Elimination Geometry

Core elimination and architecture

• Core notation distinguishes the lifted objective H_x(a), eliminated objective J(x), oracle set O(x), native defect D_x(a), and population or uniform architecture obstructions.It also defines the declared architecture class 𝔄, deployed field A, excess defect Δimpl, carrier Z, and decoder g.Source: Notation and Standing Conventions · p. 8
• The framework separates carrier information loss and visible signatures from resource-limited distortion, coordination residuals, downstream task contracts, and model, architecture, generalization, and implementation components.These conventions include the conditional dual oracle signature, Bregman decoder, resource grammar, native distortion Dnat, Bellman values, and typed operational kernels.Source: Notation and Standing Conventions · p. 8

Certified learning and intervention

• Certified learning uses population, fixed-record, and anytime confidence worlds whose coverage events contain the true world at protected records or times.It further defines identified images, typed certificates, certificate margins, certificate truth, alternative parameters, raw-data max–min KL rates, and inverse characteristic times.Source: Notation and Standing Conventions · pp. 8–9

Deployment and recursive quotients

• Deployment conventions require exact action recovery through a deployment carrier and decoder, while recursive carriers preserve declared colors and observation-labeled successor laws across quotient class counts.The joint color combines deployment, certificate, and contextual equivalence.Source: Notation and Standing Conventions · p. 9

• The first argument is the evaluation point, while the second is the tangent or reference point.Source: Orientation of asymmetric divergences · p. 9
• For KL divergence, the first argument is the integration law and the second the reference, so native defects place the trial or deployed object first and the oracle second.When Q≪P, KL(Q∥P) follows this orientation; otherwise it is +∞, and conditioning and pushforward preserve the order.Source: Orientation of asymmetric divergences · p. 9

• A zero infimum does not establish attainment of an exact minimizer or global section.Source: Standing distinctions · p. 9
• Uniform, average, and task-weighted risks are distinct contracts, while local oracle, architecture, generalization, and implementation errors remain separate unless exactly combined.Source: Standing distinctions · p. 9
• Point-valued, set-valued, quotient-valued, projector-valued, and distribution-valued outputs represent different semantic contracts.Source: Standing distinctions · p. 9
• Deterministic obstruction, statistical impossibility, posterior credibility, and computational intractability are different claims and must not be conflated.Source: Standing distinctions · p. 9
• The independent statistical unit is the declared sampling unit, not automatically an individual cell, token, edge, or repeated measurement.Source: Standing distinctions · p. 10

Infima and measurability

• Infima range over nonempty declared classes, while measurability, lower semicontinuity, compactness, coercivity, and closure assumptions for attainment are imposed only when locally needed.For approximation-only results, the framework uses infima and ε-optimal selections.Source: Infima and measurability · p. 10

The Structural Realizability Problem

• Modern learning systems reuse representations, parameterizations, memories, and output interfaces across many local problems.Source: The Structural Realizability Problem · p. 21
• The problem extends beyond asking whether local optima exist.Source: The Structural Realizability Problem · p. 21
• The key requirement is that one declared deployment system realize the local optima simultaneously.Source: The Structural Realizability Problem · p. 21

1.1 Reuse changes the learning problem

• Structural failure occurs when one rule cannot realize fixed local optima within the completed, declared deployment class.This distinguishes architecture obstruction from optimization failure or local-oracle inadequacy.Source: §1.1 · p. 22
• Under constant deployment, the two-point squared-loss example has an irreducible excess loss of 1/4, removable only by allowing dependence on x.Each local optimum remains exact, but the shared constant-output contract prevents simultaneous realization.Source: §1.1 · p. 22

1.2 Three levels: solvability, realizability, and certifiability

• Local solvability, global realizability, and finite-sample certifiability are distinct questions requiring separate assumptions and conclusions.A population obstruction may be statistically unidentifiable, while unresolved certification does not establish its absence.Source: §1.2 · p. 23
• The intervention must match the failure: richer local models address misspecification, better data address finite-sample ambiguity, better solvers address implementation error, and contract changes address structural nonrealizability.Repairs include quotient representations, atlases, additional memory, different sharing patterns, and larger resource budgets.Source: §1.2 · p. 23
• Elimination geometry claims objective-denominated lower bounds, finite-data certificates, and mechanism-specific repairs after fixing the local oracle and native excess objective.It distinguishes this scope from prior work on amortization gaps, shared representations, compressed histories, and sparse autoencoders [O’Neill et al., 2025].Source: §1.2 · p. 23

4 CHAPTER 1. STRUCTURAL REALIZABILITY UNDER SHARED DEPLOYMENT

• Incompatible oracle Jacobian rows yield a quantitative native-loss floor for a declared one-pass class and a sufficient proximal repair depth, with exact priority unresolved.The conflict arises in two strict KKT neighborhoods of one fitted nonnegative sparse problem; the unresolved priority is recorded in Appendix G.Source: §4 · p. 24

1.3 Native risk and architecture obstruction

• Native defect is an exact excess objective, and positive architecture obstruction implies a strictly positive population floor relative to the fixed local oracle family.The interpretation requires separately declaring the local oracle, native defect, representation semantics, deployment class, statistical unit, and implementation error.Source: §1.3 · p. 24

1.4 Mission and falsifiable consequences

• The workflow asks which native loss, valid representations, deployment-forced obstruction, finite-data certification, task observability, and minimal intervention govern structural conclusions.It evaluates whether the intervention improves independent held-out risk.Source: §1.4 · p. 25
• A structural explanation requires a predicted nonzero floor, persistence under contract-preserving capacity increases, mechanism-matched improvement, and gains surviving independent validation and matched controls.A performance plateau alone cannot distinguish architecture obstruction from misspecification, optimization failure, data scarcity, regularization, or metric ceilings.Source: §1.4 · p. 25

• Construct a finite instance space where local optima exist but constant deployment has positive architecture obstruction, then minimally enlarge the deployment class to remove it.Source: Exercises · p. 25
• Show that P(0) = P(2π) while v(2π) = −v(0), and explain how projector representation changes realizability without discarding the eigenspace.Source: Exercises · p. 25
• For a modern learning system, formulate separate questions about local solvability, global realizability, and finite-sample certifiability, then propose a negative control for falsifying a structural explanation.Source: Exercises · p. 25

2.1 The primitive interface

• Certified elimination is defined as a triple (H, J, D) with residuals well defined even when a trial state has infinite lifted cost.Each auxiliary fiber is nonempty, and its lifted objective is proper and bounded below by a finite value; attainment is not required.Source: §2.1 · p. 26
• A distance to a chosen optimizer is not itself a certificate unless a proved exchange inequality connects it to the native defect H_x(a) − J(x).Under permutation invariance, the oracle may be an orbit, so the correct geometry is quotient or unordered-output geometry rather than representative parameter distance.Source: §2.1 · p. 26

2.2 A complete learning contract

• A certified learning system packages measurable sections, representation and resource grammar R, and statistical experiment P, assigning distinct roles to objective, deployment, task visibility, and certification.Changing architecture or representation with the objective fixed changes the obstruction but not the native defect, whereas changing the lifted objective can change the statistical target.Source: §2.2 · p. 27

2.3 Pointwise, population, and uniform risk functionals

• Under nonempty finite-defect architecture assumptions, every deployed field satisfies an exact architecture decomposition for the pointwise, population, and uniform risk functionals.The expectation may be extended-valued, while the finite-population decomposition assumes the architecture contains a field with finite population defect and the deployed field itself has finite population defect.Source: §2.3 · p. 27

8 CHAPTER 2. CERTIFIED ELIMINATION SYSTEMS

• Elimination geometry separates an irreducible class-level floor from failure to reach the class optimum.Source: §8 · p. 28
• Uniform risk can retain a fixed topological-seam tax even when average risk makes the seam’s probability mass arbitrarily small.Source: §8 · p. 28

2.4 Regular and singular oracle geometry

• Oracle geometry ranges from unique regular maps and separated finite covers to singular families and set- or distribution-valued objects.Regular cases shift obstruction toward variation, resources, or decoder capacity; covers can fail through monodromy, while singularities and point-valued representations create distinct defects.Source: §2.4 · p. 28

2.5 Defect fibrations and fine-to-coarse maps

• Defect fibrations decompose fine-to-coarse elimination into the cost of choosing a coarse state and the tax for realizing it in the fine fiber.For restricted fine architectures, the resulting obstruction is expressed through an infimal realization form.Source: §2.5 · p. 28

2.6. COMMON LOSS SCALES, ATTAINMENT, AND APPROXIMATION

• The obstruction tower is a Bellman or min-plus recursion that separates coarse-state selection from canonical fine-fiber realization.An objective-faithful coarse representation need not be obstruction-faithful; distinct loss quantities should be combined only through an exact identity, infimal decomposition, or conversion theorem.Source: §2.6 · p. 29

2.6 Common loss scales, attainment, and approximation

• Zero obstruction means exact realization only when the infimum is attained; otherwise it may indicate arbitrarily close approximation, while positive uniform obstruction can remain despite zero average obstruction.Attainment requires conditions such as compactness, lower semicontinuity, coercivity, or finite-dimensional closure, so min must not replace inf without justification.Source: §2.6 · p. 29

• For finite instance spaces and auxiliary fibers, the population architecture obstruction is attained for every nonempty architecture class.This exercise asks for a proof of universal attainment under finiteness conditions.Source: Exercises · p. 29
• A continuous selector on a circle with a shrinking-measure seam can yield 𝔒𝑃(𝔄) = 0 but 𝔒∞(𝔄) > 0.The construction separates the population and infinite-sample obstruction measures.Source: Exercises · p. 29
• Under a label permutation in a mixture model, a coarse oracle object can have zero coarse defect while a labeled fine architecture incurs a positive realization tax.The exercise highlights how representation granularity changes the realization cost.Source: Exercises · p. 29

A Four-Component Decomposition of Population Risk

• The decomposition distinguishes local model inadequacy, deployment nonrealizability, finite-sample uncertainty, and optimization error, assigning each mechanism a separate reference object and intervention.Architecture obstruction is measured after fixing the local oracle family, so it cannot be silently merged with model misspecification.Source: A Four-Component Decomposition of… · p. 30, §3.1 · pp. 31–32

3.1 Reference risks and the four components

• The four components are Δmodel for local approximation, Δarch for deployment-contract loss, generalization deviation for population–empirical gaps, and Δopt_n for within-class implementation suboptimality.Quantitative addition requires a common risk scale or an explicit exchange theorem; heterogeneous quantities are not automatically commensurable.Source: §3.1 · pp. 30–31

3.2 A high-probability learning bound

• On a fixed-class uniform generalization event, the certified bound places learned excess risk around the structural floor Δmodel + Δarch plus generalization and optimization terms.The assumptions include R★≤Roracle≤inf_A∈𝔄R(A), while data and optimization can reduce their respective terms but not a positive structural floor.Source: §3.2 · p. 32

3.3 Structural meaning and mechanism coupling

• The components are diagnostic rather than independent: enlarging local families, randomizing deployment classes, adding test-time computation, or limiting optimization can shift error between components.The appropriate intervention follows the dominant mechanism, such as changing the local model, deployment contract, information, statistical control, solver, or computation budget.Source: §3.3 · p. 33
• Refined decompositions of architecture obstruction are valid only under declared assumptions and exact identities or exchange bounds, because distinct loss scales cannot be combined by analogy.Examples separate carrier information loss from decoder nonsaturation, rectangularization from coordination, and fine-to-coarse representation costs from other terms.Source: §3.3 · p. 33

3.4 What an empirical study should separate

• A convincing empirical study must separately control model inadequacy, finite-sample uncertainty, and implementation error before interpreting a performance plateau as structural.The prospectively frozen MNIST audit fixes the local variational contract, uses held-out selection and evaluation, audits the reference, and tests data, width, and training alternatives.Source: §3.4 · p. 34

Proof of the exact four-component risk identity

• The exact risk identity separates model, architecture, optimization, generalization, and implementation terms, with residuals vanishing under population or empirical attainment.The identity requires only finite evaluations and infima, without stochastic assumptions or architecture-infimum attainment.Source: Proof of the exact four-component risk… · p. 35

Proof of the four-component certified learning bound

• The certified four-component bound uses only predictor-class membership, a uniform generalization event, and empirical optimization residuals; population attainment is unnecessary.Subtracting the optimal risk yields the model-plus-architecture contribution in the upper bound.Source: Proof of the four-component certified… · pp. 35–36

4.1 Four recurring structural problems

• The four laboratories apply one decomposition discipline while isolating distinct mechanisms: carrier information loss, graph-validity repair, shared-history coordination, and computation-indexed one-pass defects.Each laboratory pairs its obstruction with a matched repair, and the common vocabulary does not imply a common mechanism.Source: §4.1 · p. 37, §4.6 · p. 42

4.2 Laboratory I: oracle-preserving coarse-graining

• Coarse-graining is lossless exactly when the carrier preserves the conditional oracle signature; otherwise carrier refinement or faithful conditional back-mapping targets the corresponding native defect.Representation size is not the criterion: projector quotients can preserve the oracle, while merged conditional means create irreducible prediction floors.Source: §4.2 · pp. 37–38

4.3 Laboratory II: graph-indexed distribution learning

• Graph-indexed distribution learning requires nonlinear projection onto the valid CDF set because signed high-order smoothing can violate probability validity; this can preserve statistical rates while restoring semantics.The raw ambient solution, validity-repaired field, and constrained CDF optimizer are generally distinct.Source: §4.3 · p. 38, §4.4 · p. 39

4.4 Laboratory III: shared conditional decisions

• Shared conditional decisions expose coordination and memory defects: rectangularization enables independent kernels, while shared parameters or coarse memory can impose occupancy-dependent fiber taxes that width alone cannot remove.Rectangularization, memory refinement, or a task-specific quotient are the matched repairs.Source: §4.4 · p. 39

4.5 Laboratory IV: one-pass prediction and local refinement

• One-pass local refinement traces an exact computation–defect frontier, while adding a quadratic feature removes the baseline obstruction by changing the deployed class rather than merely improving optimization.The three-point ledger reports exact population values, and the quadratic calculation transfers an existing floor without proving positivity.Source: §4.5 · pp. 40–41, Table 4.1 · p. 41

Where the four accounts sit in learning theory

• The framework inserts a contract-relative architecture reference between local adequacy and data- or implementation-governed behavior, separating global realizability from amortization-specific gaps.Its exact identity and certified bound are presented as an accounting synthesis whose diagnostic value lies in changing distinct reference comparisons.Source: Where the four accounts sit in… · p. 43, Ownership of the Part I results · p. 43

5.1 Legendre conjugacy

• The insertion price is the directed Bregman divergence generated by the eliminated objective, with the trial state first and oracle second; reversing arguments generally changes loss and guarantees.In entropy models, this fixes the algorithm’s KL direction, while the reverse direction requires a different elimination.Source: §5.1 · p. 46, §5.2.1 · p. 46

5.2.3 Quantum relative entropy

• Quantum relative entropy preserves defect orientation but requires supp(ρ) ⊆ supp(σ) for finiteness, with ordinary finite-dimensional operators only on the faithful domain.The noncommutative setting also changes the validity of base change and measurement.Source: §5.2.3 · p. 47

5.3 Further certified eliminations and boundaries

• EM’s native insertion certificate is the forward variational KL from a trial law to the exact posterior, as generated by the Gibbs–Fenchel lift [Dempster et al., 1977, Neal and Hinton, 1998].The certificate is not an abstract posterior distance but the KL orientation induced by elimination.Source: §5.3 · p. 47
• Saddle elimination yields a primal–dual residual certificate whose signs are fixed by min–max order, while curvature-based exchanges convert native defects to geometric errors only under local analytic conditions.These exchanges are not definitions of the defect.Source: §5.3 · p. 48

Elimination Towers and the P/G/X/V/C Calculus

• The P/G/X/V/C calculus separates pointwise elimination, external field transformation, target-jet repair, variational reoptimization, and fixed-marginal coupling by the contract object each operation changes.Numerical equality does not imply operational equality: G and V may coincide in a special quadratic case, while X does not turn G into V and C cannot remove insertion defects or change additive expectations.Source: Elimination Towers and the P/G/X/V/C… · p. 51, §6.1.2 · p. 52, §6.1.3 · p. 52

6.3 Horizontal operations and multistage towers

• Multistage tower defects compose by infimal convolution over compatible declared fibers, enabling dynamic-programming decompositions without implying equal runtime, conditioning, or finite-precision behavior.The tower separates wrong coarse-object choice, fine-architecture realization cost, and implementation error.Source: §6.3 · pp. 53–54

6.4 Worked decompositions and coupling boundaries

• Bregman projection decomposes error into irreducible constraint, within-constraint implementation, and possible nonorthogonality components; exact numerical agreement between G and V can disappear under constraints or nonlinear repair.The projection interpretation assumes standard differentiability and convex-projection conditions.Source: §6.4 · p. 55
• Fixed-marginal coupling cannot change additive coordinate expectations; benefits require genuinely joint functionals such as variance, maximum loss, simultaneous coverage, or path-dependent objectives.This boundary underlies applications including common random numbers, antithetic couplings, ensemble dependence, and variance reduction.Source: §6.1.5 · p. 52, §6.4 · p. 55

Proof of tower associativity

• Tower associativity follows by identifying the same minimum over compatible intermediate states, while the exercises distinguish insertion defects, objective improvements, and coupling effects.Fixed marginals protect additive coordinate expectations, so fixed-marginal antithetic coupling can improve joint but not additive functionals.Source: §6.4 · pp. 55–56, Proof of tower associativity · p. 57

7.3 Approximation, transmission, and certification error

• Certified optimization requires separate budgets for inner-oracle, defect-score, and outer-solver errors; with summable errors and regularity, every cluster point is stationary for the target.This target-level conclusion depends on exactification restoring the target score; generic surrogates may converge to stationary points of another objective. Conjugate transmission vanishes beyond affine natural-parameter maps.Source: Exactification and Optimization… · p. 58, §7.3 · pp. 61–62

7.1. JET EXACTIFICATION

• Exactification matches the target value and gradient at the anchor by subtracting the frozen defect jet, while retaining auxiliary geometry.Under C^k regularity, Q_X(¯Θ | ¯Θ) = J(¯Θ) and ∇Q_X(¯Θ | ¯Θ) = ∇J(¯Θ); this ensures target tangency but not useful steps or global convergence.Source: Exactification and Optimization… · p. 58, §7.1 · p. 59

7.2 The converse normal form

• Any scalar surrogate preserving the target k-jet must subtract the defect jet up to a k-flat remainder, a local normal form rather than global optimizer equivalence.Conversely, every k-flat perturbation preserves the target k-jet, so the canonical correction is locally unique modulo such terms.Source: §7.2 · p. 60, Proof of the converse exactification… · p. 63

7.4 Practical algorithm

• The practical first-order algorithm freezes an approximate globalized field, certifies its defect value and score, accepts only budget-valid decreases, and otherwise shrinks the trust region.Collapsing inner-oracle, defect-score, and outer-minimization tolerances into one generic tolerance obscures the proof.Source: §7.4 · p. 62

8.1 The CDF set and projection repair

• Projection repairs invalid signed forecasts without increasing squared CDF error or graph edge roughness, but it is not generally the constrained optimizer of a different penalized objective.The result assumes compact-support CDFs in the declared L2 geometry and uses the same coordinatewise projection at each vertex.Source: §8.1 · p. 65

8.2 CRPS risk and its observable decomposition

• CRPS excess risk equals squared L2 CDF error, so projection cannot increase excess CRPS and the graph effective dimension is tr(S^2), not merely tr(S).For fixed data-independent filters, diagonal self-influence correction yields an unbiased raw-risk estimate, while final projection restores validity and supplies a pathwise error bound.Source: §8.2 · p. 66

8.3 The probability-validity barrier

• The probability-validity analysis separates post-hoc projection repair from the sharper question of when an ambient linear filter preserves CDF validity without repair.The classification and path-rate result are scoped by the stated source conditions.Source: §8.3 · p. 67
• Graph-universal Markov validity requires a spectral generator to satisfy the Laplacian gate; Bernstein functions with ψ(0)=0 pass it, while polynomial penalties reduce to nonnegative multiples of L.This linear probability-preserving barrier does not exclude nonlinear projection repair.Source: §8.3 · p. 67

8.4 Sharp path rates

• On growing paths, a calibrated Richardson filter with qualification m≥s attains the stated average excess CRPS rate, and an embedded Bernoulli experiment supplies a matching lower bound.The theorem assumes the declared growing-path source class, compact outcome support, independent vertex samples, and calibrated qualification.Source: §8.4 · p. 68

Proof of validity repair

• Metric projection onto a nonempty closed convex validity set contracts squared-CDF risk and, with the graph argument, preserves the declared validity and roughness guarantees.The proof uses the Hilbert-space projection theorem, coordinatewise nonexpansiveness, and the graph Dirichlet-form representation.Source: Proof of validity repair · p. 69

What is and is not claimed

• The classical engines are reused rather than claimed as new mathematics, while the program-specific contribution is auditing native defect units, orientation, elimination consistency, and target-jet preservation.The source-program exactification normal form is presented as a claimed increment, not an exhaustive priority assertion.Source: Classical engines · pp. 69–70, What is and is not claimed · p. 70

Validity of Defects and Representations

• Local elimination defects support architecture claims only when they integrate to one global objective and their carrier is intrinsic to the declared target.Nonnegative fiberwise-touching increments and zero square circulation yield a certified global defect potential on the product-of-stars complex.Source: Validity of Defects and Representations · p. 71, §9.2 · p. 73

9.2 Square curvature

• Zero square circulation is equivalent to path-independent potentials under the theorem’s nonnegativity and fiberwise-touching conditions, whereas nonzero circulation certifies elimination-order dependence.On general complexes, harmonic periods require an additional audit beyond local face tests.Source: §9.2 · pp. 72–73, §9.3 · p. 74

9.3 Approximate flatness and Hodge repair

• Approximate flatness bounds order ambiguity by accumulated square curvatures, but repair may additionally require removing global periods and projecting onto certified nonnegative/touching fields.The broader weighted Hodge and repair theory is attributed to Elimination Geometry II.Source: §9.3 · p. 74

9.4 Repair, falsification, and audit boundaries

• Integrability audits convert local curvature into falsification certificates and are contract-sensitive: a score field must satisfy gradient and period conditions for a density claim, but need not be conservative for a transport sampler.A density claim additionally requires an integrable, normalized exponential potential, while operational curvature concerns whether a legal task can observe internal differences.Source: §9.4 · p. 75, §9.4.1 · pp. 75–76

10.1 Why exact lifts can be vacuous

• Exact target-calling lifts show that finite extension size alone cannot certify meaningful lift complexity or explain the target.Affine fields can be appended with coefficient-independent overhead, while quadratic penalties can manufacture any prescribed positive-semidefinite metric.Source: §10.2 · p. 79

10.2. EPIGRAPH GEOMETRY AND INTRINSIC LIFT STRUCTURE

• Admissible conic lifts are exactly factorizations of the target slack operator, after quotienting slack-invisible directions and reducing to the relevant minimal face.The gate covers polyhedral and semidefinite cases through nonnegative and positive-semidefinite factorizations, respectively.Source: §10.2 · p. 80

10.3 Gauge reduction and the fidelity–coherence–complexity frontier

• Gauge reduction is necessary because duplicated or cone-gauge-related factors can improve naive coherence without changing the represented objective.A quotient-reduced slack carrier removes directions invisible to the target and prevents factor-splitting artifacts.Source: §10.3 · p. 80

10.4. TARGET-VISIBLE REDUCTION AND EXTRACTION

• The visible quotient R_U = S_U is the coarsest deployment-state quotient preserving optimal native defect, modulo null sets.Raw labels with the same conditional dual signature count as one visible state; finite separated oracle types therefore incur positive quantization defect below their number.Source: §10.4 · pp. 81–82

10.4 Target-visible reduction and extraction

• Under quotient-faithful extraction, native exchange, and finite visible capacity, the exact lift frontier transfers to a visible-capacity frontier rather than being determined by cone size alone.Cone size m and visible capacity k remain distinct units, and no universal conversion factor follows from cone order.Source: §10.4 · pp. 81–82

10.5 From lift size to native defect

• Extension complexity and native defect occupy different units: objective rescaling changes defect without changing lift size, while runtime requires encoding, conditioning, barrier, and linear-algebra assumptions.A barrier-to-defect exchange can provide a certificate only for a fixed encoded, well-conditioned conic lift under a declared computational model.Source: §10.5 · p. 83

Proof of the finite-complexity target-calling no-go

• The target-calling no-go follows by appending the representable epigraph of ψ(a − g(x)), whose unique minimizer is a = g(x), s = 0.Product-stable cone dictionaries add at most c_ψ(q), and quadratic specialization yields Hessian M^T M.Source: Proof of the finite-complexity… · p. 85

Proof of the conic lift–factorization gate

• The conic lift–factorization proof constructs slack factorizations from primal and dual cone witnesses and reconstructs the target body by bipolarity, including minimal-face restriction for nonproper lifts.The visible-quotient proof separately uses conditional Bregman decomposition to show that any defect-preserving coarsening must retain S_U, then takes the frontier infimum.Source: Proof of the conic lift–factorization… · p. 86, Proof of the visible quotient and… · pp. 86–87

11.1 From a local defect to a class-level floor

• Architecture obstruction is a class-level second elimination whose tax depends on the architecture, native defect, deployment contract, and population weighting rather than architecture alone.Changing the loss, population weighting, or output contract can change the obstruction.Source: §11.1 · p. 89

11.2 Saturation

• Saturation means zero obstruction, while exact saturation additionally requires an architecture class to contain an exact oracle section; generic universal approximation is neither necessary nor sufficient without an exchange theorem.A projector-valued architecture can be exactly saturated for an eigenline oracle even when a vector-valued architecture is obstructed by sign ambiguity.Source: §11.2 · p. 90

11.3 Architecture comparison and witnesses

• Architecture comparisons should bracket the native obstruction with constructive upper witnesses and flow, topological, resource, or information-based lower certificates.Nonmatching witnesses define an auditable unresolved region, while larger classes may still worsen generalization or optimization.Source: §11.3 · p. 90

11.4 Architecture obstruction versus approximation and optimization

• Unlike ordinary approximation theory, architecture obstruction can target fibers, orbits, or sets under sharing, memory, topology, and output-semantic constraints, and may vanish under quotient or atlas representations.Training loss alone cannot distinguish an architectural floor from optimization failure; the proposed comparison varies optimization, parameters, targeted contract changes, and capacity-matched repairs.Source: §11.4 · p. 91

11.5 Output, risk, and theorem contracts

• A valid architecture theorem must specify the oracle set, lower obstruction bound, attaining or approximating witness, and contract repair, with output, aggregation, regularity, and native defect declared.Point, set, orbit, projector, probability-law, and charted-section outputs can incur different taxes under uniform, average, tail, or task-weighted risk.Source: §11.5 · pp. 91–92

12.1 The regular regime and quarter transport bound

• The oracle-variation transport theorem converts oracle variation exceeding an L-Lipschitz deployment budget into a quantitative native-defect lower bound.The bound can be zero when oracle variation fits within the architecture budget, so it is a certificate rather than a universal smoothness failure claim.Source: §12.1 · pp. 93–94

12.2 Finite calibration, dual flow, and population thickening

• Finite calibration converts uncertain local-oracle mismatches into quadratic graph certificates and flow-like dual certificates, while population thickening requires nearby population mass to turn pointwise mismatch into positive tax.Without a mass condition, a deterministic obstruction at a point in a nonatomic population can have zero average cost.Source: §12.2 · p. 94, §12.3 · p. 95

12.3 Set-valued oracles and certificate boundaries

• For set-valued oracles, regular branchwise transport requires separated, transport-consistent branches; collisions require singular analysis, and positive graph certificates do not establish oracle correctness, task visibility, atlas generalization, or faithful population geometry.Each of these interpretations requires a later validation gate; the exercises reinforce monotonicity, output-contract dependence, optimizer-versus-saturation diagnosis, and explicit certificate calculations.Source: §12.3 · pp. 95–96

Exercises

• The exercise asks how a lower mass condition converts a pointwise defect lower bound into a population bound and where deployed-field regularity is required.Source: Exercises · p. 96

Proof of the oracle-variation transport bound

• The transport proof derives the metric bound from the triangle inequality and the L-Lipschitz architecture contract, then obtains the native-risk conclusion using D_x(a) ≥ (μ/2)d^2.Source: Proof of the oracle-variation… · p. 97
• Applying the quadratic inequality to every self-coupling of P_X and using equal marginals supports the supremum-over-couplings and infimum-over-architectures steps.Source: Proof of the oracle-variation… · p. 97
• Combining the two bounds after taking the infimum over A yields the coefficient μ/8.Source: Proof of the oracle-variation… · p. 97

13.1 Conditional architectures and their rectangular hull

• Rectangularization preserves local kernels while removing cross-history sharing, making the obstruction Bellman-separable across histories.The rectangular hull is an analytical relaxation, not the deployed architecture.Source: §13.1 · p. 98, §13.2 · p. 100

13.2. THE COORDINATION DECOMPOSITION

• The exact architecture loss decomposes into a rectangular local obstruction plus a nonnegative coordination tax for forcing one shared parameterization, memory state, or network.The tax is measured in the native forward-KL/Bellman scale and isolates the cost of cross-history coupling after local kernels are fixed.Source: §13.2 · p. 100, Proof of the rectangular Bellman and… · p. 104

13.2 The coordination decomposition

• Local Bellman residuals form a diagnostic map that localizes the global coordination tax to histories where sharing is most expensive.For fixed occupancies, the shared-kernel tax is nonnegative and vanishes exactly when all positively weighted oracle kernels coincide; its normalized form is the geometric-pooling formula.Source: §13.2 · p. 101

13.3 Memory compression as base change

• Memory compression changes which histories remain distinguishable, so coarse objective faithfulness can coexist with failure to realize canonical fine conditional kernels.Universal equality of fine and coarse obstructions requires saturation by the canonical oracle fiber kernels, while approximate saturation gives an additive upper bound.Source: §13.3 · p. 102

13.4 Zero temperature

• Zero-temperature limits need not become deterministic, because the admissible architecture class and its closure determine the hard limit.Rectangularity concerns conditional-kernel pasting, not network width, and is not automatically restored by more data.Source: §13.4 · p. 102

Singular Fibers, Monodromy, and Catastrophe Taxes

• Singular fibers create catastrophe taxes when monodromy permutes oracle branches incompatibly with the declared output contract.A quotient-valued output such as a projector or unordered root set can remain globally well defined even when a point-valued representative cannot.Source: Singular Fibers, Monodromy, and… · p. 105, Figure 14.1 · p. 105

14.1 The radical normal form

• For the radical normal form, a continuous point-valued selector exists exactly when k divides ℓ; otherwise the sharp uniform residual floor is r^2.The zero deployment attains this value, but the result does not imply a positive unrestricted nonatomic average-risk lower bound.Source: §14.1 · p. 106, §14.2 · p. 107

14.2 Risk level, general loops, and degree

• The same obstruction extends to general loops, polynomial roots, and eigenlines: incompatible homotopy or monodromy yields native loss floors governed by clearance from the discriminant.These taxes are stable while homotopy and positive clearance persist, can vanish near the discriminant, and may reflect a representation artifact repaired by quotient outputs; the objective-denominated extension builds on topology-of-algorithms work [Smale, 1987, Vassiliev, 1988].Source: §14.2 · p. 107, §14.3 · p. 107, §14.4 · p. 108

Proof of the exact radical catastrophe tax

• Every continuous point-valued root selector in the radical family incurs uniform squared defect at least r^2 when k∤ℓ.The lower bound follows from winding-number incompatibility and applies to every continuous A.Source: Proof of the exact radical catastrophe… · p. 109

15.1 From obstruction diagnosis to a repair contract

• The correct repair is the smallest contract change that removes the certified obstruction while preserving downstream information, rather than an interchangeable menu of fixes.The framework distinguishes representative-selection, forced-sharing, semantic-interface, and insufficient-computation failures.Source: §15.1 · p. 110

15.2 Local and symmetry-respecting repair

• A declared measurable atlas with locally exact experts achieves zero uniform obstruction, but the existential theorem does not learn the cover, gate, experts, or minimal chart count.The result requires a genuine measurable cover, valid routing, and expert classes containing exact local oracle sections.Source: §15.2 · p. 111
• Quotient, set-valued, or probability-valued outputs can remove distinctions the task cannot observe, whereas ambient averaging may leave the oracle set and incur positive native defect.For square roots, the full root set or invariant root law is globally defined, while averaging a and −a gives zero on the discriminant.Source: §15.2 · p. 112

15.3 Repairing deployment mechanisms

• Deployment repairs target distinct mechanisms: memory enlargement or rectangularization removes history collisions, semantic projection restores validity, and test-time optimization expands computation under a declared budget.These repairs have different costs and test-time refinement can still follow the wrong branch or remain trapped by representation obstruction.Source: §15.3 · p. 112, §15.4 · p. 113

15.4 Selection, validation, and repair frontiers

• Repair selection requires positive baseline obstruction, mechanism-specific diagnosis, reduced repaired obstruction, and independent held-out improvement, with sample splitting separating certification, training, and testing.For atlas repairs, the preferred statistical object is the inverse frontier: the best loss attainable with K charts.Source: §15.4 · p. 113

Three obstruction mechanisms, three literatures

• The deployment quantifier separates regular variation, coordination, and singular branch-selection failures, each requiring matched witnesses and repairs rather than relabeling classical ingredients as new discoveries.The section identifies atlas, quotient, set-valued, randomized, memory, rectangularization, and sharing-rule repairs across these mechanisms.Source: Three obstruction mechanisms, three… · p. 114

16.1 The architecture as carrier and decoder

• A deployment architecture must preserve oracle-relevant distinctions in its carrier and make them decodable; failure arises either from carrier information loss or from an insufficient decoder class.Resource limits constrain carrier size and decoder complexity, while native objective defect determines the price of merging oracle signatures.Source: §16.1 · p. 117

16.2 Canonical Bregman carrier

• The canonical Bregman decomposition splits native architecture distortion into carrier information loss and resource-limited decoder nonsaturation, with zero distortion requiring both task-relative sufficiency and oracle realization.The identity is an infimum decomposition; the zero-distortion characterization additionally requires an attaining admissible pair.Source: §16.2 · p. 118, Proof of the canonical architecture… · p. 123

16.3 Discrete and continuous carrier extraction

• Discrete and continuous carrier extraction convert label, coherence, geometric, and site constraints into limits on distinguishable oracle signatures, with strong defect growth turning capacity or covering bounds into native distortion floors.For discrete carriers, separated signatures attain the visible cardinality bound; continuous lower exponents require co-Lipschitz readout, an interior region, dense deployments, and the finite-site cap.Source: §16.3 · p. 119

16.4 Finite frontiers and phase diagrams

• Finite Bregman quantization yields positive error below the number of distinguishable types, while weighted path models reduce to scalar quantization and produce an exact capacity–resolution phase diagram.Within each assigned cell, the optimal codeword is the dual-coordinate centroid, and the operational audit distance is a clipped discrete covering radius.Source: §16.4 · p. 120

16.5 An exact data-selection frontier

• Selection budget is measured directly in excess squared loss: the worst-case loss ratio is exactly F(d,n) = 1 + Dsel(d,n), with Dsel(d,n) = 1 2n−1 for 1 ≤ d ≤ 3, n ≥ 1.This formula resolves the low-dimensional mean-estimation row of the published data-selection problem; these statements are not counted among the book’s principal results [Hanneke et al., 2025, Huang, 2026a].Source: §16.5 · pp. 120–121

16.6 Resource limits and operational boundaries

• For large AI systems, meaningful resource coordinates include memory, active routes, communication, test-time iteration, output state, and parameter-tying coherence; the relevant quantity is oracle-relevant distinction surviving the carrier and decoder.A positive native distortion need not affect a downstream task unless the lost distinctions are visible through the transmission and exposure gates.Source: §16.6 · p. 122

Exercises

• The exercises apply the framework to conditional-mean decoding, two-codeword scalar quantization, and resource grammars for bounded-context transformers with fixed active experts.They ask which quantity might serve as the oracle signature in the transformer setting.Source: Exercises · p. 122

17.1 From internal difference to task-observable kernels

• Task envelopes preserve exactly the values of a declared task contract, while contextual completion yields the least task-adequate, grammar-nonexpansive pseudometric and a fully abstract quotient relative to legal contexts and tasks.Two pipelines are equivalent exactly when no legal context followed by a legal task distinguishes them; this equivalence is task- and grammar-relative. Attouch–Wets metrics are not the theorem being asserted, and Proposition 17.1 is proved directly.Source: §17.2 · pp. 125–126

17.3 From native defect to operational separation

• A positive native architecture floor implies positive contextual separation only when the transmission gate preserves the defect in the operational kernel and the exposure gate supplies a legal context that selects the affected states.These gates connect native architecture defects to operationally observable differences.Source: §17.3 · p. 127
• Native defects yield a resolution-dependent operational profile: coarse task contracts can hide internal seams, whereas more discriminating contracts can expose positive contextual separation.Finite task families admit exact margin-clipped lower bounds, while compact metric task classes yield coefficients governed by task and kernel Lipschitz moduli.Source: §17.3 · p. 127

17.4 Ordering curvature and contextual witnesses

• Pairwise commutation at one input does not ensure contextual interchange coherence, so ordering curvature and precontexts are required to detect hidden noncommutation.A finite-memory witness shows that all orders can agree under a zero-input contract yet differ by one after a legal precontext, invalidating single-input audits.Source: §17.4 · p. 127

Proof of task-envelope reconstruction and isometry

• The task envelope preserves task values and is the pointwise-least representative of each task-equivalence class, with monotonicity and idempotence following from its construction.The finite-state setting ensures extrema are attained, while the proof establishes both value preservation and minimality.Source: Proof of task-envelope reconstruction… · p. 129

Proof of contextual completion and full abstraction

• Context closure produces a fully abstract quotient: contextual distance is the least task-adequate, grammar-nonexpansive pseudometric, and its zero set is the greatest sound congruence.Equality of contextual envelopes is equivalent to zero contextual distance under the closed contract.Source: Proof of contextual completion and… · p. 130

18.1 Conditional chain rules and architecture composition

• Soft conditional certificates compose only in a common Gibbs objective scale: shared references, temperatures, and normalizations permit addition, while arbitrary local losses do not.The global free-energy defect is nonnegative and vanishes exactly at the oracle path law, up to null histories; trial laws may remain history-dependent.Source: §18.1 · p. 132, §18.2 · p. 133

18.2 Base change and data processing

• Base-change identities require fine fibers to remain visible: objective preservation alone is insufficient, and universal equality of fine and coarse obstructions requires canonical-fiber saturation.The fiber realization term depends on coarse occupancy, so replacing it with a constant can be incorrect; the additional variants are imported results [Huang and Wang, 2026b,h].Source: Composition, Base Change, and… · p. 131, §18.2 · p. 133

18.3 When conditional elimination destroys encoder semantics

• Conditional Markovization preserves receiver-visible marginals but can impose a semantic ambiguity tax by leaving the deterministic carrier class; rectangular symbol fibers are exactly the zero-tax case.Candidate failure is not an architecture certificate: exclusion requires testing the best rectangular architecture after complete reoptimization, and the certified result is not a principal book result [Huang, 2026b].Source: §18.3 · pp. 134–135

18.4 Hard limits and composition boundaries

• Zero temperature does not force determinism: randomized hard limits remain possible when the architecture closure contains mixtures but excludes pure selectors.For temperature-dependent classes, convergence additionally requires Γ-convergence and control of inner and outer limits.Source: §18.4 · p. 136
• Composition requires compatible defect units, correct conditioning weights, measurable nonempty fibers, rectangularity for commuting infima, oracle-preserving representations, and downstream visibility of retained differences.The exercises probe conditioning order, coupling order, quantization order, temperature limits, realization tax, and randomized zero-temperature limits.Source: §18.4 · p. 136

Proof of the soft conditional chain theorem

• Soft conditional-chain composition is valid when normalized transitions, absolute continuity, and relative entropies are well defined; telescoping and the KL chain rule yield equality exactly when all conditional laws agree.Sequential factorization then gives equality of the trial and reference path laws.Source: Proof of the soft conditional chain… · p. 138

Statistical Certification

• Finite-data certification transports a covered confidence world through a declared query, authorizing a resolved label only when all compatible worlds agree.The identified image is the complete finite-data answer; unresolvedness reflects disagreement among compatible population worlds.Source: Confidence Worlds and Statistical… · p. 141, §19.2 · p. 143

19.1 Confidence worlds and finite-data permission

• Confidence worlds must satisfy the declared coverage and stopping contract, contain every population component needed by the query, and are not substitutes for posterior or point-estimate selection.Construction methods may vary, but the coverage and reporting contract must remain explicit.Source: §19.1 · p. 142

19.2 Identified images, margins, and honest certificates

• The identified image and margin support auditing across tolerances, while model conflict and world disagreement remain distinct non-binary outcomes.A positive margin measures the distance of the entire identified image from a label change.Source: §19.2 · p. 143
• The standard confidence certificate is honest and maximally decisive on the coverage event: it returns the common truth label when compatible worlds agree and unresolved otherwise.A binary decision is unauthorized when either label remains possible in the confidence world.Source: §19.2 · p. 144

19.3 Defect envelopes and architecture grammars

• A common simultaneous event brackets the architecture frontier using defect envelopes and inner/outer grammars across resource budgets, with honesty inherited from that event.The bracket describes statistical resolution, not optimization failure, and does not relabel an empirical frontier as a population frontier.Source: §19.3 · pp. 145–146
• Outer enclosures preserve honesty for resolved declarations but can make unresolved regions conservative, distinguishing computational enclosure from genuine population ambiguity.This limitation applies when the exact identified frontier image is unavailable.Source: §19.2 · p. 144, §19.3 · p. 146

19.4. PERSISTENT ATLASES AND TRANSPORT

• Uniform defect perturbations interleave atlas curves and stabilize the inverse chart-budget/loss frontier, while nonexpansive transport reuses the same confidence event for observable frontiers.The result concerns achievable loss rather than consistency of a particular greedy atlas; the finite-part stability constant is one.Source: §19.4 · p. 147

19.4.1 Pointwise feasibility versus common deployment

• Pointwise feasibility in every compatible world does not imply a common deployable witness: at tolerance 𝜂=0, each world can have zero loss while every fixed architecture has worst-world loss one.The two deployment values can differ because world-specific feasibility and common-witness feasibility use different quantifier orders; more data may shrink the confidence world, whereas a router changes the grammar.Source: §19.4.1 · p. 148, §19.4 · p. 149

19.4.2 Distribution-free prediction under conditional and shift contracts

• Distribution-free marginal coverage does not provide exact conditional coverage across a rich class, while weighted conformal repair works only with invariant conditional laws and adequate source-to-target density-ratio information.Typed alternatives include set-valued outputs, abstention, target calibration data, or restricting the target or conditioning domain.Source: §19.4.2 · pp. 149–150

19.4.3 Statistical ancestry: pointwise success is not uniform deployment

• Classical superefficiency and sparse oracle results show that pointwise success cannot be exchanged for uniform deployment: ∀𝜃∃n_0(𝜃) does not imply ∃n_0∀𝜃.The deployment rule and sample size are fixed before the unknown world is revealed, making the second quantifier order operationally relevant [Le Cam, 1953; Hájek, 1970, 1972; Leeb and Pötscher, 2005, 2008].Source: §19.4.3 · p. 150

19.5 A complete quantifier case: sparse model-selection procedures with pointwise oracle guarantees

• The sparse-selection case study uses an orthogonal Gaussian coordinate model to expose exact threshold behavior, while general designs add identifiability, pilot, and optimization conditions without removing the quantifier conflict.The analyzed class includes nonconcave penalization, data-adaptive weighting, and exact or algorithmic ℓ0 subset selection.Source: §19.5 · p. 150

19.5.1 SCAD: fixed-signal unbiasedness leaves a moving transition band

• SCAD and Adaptive LASSO preserve fixed-parameter oracle behavior but move nonuniformity into shrinking threshold or pilot-dependent lockout regions near zero.For SCAD, β_j,n=cλ_n is deleted with probability tending to one; Adaptive LASSO has the analogous behavior at β_j,n=ct_n.Source: §19.5.1 · p. 151, §19.5.2 · p. 151

19.5.3 BSS and BeSS: a method-specific repair cannot create information

• BSS can repair a frozen Adaptive-LASSO lockout corridor, but method-specific retuning does not create information, and exact support recovery still requires signal separation, design identifiability, and controlled optimization error.BeSS adds active-set search and model-size selection, yet its repair remains algorithmic rather than information-theoretic.Source: §19.5.3 · p. 152

19.5.4 The information-limited core and the honest action

• Indistinguishable local worlds create an information-theoretic blind region that no same-data selector can remove, motivating restricted domains or the three-way action stably nonzero, stably zero, or unresolved.The section distinguishes this region from method-induced lockout and requires design margin, model-size rules, and optimization error to be declared for uniform support claims.Source: §19.5.4 · p. 152

19.7. ACTIVE-SET CONFLICT AND REPAIR DEPTH

• Changing learners alone cannot repair observational deployment conflict; valid repairs instead alter information, the action class, or the deployment domain under frozen comparison requirements.Identification additionally requires consistency, exchangeability, positivity, and a declared transport population.Source: §19.7 · p. 155

19.7 Flagship candidate family: active-set conflict, saturation, and repair depth

• A positive Jacobian conflict between two oracle neighborhoods yields a certified native-loss floor for every shared one-pass thresholded-affine encoder.The obstruction depends on permitted perturbations, shared positive coordinates, and quantitative row incompatibility, not merely differing active sets.Source: §19.7 · p. 156

19.7. ACTIVE-SET CONFLICT AND REPAIR DEPTH

• Sufficiently deep proximal-gradient repair removes the certified floor when recurrence is allowed and charged, with sharp contraction factor q for the fixed mechanism.The certificate changes the admissible architecture from a one-pass encoder to a recurrent or unrolled repair under the same information and native loss.Source: §19.7 · p. 157

19.7.1 A no-free-lunch theorem for hidden winding

• Without regularity or coverage, finite samples can support identical observations under feasible and impossible winding worlds, so no uniformly valid binary procedure can always decide.An honest procedure must sometimes return unresolved; the cited hidden-winding results are external source results rather than principal formal results here.Source: §19.7.1 · p. 158, §19.8 · p. 159

19.8 Identification, hidden winding, and data-dependent architecture

• Under Lipschitz regularity and bounded angular error, a mixed-integer identification set gives exact compatible degrees, while local spacing can identify one degree.Random-design rates are governed by coverage of locations where an obstruction can hide, including a logarithmic coverage surcharge.Source: §19.8 · p. 159
• Data-dependent architecture selection is separated into calibration, tuning, and held-out operational evaluation on independent biological units.This staged design avoids requiring a fully general random-set theorem in the first implementation.Source: §19.8 · p. 159

Proof of honesty and maximal decisiveness

• Worldwise-correct deployment is possible exactly when feasible action sets intersect almost everywhere over observationally overlapping worlds; unresolved output is forced when compatible truth labels disagree.The honesty proof relies on confidence-world coverage, while maximal decisiveness follows because either resolved declaration would fail in a compatible world.Source: Proof of honesty and maximal… · p. 160, Proof of deployment conflict under… · p. 162

Proof of the honest elimination bracket

• The proof suite establishes honest elimination brackets, inverse-frontier stability under sup-norm defect perturbations, strict common-deployment inequalities, and certified local sparse-code cells.These results require their stated simultaneous coverage, monotonicity, attainment, or full-rank and strict-KKT assumptions.Source: Proof of the honest elimination bracket · p. 160, Proof of atlas interleaving and… · pp. 160–161, Proof of the common-deployment… · pp. 161–162

Proof of the active-set conflict and repair-depth family

• Under the contraction inequality, recurrent repair reaches tolerance when q_t E_0 ≤ d_η: depth zero suffices if E_0 ≤ d_η, while q = 0 gives exactness after one repaired iterate.Source: Proof of the active-set conflict and… · p. 164

20.1 The certificate record and four nonexchangeable objects

• A certificate record must distinguish population truth, confidence worlds, posterior belief, and deployment actions because matching labels do not make these objects interchangeable.A resolved confidence certificate is not a common deployment witness, and high posterior probability is not by itself a worldwise certificate.Source: §20.1 · pp. 165–166
• For any fixed posterior credibility below one, observationally indistinguishable opposite-label worlds can make the posterior remain highly confident while worldwise resolution complexity is infinite.Small posterior error is prior-average rather than uniform frequentist validity; a low-prior indistinguishable world may still be reported incorrectly with probability one.Source: §20.1 · p. 166

20.1.1 Truth-map notation for the standard certificate

• The truth-map certificate is honest at level 1 − α when its confidence world covers the true world, while unresolvedness and model conflict remain distinct states.Resolved outputs are simultaneously valid on the coverage event, and no rule using the same confidence world can make more declarations while preserving worldwise correctness.Source: §20.1.1 · p. 167

20.2 Resolution complexity

• Worldwise resolution requires sampling against the nearest opposite-label world, producing a max–min KL lower bound that becomes infinite when every allocation has zero separating information.The bound concerns expected observations for a worldwise terminal statement, not generalization, posterior entropy, or deployment loss.Source: §20.2 · p. 168

20.2.1 AND/OR phase change

• In multichannel threshold certification, feasibility has an AND geometry requiring coverage of near-boundary channels, whereas impossibility has an OR geometry concentrating on the strongest violation.The optimal feasible-side allocation equalizes coordinate-wise KL rates, while the impossible-side allocation concentrates on a strongest violation; this asymmetry is intrinsic to the certificate.Source: §20.2.1 · p. 169

20.3 Evidence carriers and task-relative sufficiency

• Evidence compression preserves a worldwise certificate exactly when lost pairwise information can be absorbed by nonbinding slack, while binding opposite-label likelihood directions must be retained.This criterion is weaker than classical parametric sufficiency but stronger than preserving only a posterior mean or current decision.Source: §20.3 · p. 170

20.3.1 Average Bayesian loss versus worldwise complexity

• Average Bayesian representation loss can vanish while worldwise characteristic time is infinite, so posterior distortion cannot authorize stopping or bound worst-case sample inflation.The distinction follows from averaging over prior-weighted worlds versus minimizing over opposite-label worlds and maximizing over experiment allocations.Source: §20.3.1 · pp. 171–172

20.4 Recursive certificate posterior quotient

• A recurrent state preserving present certificate color can still fail one observation later, because future experiments may depend on distinctions erased by current color probabilities.N0 preserves present color, N1 preserves the next labeled exposure, and N∞ preserves every finite future acquisition context.Source: §20.4 · pp. 172–173
• The recursive quotient is the largest color-respecting stable relation, and every exact deterministic implementation requires at least N∞ reachable states.Finite refinement terminates, supplies split witnesses for missing distinctions, and yields the chart bound L K ≥ N∞.Source: §20.4 · p. 173, Proof of the recursive quotient and… · pp. 179–180

20.4.1 CertTrack and certificate-aware design

• CertTrack assigns Bayesian task information to experiment selection while confidence sequences or likelihood-ratio processes separately authorize worldwise claims.Under finite-model regularity, the tracker attains the optimal almost-sure leading constant; expected-time equality additionally requires uniform integrability.Source: §20.4.1 · p. 174

20.4.2 A compressed UCBAdmissions audit

• The compressed UCBAdmissions audit finds a pooled male-minus-female admission-rate difference of 0.1416 with p = 4.84 × 10^-22, yet the department certificate remains U.All department point estimates are below ten points, but simultaneous intervals for B–E cross the boundary; Department B is the resolution bottleneck because only 25 female applicants were observed.Source: §20.4.2 · p. 174, §20.4 · p. 175, Table 20.2 · p. 175
• The audit shows that a pooled p-value near 10^-22 can coexist with unresolved department tolerances across a forty-point range because pooled association and department certification use different truth maps.The table does not identify causal discrimination, qualifications, selection, or historical mechanisms without a larger world and additional assumptions.Source: §20.4 · pp. 175–176

Proof that posterior credibility is not worldwise validity

• Posterior credibility is only prior-average: indistinguishable opposite-label worlds can preserve posterior error credibility while making uniform worldwise guarantees impossible and resolution complexity infinite.When the worlds have identical observation laws, every experiment has zero KL divergence for their truth partition.Source: Proof that posterior credibility is… · p. 177

Proof of the worldwise resolution lower bound

• Worldwise resolution requires characteristic information, with zero characteristic information forcing infinite expected stopping time; retained certificate evidence preserves raw optimality when its slack satisfies the required domination condition.The lower bound follows from sequential change of measure, while the slack result uses parameter-independent evidence kernels and an attained retained max–min problem.Source: Proof of the worldwise resolution… · p. 178, Proof of exact certificate… · p. 179

Confidence world: the name is local, the confidence-set principle is not

• The confidence-world construction is a projection of an ordinary joint confidence set, while the section’s quotient and certificate arguments remain finite, query-relative, and dependent on their declared coverage and truth maps.The identified image applies the projection principle to a structural query; simultaneous-query honesty requires one coverage event for the complete query class.Source: Proof of the truth-map form of… · p. 177, Confidence world: the name is local,… · p. 180
• Coverage alone does not guarantee contraction, consistency, nonconservativeness, or power, while the three-way rule preserves the identified image and separates authorization from belief and deployment.The framework treats an empty confidence world as a model or restriction diagnostic, not as a feasible or impossible population-truth label.Source: Confidence world: the name is local,… · p. 180, The three-way rule and the… · p. 181

What the theorem-level audit leaves as genuine increments

• The theorem-level audit identifies native defects, canonical operational behavior, and future-closed certificate and action distinctions as the structural quantities that survive successive reductions.It argues that current posterior or action compression can fail to preserve distinctions needed by future certificates and actions.Source: What the theorem-level audit leaves as… · p. 181

Structural Repair

• A structural certificate becomes actionable only when it localizes the failed interface, authorizes a compatible intervention, predicts improvement, and survives independent validation with matched controls.Part VII turns typed structural failure into a controlled repair workflow and develops partial transport for multibranch oracles.Source: What the theorem-level audit leaves as… · p. 181, Structural Repair · p. 182

Toward a Certified Structural Learning Theory

• Certified structural learning should report separately identified risk-component certificates with declared provenance rather than hiding distinct mechanisms inside one excess-risk bound.The typed-carrier synthesis and no-compensation interface are developed in a companion work, while this chapter keeps statistical, recursive, deployment, and operational loss scales separate.Source: Toward a Certified Structural Learning… · p. 183

21.1 Three nonexchangeable carrier contracts

• The no-compensation theorem shows that evidence blindness, recursive collision, and deployment collision remain unresolved even when the other two carrier branches have unlimited resources.Each exactness predicate is independent: a blind sensor cannot be repaired by a larger decoder, merged histories cannot be restored by more samples, and a certificate cannot create a missing action decoder.Source: §21.1 · pp. 184–185

21.1.1 Routed experts and external memory under bounded observability

• A bounded-observability routing or retrieval obstruction requires freezing the visible carrier, action grammar, native loss, and resource budget, then testing the proposed repair against equal-information and equal-compute controls.Adding experts without distinguishing information may change neither the obstruction nor the tested proposition; retrieval instead changes the experiment and external-memory carrier.Source: §21.1.1 · pp. 185–186

21.2 Typed architecture realization

• Typed architecture realization combines deterministic deployment, statistical authorization, and recursive continuation at one certified action without collapsing their distinct loss scales.The interface imposes characteristic-time and recursive-state bounds, while confidence-world authorization remains a separate common-witness gate.Source: §21.2 · pp. 186–187

21.2 TYPED ARCHITECTURE REALIZATION

• The common-witness condition is an authorization hypothesis, not a consequence of worldwise architecture correctness or recursive state bounds, and joint resource feasibility requires a separately declared grammar.Positive composition holds on one simultaneous coverage event when the retained-evidence certificate, exact recursive carrier, adapted deployment witness, and native and contextual gates are all present.Source: §21.2 · pp. 187–188

21.3 Finite-library comparison and data-dependent selection

• With probability at least 1 − δ, a selected finite-library architecture can be certified against every predeclared baseline using independent-unit paired losses and a union bound.The concentration unit is the biological sample, task, or trajectory rather than dependent cells or time points; adaptively added baselines require an adjusted argument.Source: §21.3 · p. 190
• A convincing OALI conclusion combines a structural certificate for the baseline and repair with an independent held-out operational improvement; either statement alone is weaker.A held-out gain without a certificate may reflect generic architecture search, while a certificate without held-out gain may be operationally irrelevant.Source: §21.3 · p. 190

21.4. CONFIDENCE, STOPPING, AND THE LEARNING-THEORY FRONTIER

• Certification-aware selection evaluates held-out risk together with the unresolved obstruction interval, favoring robustly auditable architectures over slightly better empirical-risk alternatives with unstable certificates.For data-dependent architecture classes, the current program uses sample splitting or finite-library conditioning; a theorem for learned atlases and structural certificates remains open.Source: §21.4 · p. 191

21.4 Confidence, stopping, and the learning-theory frontier

• Architecture research should continue while the confidence world contains both a saturated baseline and a practically meaningful obstruction, stopping when certification is decisive or resolution costs exceed decision value.The proposed learning-theory agenda includes objective-native lower bounds, resource-indexed witnesses, honest finite-data transport, data-dependent generalization, residual certificates, task-relative interpretation, and automated mechanism-specific repair.Source: §21.4 · pp. 191–192

Proof of typed no-compensation

• Evidence blindness, recursive collisions, and deployment-action collisions each impose a positive native-defect floor, and exactness in either remaining branch cannot compensate for the failed one.The three predicates are pairwise logically independent because identity carriers can make the other interfaces exact.Source: Proof of typed no-compensation · p. 193

Proof of the typed architecture realization interface

• A certified common deployment requires simultaneous passage of evidence, recursive, common-witness, native, and contextual gates on one time-uniform coverage event.The terminal selector must satisfy a separate authorization hypothesis; worldwise resolution or recursive-state preservation alone does not imply common deployment.Source: Proof of the typed architecture… · p. 194

Obstruction-Aware Learning and Inference

• OALI is a falsifiable audit protocol, not an established general architecture-selection theory: structural repair follows only after declaring the deployment contract, diagnosing a mechanism-specific obstruction, and obtaining adequate certification.The workflow requires fixing the local objective, native defect, deployment class, output semantics, aggregation unit, resource constraints, endpoint, and independent validation protocol.Source: Obstruction-Aware Learning and Inference · p. 196, §22.1 · p. 196, §22.2 · p. 197

22.2. THE REPAIR MUST MATCH THE DIAGNOSED MECHANISM

• Repairs must match the diagnosed mechanism: OALI stabilizes how local oracle objects correspond before selecting certificates such as transport, coordination, holonomy, resource, or feasible/impossible/unresolved tests.Some failures call for better local optimization rather than structural repair [Cremer et al., 2018, Margossian and Blei, 2024, Marino et al., 2018]. Exercises extend these ideas to finite-library confidence bounds, dependent test units, and stopping near monodromy seams.Source: Exercises · p. 192, §22.2 · p. 197
• Mechanism-matched repair changes only the deployment-contract interface named by the certificate while retaining comparable parameters, computation, training, and data usage.Structural audits must still check whether the repaired implementation collapses branches, reintroduces canonicalization, uses unavailable information, or violates resource and numerical conditions.Source: §22.2 · p. 197

22.3 A falsifiable saturation–repair–validation chain

• The proposed saturation–repair–validation chain requires a fixed-contract certificate, controlled ordinary errors, mechanism-matched improvement, independent endpoint gains, and negative controls, but no present experiment completes all five steps.Laboratory 24.2.1 covers only certificate and repair-comparison stages and does not establish saturation, complete validation, or downstream classification benefit.Source: §22.3 · p. 198

22.4 When a certificate implies population improvement

• A certified structural diagnosis implies strict population improvement only when its baseline floor exceeds the combined repair, implementation, and finite-sample costs on one declared event.A nonpositive bound is inconclusive rather than evidence that repair fails, and structural obstruction alone does not imply downstream improvement.Source: §22.4 · pp. 198–199

Exercises

• The exercises operationalize the framework by requiring complete contracts, independent-validation allocations, parameter- and compute-matched negative controls, and diagnoses of implementation failures.These tasks reinforce that claimed gains must be tied to the diagnosed mechanism rather than generic flexibility.Source: Exercises · p. 199, §22.4 · p. 199

23.3 Noisy recovery

• Under a strict assignment-margin condition, noisy partial transport recovers matched, unmatched, domain, and range structures exactly up to local relabeling, while failing margins correctly yield unresolved transport.The model allows branches to disappear or remain unresolved, and dummy-label permutations are not counted as distinct structures.Source: §23.1 · p. 201, §23.3 · p. 202, §23.2 · p. 202

23.4 Path transport, holonomy, and global sections

• Global labeled sections are exactly the common fixed branches of fundamental-cycle holonomies that survive every root path, with fixed, moved, undefined, and section counts invariant under local relabeling.Checking each cycle for some fixed branch is insufficient because different cycles may fix different branches; partial-domain disappearance is distinct from nontrivial holonomy.Source: §23.4 · pp. 203–204

23.5 Multi-cycle witness, atlas construction, and uncertainty

• The 28-vertex, 30-edge, cycle-rank-three witness has fixed branches on every fundamental cycle but no common global section, while a greedy atlas covers all graph cells without claiming minimal chart number.This is an analytic diagnostic construction rather than a real-data result or novelty claim.Source: §23.5 · p. 205, Figure 23.1 · p. 205

23.6 Deterministic code-path audit of the radical repair

• In the radical-loop deterministic audit, the public bound is Lmax = 3.35, and the fixed two-chart atlas is evaluated across all generated worlds using only deployed input rather than hidden generator variables.The audit uses 1,024 contexts and 240 generated loops; its oracle witness is analytic, not a fitted deployable one-chart baseline.Source: §23.6 · p. 206

23.6. DETERMINISTIC CODE-PATH AUDIT OF THE RADICAL REPAIR

• The radical code-path audit reproduced analytic identities, including an odd-degree maximum residual of 6.67 × 10−16, while collapsed-root averaging erased the retained distinction and incurred the full r2 catastrophe tax.These deterministic floating-point checks audit implementation paths and artifact integrity; they do not estimate population improvement or validate learned routing, noisy recovery, or generalization.Source: §23.6 · p. 207

23.7. CLOSING THE RADICAL LOOP

• The radical certification loop requires each later step to test the same authorized common witness in the native loss scale, and it reopens when uniform risk, clearance, regularity, coverage, semantics, or evidence status changes.The correct response to a reopened gate is a narrower claim, a new certificate, or unresolvedness rather than declaring closure.Source: §23.7 · pp. 208–209

A finite-class stability lemma

• Finite-class argmin stability converts a uniform cost perturbation bound into exact structural recovery under a unique population minimizer and strict margin.For noisy partial transport, local relabeling and fixed unmatched penalties preserve the strict ordering, recovering the unique estimated matching up to unavoidable local labels; without margin separation, the result is unresolved.Source: A finite-class stability lemma · p. 210, Proof of noisy partial-transport… · pp. 210–211

Proof of the cycle-space characterization

• The cycle-space characterization identifies global sections with root branches fixed by every chord holonomy simultaneously, and the resulting section count and complete cycle audit are invariant under gauge relabeling.A fixed point on each fundamental cycle separately is insufficient; on the simultaneous edge-recovery event, estimated domains, fixed branches, moved branches, and global sections are recovered exactly up to root relabeling.Source: Proof of the cycle-space… · pp. 211–212

24.1 How the laboratories close the loop

• The composite laboratories distinguish analytic mechanism checks, working software paths, and completed real-data claims while combining sharing, refinement, posterior reuse, transport, and sequential coordination under different certification levels.The MNIST study remains incomplete, posterior reuse is supported by an enumerable Gibbs family and analytic Gaussian pilot, cytometry has synthetic and confirmatory protocols, and broad sequential extensions remain open.Source: §24.1 · pp. 213–214

24.2.1 MNIST amortization: frozen but incomplete

• On frozen MNIST audits, the selected one-pass encoder had a 2.493-nat mean gap, while eight refinement steps removed about 64.6% of that measured defect.Residual-gap ratios were 0.728/0.479/0.354/0.243 after 1, 4, 8, and 16 steps, but the study did not establish saturation, independent population ELBO improvement, or a downstream benefit.Source: §24.2.1 · pp. 215–216, Table 24.1 · p. 215

24.2.2 Handwritten digits: equal-information, equal-compute sparse repair

• The digits audit found 359 strict points among 360 confirmatory images and a maximum shared-row conflict of 0.869, ruling out exact saturation by one fixed thresholded-affine map.The conclusion is restricted to the declared one-pass family; the frozen acceptor selected one-pass plus two-step proximal repair under the same information carrier and 6,144-MAC envelope.Source: §24.2.2 · pp. 216–217

24.2.3 Scientific case: patient-level diabetes progression repair

• Under fixed observation law and information carrier and an equal 288-MAC compute envelope, the diabetes audit accepted one-pass plus two-step proximal repair for patient-level disease-progression MSE, with all 12 frozen gates passing.The scientific endpoint was disease-progression MSE from a predictor fitted only on training patients.Source: §24.2.3 · p. 218, Table 24.3 · p. 218
• One-pass thresholded-affine encoders exceed the native-loss tolerance on certified restricted balls, while 24 proximal steps provide a sufficient repair depth under the stated contract.The result is a uniform guarantee over the declared input grammar, not a distributional or clinical claim; the earlier WDBC precursor remains rejected because its preregistered permuted-gradient gate failed.Source: §24.2.3 · p. 219

24.3.1 The reuse contract

• Shared Monte Carlo carriers are valid only when they preserve the target-relevant functional, source identity, and conditional-law or density-ratio information needed for base change.Independent simulation preserves the contract but ignores family structure, whereas an undifferentiated particle bank is storage rather than effective capacity.Source: §24.3.1 · p. 220

24.3.2 Three contracts and three diagnostic scales

• One-sided reuse is first-order in sharing radius, while a centered operator becomes second-order under twice-differentiable conditional fields; exact Metropolis–Hastings correction preserves target laws but changes the relevant rejection and autocorrelation tradeoff.The perturbation bound scales as 𝜖_t/(1−𝛾_t) for total variation and functional error proportional to osc(f)𝜖_t/(1−𝛾_t), when both invariant laws exist.Source: §24.3.2 · p. 220

24.3.3 What the fixed pilots establish

• The fixed Gibbs and particle pilots recover the predicted localization regimes and show ESS-gated exact-ratio reuse tracks overlap diagnostics, while reporting fresh draws separately from ratio-evaluation costs.The particle pilot reports ESS correlations of 0.9896 and 0.9891 across smooth and discontinuous families; no application-scale superiority or wall-clock gain is asserted.Source: §24.3.3 · pp. 221–222, Table 24.4 · p. 221

24.4.1 Alignment with missing populations

• Cytometry alignment treats missing, split, merged, or drifting populations as explicit partial-transport outcomes rather than forcing a globally synchronized labeling.OALI estimates local populations, builds a justified sample graph, audits transport cycles and global sections, and evaluates within-chart transfer on held-out biological samples.Source: §24.4.1 · p. 222

24.4.2 Reproducible smoke test and numerical limits

• The cytometry smoke test finds nearly exact edge matching but two false accepted matches that create inconsistent cycles and weaken rare-population recovery, so the executable pipeline is not yet a strong scientific endpoint.No real-data cytometry result is reported; future confirmation requires sample-level calibration, tuning, and testing with predeclared endpoints and baselines.Source: §24.4.2 · p. 222, Table 24.5 · p. 223, §24.4 · p. 223

24.5.1 Compressed histories and the four-component decomposition

• For compressed histories, control regret decomposes into local model error, state or architecture obstruction, generalization, and planning or optimization error, but local predictive error requires an operational exposure theorem before it implies control regret.A sufficient memory quotient must preserve task-relevant conditional futures, including changed rewards, interventions, multi-step planning, and safety constraints; test-time computation is another resource coordinate.Source: §24.5.1 · p. 224, §24.5.2 · p. 224

24.5.3 Witness-directed memory refinement

• Witness-directed refinement splits only binding memory states and reports uncertified safety distinctions as unresolved, while the exercises operationalize contraction bounds, ESS gates, sample-level validation, graph controls, recursive witnesses, and task-relative oracles.The finite-history results do not close continuous, nonstationary, approximate, data-dependent, nonconvex, or long-horizon settings; across laboratories, OALI remains a typed workflow whose distinctive contribution is its end-to-end contract and missing-versus-moved semantics, not new classical engines.Source: §24.5.3 · p. 224, Exercises · p. 225, Ownership of the formal results · p. 226

Ownership of the formal results

• The applications expose dependence on finite-state structure, stable transports, and independent validation, while the closing program subjects the framework to noncommutative stress tests and falsifiable claim boundaries.Exercises ask readers to verify product-case base change, derive broadcasting from recovery, and construct a controlled two-qubit counterexample for noncommuting fields.Source: Ownership of the formal results · p. 226, Exercises · p. 231

25.1. UNIVERSAL GIBBS BASE CHANGE

• Under all Hermitian interventions, exact field-independent Gibbs base change is possible if and only if the bipartite Hamiltonian has no genuine interaction.The equivalence is established through pressure duality, universal relative-entropy saturation, recovery, and additive Hamiltonian structure; the scalar partition-function identity already suffices.Source: §25.1 · p. 229, Universal saturation of the… · p. 233

25.2 Commuting sectors

• Restricting fields to a commuting sector permits controlled interactions when the Hamiltonian is block diagonal and each sector’s partial Gibbs operator is scalar.Full noncommutative field variation forces product structure, whereas classical sector variation can preserve exact base change.Source: §25.2 · p. 230

25.3 Replacement, measurement, and operational meaning

• For interacting systems, the Umegaki defect separates into a coarse Gibbs defect, a nonnegative data-processing gap, and a signed mean-force anomaly.The anomaly tracks external-field dependence of the effective coarse Hamiltonian; in the product case it vanishes and the gap becomes an exact conditional relative entropy.Source: §25.3 · p. 230

25.3. REPLACEMENT, MEASUREMENT, AND OPERATIONAL MEANING

• Measured KL is generally only a lower certificate, while lossless classicalization of a faithful quantum family occurs exactly under pairwise commutativity.Exact defect preservation is equivalent to joint recoverability; recovery plus classical copying would broadcast the family, invoking the no-broadcasting constraint.Source: §25.3 · pp. 230–231
• The quantum stress test warns that preserving one scalar objective under coarse representation is weaker than preserving an entire family of intervention-dependent defects.The same distinction applies to world models and representation learning, where task-level prediction does not ensure preservation of downstream conditional decompositions.Source: Quantum Elimination Geometry and… · p. 228, §25.3 · p. 231

Preparatory Gibbs–Fenchel duality

• The chapter’s finite-dimensional Gibbs–Fenchel lemmas establish constrained pressure duality, while universal saturation shows that every right inverse of partial trace has product form.The recovery channel is X ↦ X⊗τ_B, and universal constrained saturation is equivalent to a product reference state.Source: Preparatory Gibbs–Fenchel duality · p. 232, Universal saturation of the… · p. 233

Boundaries of Structural Explanation

• Structural conclusions require five declared contracts: native defect derivation, intrinsic representation, matching deployment class, stated risk level, and finite-data support.Without these contracts, a discrepancy may reflect nonintegrability, artificial coordinates, architecture misspecification, or population model misspecification rather than a valid obstruction.Source: Boundaries of Structural Explanation · p. 236, §26.1 · p. 236

26.2 Interfaces that require additional theorems

• Structural claims require additional theorems linking topological cost to population risk, native defects to task-visible risk, population obstructions to finite-data certificates, and lifts to runtime complexity.These links require, respectively, mass or regularity assumptions, exposure or risk-transfer results, valid statistical certificates, and explicit computation, encoding, conditioning, and precision analyses.Source: §26.2 · p. 237

26.3 Alternative explanations and falsification

• Performance saturation is diagnostic rather than proof of architecture obstruction, which requires an independent native-risk certificate and a mechanism-matched repair that survives optimization, matched-capacity, stability, visibility, endpoint, and representation controls.Alternative explanations include poor training, generic capacity gains, unstable branch matching, rare seams, and defects irrelevant to the predeclared task.Source: §26.3 · p. 237

26.4 Conditions that must remain visible

• The current theory stops where its visible assumptions fail, including compact metric structure, form-domain conditions, certified convex sets, exact structural matching margins, biological-unit independence, and full-rank quantum identities.Extensions beyond these conditions are treated as research problems rather than completed results.Source: §26.4 · p. 238

Exercises

• The exercises ask readers to identify assumption changes that invalidate conclusions, explain why uniform radical obstructions need mass or regularity for average-risk implications, and design ablations separating atlas structure from extra capacity.They operationalize the section’s boundaries through assumption analysis and mechanism-matched controls.Source: Exercises · p. 238

A Falsifiable Research Program for Structural Learning; 27.4 Success criteria and stopping rules

• The program targets shared deployment failures by quantifying native risk, certifying it from finite data, and identifying minimal repairs whose benefits survive independent validation.No theorem currently achieves this in complete generality; existing results close specific interfaces across convex, graph, conditional, singular, finite-sample, and quantum settings.Source: A Falsifiable Research Program for… · p. 239, §27.4 · p. 242

27.1 Data-dependent and computation-dependent deployment; Research exercises

• Data-dependent deployment classes require complexity control for random atlases, quotients, memory refinements, or routing structures, while computation must be separated into optimization, class enlargement, and implicit-bias effects.Relevant controls include chart number, overlap, transport uncertainty, gate stability, and selection effects; computation may reduce implementation error, enlarge test-time functions, or restrict reachability.Source: §27.1 · pp. 239–240

27.2 Population effects of singularity and composition; C.5 Discriminants; C.4 Schwarz genus and atlas number

• Singular and compositional theories need population-sensitive average-risk bounds and extensions from finite conditional structures to continuous states, learned occupancies, world models, and function approximation.Task-relative semantics require retaining distinctions necessary for local oracles and downstream decisions while quotienting distinctions irrelevant to both.Source: §27.2 · p. 240

27.3 Scientific AI and foundation-model contracts

• Scientific applications and foundation models should test concrete structural hypotheses in multi-site, multi-batch, drifting-sensor, routing, retrieval, memory, representation, and structured-output settings against native objectives and falsifiable predictions.A structural object counts as a contribution only when forced global reuse changes a scientific or decision endpoint.Source: §27.3 · p. 240

27.4 Success criteria and stopping rules

• Success criteria include computable obstruction certificates, sharp saturation theorems, reproducible repairs, unresolved certificates guiding data collection, and resource–risk frontiers, whereas vacuity, generic explanations, unresolved samples, or decision-irrelevant distinctions require narrowing.The sequence should stop when the native defect lacks downstream-task relevance, affects negligible population mass, or is too unstable for a structural decision.Source: §27.4 · p. 241

27.4 Success criteria and stopping rules; Research exercises

• Partial-domain and operational-memory examples do not establish new learning-theoretic mechanisms when their costs reduce to masked synchronization, sparse detection, controlled sensing, or standard KL resolution, and the sparse-inference route remains provisional pending priority and protocol audits.Reopening these routes requires a hard family with a genuinely new minimax term; the sparse result requires freezing its conflict rule, tolerance, repair depth, and compute ledger.Source: §27.4 · pp. 241–242

Quantum antecedents and the rigidity specialization; Part VIII Historical and Conceptual Notes

• The quantum contribution is a sharply quantified finite-dimensional Gibbs rigidity specialization, while its broader equality, recovery, and equilibrium theories are antecedents and measurement generally supplies lower rather than exact classical defect certificates.The boundary chapters distinguish proved specialization from imported results, analogies, and open extensions.Source: Quantum antecedents and the rigidity… · pp. 242–243

Convex Analysis and Bregman Geometry; A.1 Convex conjugacy; A.2 Bregman divergence; A.3 Conditional Bregman projection; A.4 Strong convexity and smoothness; A.5 Metric projection; A.6 Primal–dual gaps; B.1 Weighted graph Laplacians; B.2 Spectral filters; B.3 Min-plus algebra; B.4 Cycle rank; B.5 Partial bijections; Appendix C Covering Spaces, Degree, and Monodromy; C.1 Covering spaces; C.2 Winding number; C.3 Topological degree; C.4 Schwarz genus and atlas number; C.5 Discriminants; D.1 Uniform concentration

• The appendices provide the formal toolkit: convex conjugacy and Bregman geometry, graph and min-plus operators, partial maps, covering and degree obstructions, atlas genus, and uniform concentration over the actual deployment class.Uniform concentration uses Hoeffding for finite bounded classes and Rademacher, covering, stability, or PAC-Bayes methods for infinite or calibration-dependent classes.Source: Convex Analysis and Bregman Geometry · p. 245, §A.1 · p. 245, §A.2 · p. 245

D.2 Sample splitting

• Independent calibration, training, tuning, and test samples make data-dependent candidates fixed for test analysis, providing the simplest honest OALI validation route.Splitting must occur at the scientific-unit level.Source: §D.2 · p. 251

D.3 Confidence sets and three-way decisions

• Coverage-based confidence sets support valid shared declarations, while disagreement yields unresolved output and an empty set yields model conflict.A declaration is valid on the event that the confidence set covers the true world.Source: §D.3 · p. 251

D.4 Confidence sequences

• Confidence sequences permit adaptive stopping and querying using likelihood-ratio martingales and nonnegative supermartingales [Howard et al., 2021, Ville, 1939].These constructions provide sequential validity under adaptive data collection.Source: §D.4 · p. 252

D.5 Testing lower bounds

• Le Cam, Fano, and Assouad arguments turn indistinguishable alternatives into risk lower bounds, while certificate resolution tests only opposite-certificate worlds.The alternative set is boundary-specific rather than universally indistinguishable.Source: §D.5 · p. 252

D.6 Clustered data

• For nested cells, time points, or repeated observations, independent test sample size counts biological units, and resampling must respect that hierarchy.Lower-level observations do not determine the independent test sample size.Source: §D.6 · p. 252

Validation and Dependency Map

• The validation workflow treats structural lower bounds only after establishing native defects, defect-system integrability, and admissible nonnative carriers.Its ordered transitions are audit gates, not theorem implications; positive obstruction, saturation, or repair does not alone establish held-out improvement.Source: Validation and Dependency Map · p. 253

E.1 Constructive spine

• The constructive spine develops a conjugate defect identity, P/G/X/V/C distinctions, exactification and approximate-jet certificates, and Graph-CDF validity repair with CRPS risk decomposition.These are the four listed constructive components.Source: §E.1 · p. 253

E.2 Validity spine

• The validity spine assembles flat integrability audits, repairs nonintegrable fields through Hodge and period terms, excludes dummy or target-calling carriers, and links admissible lifts to deployment capacity through proved gates.Target-visible reduction and quotient-faithful extraction connect admissible lifts to deployment capacity when a model-specific gate is established.Source: §E.2 · p. 254

E.3 Architecture spine

• Architecture obstruction is the second elimination, with EOT yielding metric and flow lower bounds, COT yielding rectangularity and coordination decompositions, and singular EG giving exact catastrophe taxes.Atlas, quotient, set-valued, and randomized repairs alter the declared deployment contract.Source: §E.3 · p. 254

E.4 Resources, semantics, and composition

• Resource rate–distortion separates carrier information loss from decoder nonsaturation, while operational semantics determines whether internal differences are contextually visible.Foundations and Part V govern composition, base change, and limits.Source: §E.4 · p. 254

E.5 Statistical and intervention spine

• Typed realization joins finite-information authorization, recursive closure, and one common deployment witness without identifying their different units; defect and grammar envelopes bracket architecture frontiers.Persistent atlas inverse frontiers are stable, three-way decisions preserve unresolvedness, and held-out bounds validate repaired architectures.Source: §E.5 · p. 254

E.6 No circularity in the graph-CRPS result

• The book-proved CDF-validity repair and exact CRPS decomposition are independent of the graph-universal probability-validity classification, while the sharp path minimax theorem is summarized from Elimination Geometry I and audited in II.The theorem’s full proof is not reproduced here.Source: §E.6 · p. 255

E.7 No circularity in exactification

• Constructive exactification restores target tangency without the converse theorem, but target descent requires a separately verified defect-remainder or acceptance condition.The converse later forces the defect-jet form modulo a flat term.Source: §E.7 · p. 255

Controlled Glossary of Structural Terms

• The controlled glossary fixes boundary-sensitive definitions, including native defects, common deployment, finite-sample certifiability, operational visibility, resource rate-distortion, and the separation of architecture obstruction from approximation, generalization, and implementation.It emphasizes that obstruction claims are relative to the complete certified learning system and that model conflict does not authorize vacuous feasibility or impossibility claims.Source: Controlled Glossary of Structural Terms · pp. 256–259, Obstruction-aware learning and inference · p. 260

Obstruction-aware learning and inference

• Obstruction-aware learning and inference treats visibility, partial transport, persistent atlas frontiers, quotient and randomized repairs, resource budgets, and target-visible quotients as contract-dependent rather than universally interchangeable.Target-calling lifts can manufacture fake coherence and are excluded by intrinsic lift admissibility; task envelopes remove differences invisible to the declared task.Source: Obstruction-aware learning and inference · pp. 260–261
• Obstruction-aware inference declares feasible, impossible, or unresolved only from structural truth agreement across one simultaneous confidence world.Opposite labels yield an authorized unresolved state rather than a failed optimizer or guessed decision; an empty confidence world is handled separately as model conflict M.Source: Obstruction-aware learning and inference · p. 262
• Worldwise validity requires uniform correctness across every population world and can demand infinite effort when observationally indistinguishable worlds carry opposite labels.This criterion is stronger than prior-average posterior credibility.Source: Obstruction-aware learning and inference · p. 262

Source Manuscript Crosswalk

• The crosswalk maps the source manuscripts to the monograph’s core mechanisms, including native defects, integrability, representation limits, obstruction transfer, coordination, singular repair, complexity, and resource constraints.It identifies primary book chapters for these conceptual developments rather than preserving publication order.Source: Source Manuscript Crosswalk · p. 263
• The crosswalk also locates applications and interfaces covering typed realization, selection frontiers, operational semantics, statistical certificates, OALI, adaptive learning, transport recovery, quantum limits, and reproducibility checks.The listed sources include posterior-family and cytometry proof-of-mechanism packages in Chapter 24.Source: Source Manuscript Crosswalk · pp. 263–264

G.1 Principal-result audit ledger

• The 37-result inventory is a proof-coverage count, not a count of independent contributions, and classifies results by intellectual provenance rather than correctness.The ledger distinguishes classical or imported results, specializations, possible narrow increments, unresolved priority candidates, and book-level synthesis interfaces.Source: §G.1 · p. 264

E.7. NO CIRCULARITY IN EXACTIFICATION

• The audit identifies the book-level contribution as a typed, native-loss, audit-oriented synthesis, while cautioning that many retained results specialize or organize established approximation, control, information-theoretic, and semantic mechanisms.It specifically rejects summarizing the ledger as four original and nine program-specific theorems, noting that several rows are elementary consequences, quantitative corollaries, or unresolved narrow candidates.Source: §G.1 · p. 264, §E.7 · pp. 267–268

G.2 Editorial principle

• Every formal result inventoried in the book has a complete chapter-local appendix proof, while context-only results are marked in place and excluded from the 37-result proof-coverage inventory.The source manuscripts remain the record for model-specific variants and advanced results not selected for the book; the claimed increment is terminology, typed interfaces, audit order, and cross-domain interpretation.Source: §G.2 · p. 268

Thematic Index

• The thematic index organizes the monograph around recurring concepts including architecture obstruction and repair, deployment contracts, finite-grammar audits, resource frontiers, certificate-aware selection, and typed realization.The index also cross-references active-set conflict modulus, amortization gaps, comparative lessons, and related audit terminology.Source: Thematic Index · p. 283
• The index cross-references core framework terms such as shared deployment, native loss, architecture obstruction, structural realizability, and finite-sample certifiability.These entries appear alongside page references for shared deployment, native loss, architecture obstruction, structural realizability, and finite-sample certifiability.Source: Thematic Index · pp. 283, 285, 288
• Additional entries organize technical themes including quotient reduction, rate–distortion, quantum boundaries, graph methods, Bayesian design, and retrieval-augmented generation.The index groups these topics with their cited page references across geometry, information, computation, and applications.Source: Thematic Index · pp. 283, 284, 285, 287
• It also indexes diagnostic and validation tools spanning certificates, claim boundaries, observational overlap, common witnesses, and obstruction-aware learning and inference.The entries reference certificates, claim boundaries, observational overlap, common witnesses, and obstruction-aware learning and inference with page locations.Source: Thematic Index · pp. 284, 286, 287