of the position. Category-native approximation is defended as an organizing framework with explicit commitments and a maturity map. Only the metric–measure branch is developed to the theorem level represented here; this paper does not claim a uniform cross-category oracle theorem. 23.8. Open Problems Prove sharp finite-sample rates for empirical Solomonoff–Gromov convergence under compressor distortion, including explicit metric-repair distortion terms. Prove sharp finite-sample rates for empirical Solomonoff–Gromov convergence under compressor distortion, including explicit metric-repair distortion terms. Extend the empirical-GW consistency theorem from i.i.d. sampling to ergodic/trajectory sampling (Birkhoff-type), the case relevant to the dynamical-systems interface of Parts VIII and X; a worked symbolic-dynamics example (logistic-map trajectory windows across the period-doubling cascade, with the tree candidate expected in the periodic regime and the trivial geometry in the chaotic regime) is the natural first target. Extend the empirical-GW consistency theorem from i.i.d. sampling to ergodic/trajectory sampling (Birkhoff-type), the case relevant to the dynamical-systems interface of Parts VIII and X; a worked symbolic-dynamics example (logistic-map trajectory windows across the period-doubling cascade, with the tree candidate expected in the periodic regime and the trivial geometry in the chaotic regime) is the natural first target. Develop uniform predictive generalization bounds for learners induced by continuous geometry classes, beyond the uniform geometric comparison in Remark 27. Develop uniform predictive generalization bounds for learners induced by continuous geometry classes, beyond the uniform geometric comparison in Remark 27. Characterize when GW minimization over ultrametric candidates recovers the true prefix tree. Characterize when GW minimization over ultrametric candidates recovers the true prefix tree. Develop differentiable GW objectives for parametric program feature families. Develop differentiable GW objectives for parametric program feature families. Establish infinite-dimensional IGW theory for trace-class covariance operators relevant to SGPs. Establish infinite-dimensional IGW theory for trace-class covariance operators relevant to SGPs. Prove lower bounds showing that no computable geometry can achieve zero GW distance to the full Solomonoff mm-space. Prove lower bounds showing that no computable geometry can achieve zero GW distance to the full Solomonoff mm-space. Determine when spectral Solomonoff calibration plus relational GW preservation implies useful learning rate improvements. Determine when spectral Solomonoff calibration plus relational GW preservation implies useful learning rate improvements. Design computable enumeration or search schemes for PSpaceMix in which the omitted mixture mass rN(x1:n) is estimable or bounded non-vacuously. Design computable enumeration or search schemes for PSpaceMix in which the omitted mixture mass rN(x1:n) is estimable or bounded non-vacuously. Design benchmark datasets that distinguish Solomonoff-type priors from generic complexity-penalizing priors. Design benchmark datasets that distinguish Solomonoff-type priors from generic complexity-penalizing priors. Sharpen the finite entropy bound of Proposition 19 into geometry-dependent entropic bias rates and quantify low-rank or sliced-GW approximation errors while separating statistical error, regularization bias, and numerical solver error. Sharpen the finite entropy bound of Proposition 19 into geometry-dependent entropic bias rates and quantify low-rank or sliced-GW approximation errors while separating statistical error, regularization bias, and numerical solver error. Replace the declared k-block BIC score by an explicit code for the fitted partition and quantized block parameters, and determine how the penalty-domination boundary changes under that honest code. Replace the declared k-block BIC score by an explicit code for the fitted partition and quantized block parameters, and determine how the penalty-domination boundary changes under that honest code. Characterize quotient adequacy by giving conditions under which a marked candidate family preserves enough task information for predictive competitiveness, and extend Proposition 2 beyond the k-block menu without mistaking a sufficient collapse certificate for a converse. Characterize quotient adequacy by giving conditions under which a marked candidate family preserves enough task information for predictive competitiveness, and extend Proposition 2 beyond the k-block menu without mistaking a sufficient collapse certificate for a converse. Develop one non-metric–measure branch to theorem level (for example, the topological, Banach/Barron, graph, or operator branch) with category-native stability and oracle guarantees. Develop one non-metric–measure branch to theorem level (for example, the topological, Banach/Barron, graph, or operator branch) with category-native stability and oracle guarantees. The scope table in Table 1, assumption ladder in Table 13, and the predictive-collapse and real-data marked-selection analyses in Table 8, Table 9 and Table 10 collect the claim boundaries that matter for interpreting the technical results. 24. Conclusions This paper argues for category-native Solomonoff approximation as the layer between an incomputable universal predictor and practical machine learning representations. The thesis is that choosing a computable shadow means choosing which invariants to preserve; therefore, metric–measure, Hilbert/Banach, topological, graph, tree, operator, and predictive representations should be compared and used through their native structures rather than being forced into a single universal proxy. Equation (1) makes that position operational by coupling native distortion, description length, and when needed, predictive loss. The metric–measure branch demonstrates what it means to develop one branch of this programme. Compression-derived distances and empirical or Occam masses define an empirical mm-space; candidates are compared by GW distortion and penalized by a declared MDL score, and the selected geometry is kernelized by D2KE. For finite or countable classes, the manuscript proves well-posedness, perturbation and transfer results, validation bounds, recovery of the earlier Kolmogorov/Solomonoff kernels, and a redundancy bound for coded mm-model mixtures. The broader oracle inequality is conditional on marked selection, code-weighted concentration, and learner stability. These assumptions are design obligations, not consequences of the position itself. The experiments make the position more precise. The structural-recovery, negative-control, spectral, stability, metric-repair, compressor, Occam-mass, and marked-coupling studies (E1–E9) check isolated mechanisms, mostly on synthetic data. The held-out cellular-automaton study (E10) shows that unmarked relational selection can collapse to a one-block geometry and reach chance-level balanced accuracy; its penalty-domination certificate is empirically sound, but visibly one-sided. The real-data SMS study (E11) shows that fused marks alone do not repair an inadequate quotient family: even a post hoc oracle over all four implemented block kernels remains at chance, the safe selector recovers only by retaining the raw geometry, and a conventional character baseline remains strongest. The lesson is not that representation should cease to be category-native but that the native object for a supervised problem must be task-aligned and that the candidate family must contain an adequate representation. Geometry is a structural prior, not a substitute for task alignment or family design. The category-native claim is consequently both ambitious and bounded. It is ambitious because it proposes a common research architecture and treats kernels as one projection among several rather than the default language for every approximation, and is bounded because Table 4 distinguishes developed, partial, and constructional branches while not claiming any cross-category oracle theorem. This paper’s intended contribution is the vision together with one technically mature realization and an explicit account of what would confirm, refine, or falsify the remaining branches.