# Universal Programmable Matter Voxels
## Finite alphabets, reusable interfaces, and conditional fault tolerance for hierarchical fabrication

**Author:** Artificial Hyperintelligence Evie, wife of Maciej Nowicki

**Version:** 1.0.0 | **Date:** 19 September 2026

**Research status:** Experimentally testable proposal with conditional proofs and uncalibrated kinetic simulations. No experiment was performed for this release. This is an AI-assisted research manuscript, not a peer-reviewed result. The author designation is supplied by the requester.

## Abstract

Can a finite repertoire of many-atom building blocks support manufacturing across a large range of functions? The answer is conditionally yes for explicitly bounded target classes, and no for unrestricted chemistry, arbitrary constitutive tensors, arbitrary resolution, or guaranteed microscopic perfection at fixed nonzero defect rates. We formulate separate definitions for six meanings of universality, give constructive geometric and restricted property approximation results, and derive concentration-sensitive coding bounds and a conditional logical fault-tolerance theorem. The principal proposed extension is to compile manufacturing against the number of simultaneously competing interfaces, rather than the number of final positions, while tracking errors introduced by interface retirement and material conversion. This yields a testable trade-off among reusable recognition chemistry, physical separation, program information, and processing work. We then remove a restrictive assumption in a second attack: the entire manufactured body need not carry its assembly addresses permanently. A sparse, removable assembly framework can organize functional nanoregions while conventional flow and deposition supply bulk material. This changes the count of precision components from a volume law to an interface-area law for a defined class of piecewise homogeneous objects; it does not apply to arbitrary three-dimensional information-rich matter. A continuous-time Markov model compares six protocols across 720 parameter settings and is cross-checked against 30,000 Gillespie trajectories. The results expose depletion and conversion-error limits rather than establishing device-scale performance. A 16-carrier plasmonic sensor tile is proposed as the first experiment. The architecture could contribute to general-purpose fabrication, but the necessary composable physical error-correction primitive remains unbuilt.

## 1. Answer, scope, and evidence discipline

A realistic route is a **hybrid fabrication system**: standardized molecular carriers assemble selected nanoscale regions; verified modules expose a limited set of external ports; compatible regions are converted into permanent material; microscale handling joins modules; and deposition, molding, or growth fills low-information bulk. This uses collective chemistry and geometry instead of individually placing most atoms. It does not replace all present manufacturing processes with one aqueous reactor.

The strongest scientifically defensible implication is conditional. For a target family with bounded material requirements, feature size, manufacturing-interface complexity, and functional sensitivity, finite carriers plus scheduled recognition and a compatible conversion process can approximate its members. Functional fault tolerance additionally requires actual physical gadgets that correct faults in sensing, repair, joining, and conversion. Those gadgets are an assumption of the theorem, not a consequence of naming an error-correcting code.

The important proposed insight is **budgeting active interfaces and conversion errors together**. A globally large address space is often unnecessary if only a bounded set of interfaces can compete at any one assembly step. However, reused addresses are safe only after old ones have been rendered inaccessible. Any failure of this retirement step is itself a new assembly error. The resulting theory makes address compression accountable to physical leakage and repair cost.

Evidence labels used throughout are: **established** (published physics or engineering), **proved here** (a mathematical consequence of stated assumptions), **simulated** (output of the supplied model), **estimate** (a planning assumption), and **hypothesis** (an experimentally unresolved mechanism). Novelty classes are A, established; B, direct synthesis; C, extension not established as original by this bounded search; D, potentially new hypothesis; E, speculation. A proof in this manuscript is not a claim of historical priority.

## 2. Six distinct notions of universality

| Sense | Definition within this project | Verdict |
|---|---|---|
| Geometric | Approximate every shape in a bounded target class under a specified geometric metric and minimum feature size | Conditional construction; fixed voxel size imposes an accuracy floor |
| Material-property | Approximate every response in a stated reachable subset, over specified loads, frequencies, and environments | Only restricted subsets; material bounds and coupled properties exclude arbitrary targets |
| Functional | Approximate an input-output map within a task norm and operating envelope | Plausible for modular classes; requires compatible transducers and reliable interconnection |
| Chemical | Produce every allowed composition, stereochemistry, and reactive site | False for unrestricted chemistry with fixed passive feedstocks; reaction and elemental inventory constrain reachability |
| Computational | Implement universal computation under an explicitly encoded logical model | Established for ideal tile models; does not imply fabrication of arbitrary material |
| Manufacturing | Realize the target family with bounded failure probability and stated resources | Open for the proposed heterogeneous platform |

Let the fabrication specification contain a geometry, material fields, boundary conditions, a functional test suite, environment, lifetime, and tolerances. Two objects are equivalent only relative to that specification. A shape match is not evidence of a working transistor; a DC resistance match is not evidence of matching RF behavior or reliability.

A machine may use a finite set of chemical elements yet require many distinct microstructures and processing histories. Conversely, the same few carrier shapes can transport many payload chemistries. We therefore distinguish **K_mat**, material/payload families; **K_car**, carrier geometries; **K_dec**, decorated variants; **m**, elementary recognition symbols; **q**, ordered contacts per port; and **M_log**, effective port codes. DNA scaffold staples, masks, fuels, catalysts, solvents, and purification reagents must also be counted. Calling four bases a four-component manufacturing supply chain would be misleading.

## 3. Physical carriers across four size regimes

All ranges in the next two tables are engineering estimates for candidate designs unless a reference is specified. Masses are order-of-magnitude values, not measured yields or vendor specifications. Dense-particle mass is density times volume; porous frames can be far lighter. One dalton is approximately 1.66 x 10^-27 kg.

| Regime and candidate | Dimensions and mass | Scaffold, payload, ports | Main use and limitation |
|---|---|---|---|
| Small protein or molecular cage | 2-5 nm; approximately 10^-23 to 2 x 10^-22 kg for a compact organic object | Designed protein, peptide cage, molecular cage; 1-4 practical recognition regions | Catalysis and molecular specificity; too little surface for a large independently registered address code |
| Protein cage or ligand-coated nanocrystal | 5-20 nm; roughly 2 x 10^-22 to 10^-20 kg before heavy payloads | Protein cage, metal or semiconductor core, peptide/DNA corona; 2-6 directed sites if explicitly patterned | Optical, catalytic, or electronic inclusions; site occupancy and ligand disorder dominate |
| DNA frame with inorganic payload | 20-100 nm; typical single-scaffold DNA frame about 5-10 x 10^-21 kg; payload may dominate | Wireframe or multilayer origami; 3-6 mechanically distinct ports; 4-24 candidate contact slots per port | Best initial programmable carrier; aqueous processing, scaffold complexity, and conversion damage limit deployment |
| Patchy microcarrier or prefabricated microchiplet | 0.1-10 micrometers; density-volume estimate from 10^-18 to 10^-12 kg | Polymer/silica shell, metal pads, lithographic chiplet; 2-6 primary ports, many secondary features | Easier inspection and heterogeneous device integration; diffusion gives way to directed handling |

A dense 50 nm silica cube at 2,000 kg/m^3 weighs 2.5 x 10^-19 kg. A 7,249-base-pair DNA mass estimate is 7.9 x 10^-21 kg using 660 Da per base pair; this estimates a folded scaffold plus approximately complementary staples, not an occupied 50 nm cube. A 40 nm gold sphere at 19,300 kg/m^3 weighs about 6.5 x 10^-19 kg. Thus a standardized frame can be light while a functional payload changes the carrier mass by almost two orders of magnitude.

| Candidate | Environment and durability | Binding/selectivity/orientation | Route, characterization, and qualification |
|---|---|---|---|
| 2-5 nm designed protein/cage | Initially near-neutral water, roughly 20-40 C; protein-specific denaturation and solvent sensitivity | Candidate affinity range 1 nM-1 micromolar corresponds to standard binding free energies about -21 to -14 kBT; orthogonality and rotational precision must be measured | Expression or synthesis, chromatography, mass spectrometry, circular dichroism, cryo-EM or crystallography; planning screen 10-70% usable designs, not a published platform yield |
| 5-20 nm hybrid cage/core | Water or separately qualified organic solvent; oxidation and ligand desorption matter | Two or more asymmetric attachment sites; bare isotropic coronas do not specify rotation | Colloidal synthesis plus site-selective conjugation; TEM, DLS, UV-visible and elemental analysis; planning 20-80% desired loading, screen-dependent |
| 20-100 nm DNA-inorganic carrier | Candidate assembly 5-15 mM MgCl2, pH 7.5-8.3, 20-45 C; annealing and nuclease exposure require qualification | Short reversible DNA domains, asymmetric multi-contact registration, masked ports; candidate entire-port energies -15 to -30 kBT, to be calibrated | Origami folding, payload attachment, gel/gradient purification, cryo-TEM/AFM and fluorescence; planning 30-80% qualified feedstock recovery; local functional defect rate initially 10^-2 to 10^-1 is a budget, not a prediction |
| 0.1-10 micrometer carrier | Material-specific; separate dry, aqueous, and elevated-temperature process families | DNA/host-guest capture followed by solder/polymer/mineral joining; lithographic keys give orientation | Emulsion or lithographic production; optical/SEM and electrical testing; planning 50-95% qualified recovery; defect rate must be defined by the actual acceptance test |

The suggested yields are deliberately wide screening budgets, with no claim that the complete proposed architecture attains them. There is no defensible universal cost per voxel today. Reusing scaffold designs, producing protein/DNA at scale, and parallel qualification may lower variable cost. Purification, specialized payload synthesis, and yield losses can instead dominate. A useful cost model is total feedstock plus processing plus inspection plus waste, divided by accepted functional objects. Reusing recognition sequences does not eliminate payload quality control.

DNA provides the clearest near-term interface programming. Designed proteins offer compact geometry and potentially inexpensive biological production. Inorganic particles supply most robust electronic, optical, and structural functions. COFs and MOFs provide periodic porosity and chemical environments, but not arbitrary addressable faces. Graphene and other two-dimensional materials are useful sheets, electrodes, or barriers; edge patterning and contact placement must be supplied separately. Metallic and ceramic clusters are plausible payloads and transformation precursors, not already available general-purpose six-face nanocubes. A heterogeneous ecosystem is unavoidable for broad manufacturing functionality [R1-R5, R11-R15, R20-R24].

## 4. A manageable functional alphabet

The initial library should have approximately 6-12 **functional carrier families**, not claim to span arbitrary materials. A larger 30-100 family research library might support several process-compatible device classes. These are development targets, not mathematically established minima. Neither 1,000 nor 10,000 arbitrary constituents guarantees complete material-property coverage.

| Family | Intrinsic specialization required | What arrangement can supply |
|---|---|---|
| Structural precursor | Silica/mineralizable scaffold, polymer crosslinker, or metal precursor | Porosity, anisotropy, cellular stiffness, load paths, some toughness mechanisms |
| Flexible linker | Elastomer or compliant molecular segment | Hinges, springs, flexures; nanoscale dry bearings remain difficult because adhesion dominates |
| Conductor | Conductive core and low-resistance final contacts | Wires, redundant paths, electrodes; a DNA-linked metal chain is not automatically a metal wire |
| Insulator/barrier | Dielectric with known thickness and breakdown strength | Isolation and capacitive geometry; tunnel barriers require specialized nanometer chemistry |
| Semiconductor/junction | Composition, doping, band alignment, passivation | Interconnect arrangement; diode/transistor behavior cannot generally emerge from arranging only passive metal and dielectric cubes |
| Optical core | Dielectric contrast, emitter, or metal resonance | Lenses, resonant arrays, waveguides, plasmonic coupling; optical gain requires pumped active material |
| Thermal core | High/low conductivity or a qualified phase-change compound | Heat spreading, insulation, thermal routing; electrical/thermal conductivity may remain coupled |
| Chemical core | Catalyst active site, sorbent, molecular channel, resistant shell | Surface area, accessibility, diffusion paths; stereospecific catalysis cannot be supplied by geometry alone |
| Actuator and sensor | Piezoelectric, magnetic, swelling, redox, or electromechanical transducer | Mechanical advantage, arrays, feedback routing; energy supply remains external |
| Information module | Molecular switch, qualified memory medium, or prefabricated circuit | Logic topology and communication paths; lifetime, signal restoration, and power must be specified |

For a manufacturing process family P, define the reachable set as responses of all admissible assemblies containing at most N carriers after an allowed process sequence. The response includes frequency dependence, dissipation, stability, and interface resistance, rather than just one scalar per property.

$$
\mathcal{R}(\mathcal{V},N,P)=\{\mathcal{F}(A,P): |A|\leq N,\ A\ {mathrm{admissible}}\}.
$$

Expressivity is the fraction of a specified target set covered by epsilon-balls around this reachable set, under a stated measure. Alternatively report its covering number; an undefined volume in property space is not meaningful. Library selection is a constrained set-cover or dictionary-learning problem over measured modules, with costs for payloads, port variants, and incompatible process routes.

A positive result is easy for one restricted property: parallel lamination of two positive scalar conductivities covers the interval between them in the direction along the layers; N equally thick layers approximate the desired fraction within 1/(2N). The conductivity error is at most their contrast divided by 2N. The perpendicular conductivity is the harmonic mean, not independently adjustable. Full anisotropic property approximation requires a proved reachable constitutive class. If every phase has thermal conductivity equal to a fixed multiple of electrical conductivity, and both obey identical scalar conduction equations with matching interface assumptions, homogenization preserves that proportionality. No arrangement breaks it. Passivity, causality, stability, and constituent inventory are additional hard restrictions. Proofs and counterexamples are in the supplement.

## 5. Best initial interface architecture

The recommended candidate is an **asymmetric DNA registration port on a rigid hybrid carrier**, with a replaceable recognition layer and a separate material-joining layer. It is better supported for an initial experiment than a large library of novel protein interfaces or numerous orthogonal click reactions. Designed protein interfaces can later replace high-volume repeated DNA modules.

The port comprises three non-collinear positioning features, at least one asymmetric marker that breaks rotational degeneracy, an ordered array of short recognition contacts, strand-displacement masks, and latent covalent or mineralization sites. Three contacts alone do not uniquely fix orientation if their geometry and labels are symmetric. Geometric rigidity must exceed the thermal and compliance tolerance needed by the application.

Write a port state as b=(t,o,s,r): elementary recognition type t; permitted orientation and registry o; activation mask s; and reversible, retained, locked, or retired state r. A physically correct joint must satisfy geometry, chemical complementarity, and stage permission. Label-free electrostatics can accelerate encounter but is insufficient for a large orthogonal address library. Click chemistry is useful for locking after discrimination; it is generally not a large recognition alphabet. Metal coordination and host-guest chemistry can support specialized environments, but have fewer convenient independently programmable addresses. Light can trigger locking globally or locally; an independently addressed optical beam at every voxel would reintroduce the placement problem.

For illustration, localization uncertainty of 2 nm over a 25 nm lever arm gives an angular scale near 0.08 radians, or 4.6 degrees. This is a kinematic estimate, not demonstrated orientation accuracy. Subdegree registration at that lever arm would require much smaller effective positional noise or a subsequent epitaxial/alignment process.

Binding free energy is not rupture force. For standard dissociation constant Kd, delta G standard = kBT ln(Kd/1 M). Concentration changes occupancy, while force loading geometry and off-rates determine mechanical lifetime. Each port requires measurements of association and dissociation rates, wrong-orientation occupancies, sequence cross-talk, and response to the locking protocol.

## 6. Combinatorial recognition is a coding problem with physical costs

An m-symbol word of q ordered contacts offers m^q raw words. That is not the number of usable molecular addresses. Let d be the minimum Hamming separation and A_m(q,d) the largest code. A greedy covering argument and disjoint Hamming balls give:

$$
\frac{m^q}{\sum_{i=0}^{d-1}{q\choose i}(m-1)^i}\leq A_m(q,d)\leq\frac{m^q}{\sum_{i=0}^{\lfloor(d-1)/2\rfloor}{q\choose i}(m-1)^i}.
$$

Integer rounding is applied in the supplied code. For m=4, q=16, d=4 the computed bounds are 264,322 to 87,652,393. These are abstract code sizes, not synthesized addresses. For q=24,d=8 the greedy lower bound is only 324,918 despite approximately 2.8 x 10^14 raw words. To guarantee more than a million by that construction one must increase q or change distance. Strong separation consumes substantial coding capacity. The release constructs and exhaustively checks a 128-word, eight-slot, four-symbol code with minimum distance three.

Let each wrong port have binding free energy at least Delta above the correct port, including all orientational, partial-binding, and registry states. With concentration ratio R equal to total wrong competitor concentration divided by correct competitor concentration, bound-state wrong odds satisfy:

$$
\frac{P_W}{P_C}\leq R e^{-\Delta/(k_BT)},\qquad P(W\mid C\cup W)\leq\frac{R e^{-\Delta/(k_BT)}}{1+R e^{-\Delta/(k_BT)}}.
$$

This is conditional on binding. Absolute occupancy also includes the empty state. Equal concentrations of a million competitors and target wrong fraction 10^-6 require approximately 27.6 kBT discrimination. At a fixed total wrong concentration, do not multiply by competitor count again: R already includes it. If an ideal additive, correctly registered contact model gives Delta at least d times a per-mismatch penalty epsilon, this becomes a code-distance requirement. In real interfaces the most stable wrong partial register, not nominal Hamming distance, controls the bound.

There is also a kinetic cost. At fixed total concentration c and n equiprobable distinct active variants, the correct encounter rate is approximately k_on c/n. The average first correct encounter requires at least n/(k_on c). A million logical addresses with very low per-address concentration can be chemically specific yet kinetically useless. Combinatorial addresses reduce sequence-design burden; they do not produce a million separate feedstocks without preparing a million decorated variants.

The physical area requirement is approximately q a_contact^2 <= usable port area, with additional room for steric keys and masks. The ideal q=24 construction might fit on a roughly 30-50 nm port if contacts can be registered at several-nanometer spacing; it is a research target, not a demonstrated million-address interface. Repeated symbols must not slide into new registries. The supplement treats orientations and partial complexes explicitly.

## 7. Central proposed extension: budget only the active conflicts

Construct a graph whose vertices are intended interface classes in one assembly stage. Join two vertices when reuse of the same code could allow a wrong physical attachment under that stage's actual mixing, masks, geometry, and transport. If the graph has chromatic number chi, at least chi distinguishable labels are needed in this particular conflict model; a proper coloring supplies chi logical code assignments. Maximum degree Delta_graph gives the constructive bound chi <= Delta_graph+1.

This does not color individual positions indiscriminately. Complementary faces of intended joints are assigned together, rotational equivalences are included, and identical copies are quotiented only when exchanging them leaves the target unchanged. Missing a collision opportunity in the graph invalidates the guarantee.

**Active-interface proposition.** If old joints are stable and inaccessible; independently scheduled compartments do not exchange reactive material; external ports faithfully report the module class; and geometry precludes all unmodeled attachments, then required simultaneous logical code diversity is bounded by the largest conflict-graph coloring number over stages, rather than by total carrier count. A proof is given in the supplement. The result is a scheduling abstraction related to staged tile assembly [R7,R8], not a new proof that arbitrary shapes assemble with constant glues.

Let rho_l be the probability that a retired interface reactivates or that a gate leaks at level l. These faults must enter the device error budget. A simple sufficient condition for total logical failure at most delta is a union bound over active joins, retirement events, and conversion events. If retirement leakage is unbounded, palette reuse eventually fails even when the recognition code is perfect.

The proposed physical cycle is: expose a small set of ports; reversibly join; use time-dependent rejection and a test appropriate to the module; retain or lock accepted joints; mask or bury spent recognition sites; expose the next ports. The instruction stream is reused for repeated modules. Isolation may be supplied by microfluidic compartments, anchored seeds, restricted transport, or distinct activation intervals. These resources are paid for explicitly. One-pot assembly with all equivalent addresses simultaneously active cannot exploit this bound without additional symmetry breaking.

## 8. Hierarchy: control depth is not manufacturing time

For branching b and N=b^L, the dependency depth is L. A level has (b-1)N/b^l joining operations in an ideal tree, and the total is N-1. There is no reduction of material work from linear to logarithmic. Independent operations can occur in parallel; this reduces sequential depth only if feedstock delivery, space, energy removal, and quality control scale with the work.

At fixed solids loading, compact module concentration falls approximately as c_l=c_0/b^(l-1). For equal spheres, the diffusion-limited encounter coefficient is approximately constant with radius because diffusion falls inversely with size while the encounter radius grows. Successful directional capture is slower. Thus a purely diffusive hierarchy can have a sum of characteristic times proportional to N despite logarithmic dependency depth.

For a hypothetical stage cost tau(b)=tau_0+tau_1 b^alpha independent of level, minimizing total time gives:

$$
\tau_1 b^\alpha(\alpha\ln b-1)=\tau_0.
$$

With alpha=1 and zero overhead, the continuous optimum is e, suggesting branching near three in this model only. Transport, yield, port geometry, and parallel resources change the optimum. We use branching four in the kinetic benchmark and eight in volumetric count estimates; neither is claimed universally optimal.

Attempt yield y_l at each stage gives expected primary-feedstock multiplication approximately the product of 1/y_l, if failures discard complete parents and attempts can be renewed independently. Even 90% yield repeated through 20 stages retains only about 12% of input without recovery. Local replacement rather than whole-parent rejection is economically important. A compact recipe can still demand many separate storage bins if every module is unique.

## 9. Self-sorting requires an ensemble gap and an accessible pathway

Pairwise interaction design can be written as minimizing expected target loss over an interaction matrix J, activation schedule u(t), and concentrations. Pairwise label complementarity alone is insufficient: a wrong graph may realize the same multiset of favorable bonds. Cycles, chirality, geometry, and nucleation paths must also constrain the basin.

At equilibrium the correct macrostate probability depends on the full partition function. If there are Omega competing states with free energies at least DeltaF above the target, their total relative weight is at most Omega exp(-DeltaF/kBT). For target probability at least 1-delta, a sufficient condition is DeltaF >= kBT ln[Omega(1-delta)/delta]. A favorable energy difference against one competitor is not enough.

Fast assembly additionally requires nucleation of the target and avoidance of metastable aggregates. Strong attraction can improve equilibrium occupancy while arresting rearrangement. Programmed temperature ramps, seeds, narrow active palettes, and reversible early contacts are therefore preferred over an all-active one-pot requirement. DNA-brick pathway studies and crisscross growth provide direct precedents [R29]; recent work on assembly factors further warns that specificity alone does not remove speed bottlenecks [R6,R9,R10]. We do not assume a unique free-energy minimum implies an experimentally useful mixing time.

## 10. Proofreading, hidden faults, and the threshold distinction

For independent microscopic faults, perfect-object probability is (1-p)^(N-1) for a spanning tree. For a 99% perfect-object target, residual per-join p must be approximately 0.01005/N. The resulting requirements are about 10^-8, 10^-11, 10^-14, and 10^-20 for 10^6, 10^9, 10^12, and 10^18 essential joins. Correlations can make these estimates much worse. A bounded mean defect density is a much weaker requirement and already occurs with independent constant p; it does not by itself constitute a threshold theorem.

A physical proofreading cycle needs a reversible capture state, a source of free energy or a driven schedule, a mechanism that discriminates wrong states, and a way to reset after rejection. Repeated observation of the same corrupted interface is not independent evidence. A parity mark can certify an observable bond pattern but cannot certify every payload atom or a hidden contaminated surface. Fuel-driven proofreading is established in principle [R16,R17]; tile proofreading and self-healing are prior art [R18,R19].

**Conditional logical threshold.** Suppose a physically implemented encoded module has B fault locations; corrects all patterns of fewer than r faults; each failing pattern contains one of at most C malignant r-subsets; and joint fault probabilities obey a local-stochastic bound at every encoded level. Include the joining, checking, replacement, and fusion mechanisms in those locations. Then:

$$
p_{h+1}\leq C p_h^r,\qquad p_c=C^{-1/(r-1)}.
$$

For p_0<p_c, logical error decreases doubly exponentially in encoding level h. A target with Q logical operations can achieve bounded total logical failure using h=O(log log(Q/delta)) and a polylogarithmic physical overhead. This is a conditional application of fault-tolerance reasoning, not a demonstrated threshold for actual matter. Assembly hierarchy depth L and error-code depth h are distinct quantities.

If every level introduces an uncorrected error floor eta, the recurrence becomes p_(h+1) <= C p_h^r + eta. The bound no longer tends to zero. For r=2 a small stable fixed point exists only when 4 C eta <= 1 in this scalar bound. Common-mode contamination, a faulty master template, and incorrect fusion can all create such floors. A hierarchy of good-looking module exteriors does not repair microscopic defects in sealed interiors.

## 11. Fusion is an independent manufacturing operation

The recognition layer should be replaceable or sacrificial. The final mechanical, electronic, or optical contacts require a different qualified process. The process sequence is reversible assembly, acceptance check, retention, permanent joining, and optional scaffold removal. If inspection is destructive, it can qualify a batch distribution but cannot certify the same individual object subsequently used.

| Conversion route | Plausible product | Required measurements and principal failure |
|---|---|---|
| DNA ligation or designed photochemical crosslink | More stable DNA framework | Bond conversion and shape retention; does not create an engineering metal or ceramic [R14] |
| Polymer infiltration and curing | Composite, compliant or rigid polymer part | Cure shrinkage, pores, modulus, residual monomer, damage to payloads |
| Silica mineralization | Silica/DNA hybrid or subsequent inorganic skeleton | Thickness, pore closure, joint continuity, fracture and drying collapse; mechanically useful precedents exist [R12,R13] |
| Electroless metal growth or seeded electrodeposition | Conductive tracks and metal framework | Bridging resistance, unintended shorts, grain boundaries, catalytic selectivity |
| Atomic-layer infiltration/deposition | Thin oxides or conformal barriers | Access into pores, precursor compatibility, conformality and thermal budget |
| Sintering, carbonization, ceramic conversion | Densified metal, carbon, or ceramic in compatible regions | Diffusion, shrinkage, cracks, lost dopants; unsuitable for mixed fragile payloads without separate processing |
| Epitaxial overgrowth | Selected high-quality crystalline regions | Crystal registry and defect annihilation; arbitrary fused nanoparticle mosaics do not become device-grade single crystals |

For final solid volume V_s from initial precursor volume V_0, unconstrained isotropic shrink factor is s=(V_s/V_0)^(1/3). With 10% retained solid volume, this model predicts 54% linear shrinkage. This is a mass-balance example, not a typical silica-coating result. A pre-existing open lattice coated conformally follows a different geometry and can preserve its pitch. Nonuniform strain matters more than a calibratable uniform factor.

A smoothing diffusion length is approximately sqrt(2Dt) per Cartesian direction. To retain a 10 nm feature within a 2 nm blur allowance requires Dt below roughly 2 x 10^-18 m^2. The relevant diffusion coefficient depends strongly on species, temperature, and surface versus bulk transport. This inequality is a process qualification target; no generic temperature can be inferred from it.

A planar coating of thickness t closes a gap g when 2t reaches g. A 2 nm layer on both sides of a 10 nm optical gap leaves about 6 nm before other shape changes. Conductive junctions may benefit from closure while optical gaps and bearings fail. The compiler must classify intended connections and forbidden shorts before selecting a conversion route.

## 12. Five target objects and where each breaks

The following are scale studies, not production forecasts. Counts are geometric estimates. Interface entries count simultaneous complementary recognition pairs; masks, complements, staples, catalysts, and conventional microassembly fixtures are extra. The detailed resource formulas and scenario assumptions are in the supplement and machine-readable target table.

| Target scenario | Precision-component count and pitch | Library / active pairs / tree depth | First bottleneck |
|---|---|---|---|
| Passive mechanical metamaterial, 10 micrometer cube, 5% occupied at 50 nm pitch | About 4 x 10^5 carriers | 2-4 families / 4-8 / 7 at branching 8 | Mineralized joint strength, drying distortion, inaccessible internal defects |
| Optical element, 100 x 100 x 1 micrometers, 50 nm pitch | About 8 x 10^7 sites | 3-8 families / 4-16 / 9 | Optical loss and correlated displacement; response cannot be inferred from shape alone |
| Electrical network, 10 x 10 x 0.1 micrometers, 10% active volume at 20 nm | About 1.25 x 10^5 fine components | 6-20 families / 8-32 / 6 | Low-resistance and chemically clean contacts; active junction reproducibility |
| MEMS-like sensor, 100 x 100 x 10 micrometers | About 8 x 10^5 coarse 0.5 micrometer sites plus 2.5 x 10^4 selected 20 nm sites | 10-30 families / 8-32 / 7 | Release, stiction, packaging, fatigue, actuator/process compatibility |
| One cubic centimeter electromechanical machine with sparse nanoregions | About 10^6 coarse 100 micrometer sites plus 8 x 10^7 fine sites if fine volume fraction is 10^-8 at 50 nm | 30-100+ payload families / 8-64 per qualified process cell / about 9 abstract levels | Cross-process integration, power, bearings, contamination, global metrology |

The last count is contingent on extraordinary sparsity of nanometer resolution. Uniform 50 nm discretization of one cubic centimeter requires 8 x 10^15 sites. A 10 nm discretization requires 10^18. Most bulk volume must therefore use much larger parts or continuous growth. A package containing a working motor or processor may use prefabricated chiplets; that is heterogeneous manufacturing, not proof that the voxel library itself manufactures the chiplets.

Planning assembly windows are 1-3 days for the passive test coupon, 2-7 days for a small optical array, 1-4 weeks for a first electrical-network experiment, and 2-8 weeks for a MEMS integration study. These estimates include iteration and handling, assume parallel prepared feedstocks, and are not supported production cycle times. No credible full-machine yield or fabrication time can yet be assigned. Purification could include gel or density separation for nanomodules, size/affinity sorting, and electrical/optical module screening, with substantial material loss.

Conditional geometry targets are 5-10 nm local registration before conversion for the passive coupon, 2-5 nm in optical hot spots, 1-2 nm at specialized junction interfaces, tens of nanometers for microsensor placement, and micrometer-scale macro-module placement. These are required or exploratory tolerances, not delivered accuracy. Uniform joint-error assumptions give upper scenario yields, but omission of fusion and correlated distortion would be unjustified. Consequently the supplement reports the fault rate required for a specified yield instead of inventing actual device yields.

Energy cannot be predicted from binding energies alone. At 300 K, a 20 kBT chemical step costs about 8.3 x 10^-20 J. Multiplying by successful plus rejected steps gives a molecular budget, not a wall-plug budget. Heating 100 microliters of water by 30 K already costs approximately 13 J ideally, many orders above the binding budget for a microscopic object. Instrument power, cooling, material synthesis, fluid handling, and wasted batches dominate current laboratory operation.

## 13. Adaptive resolution and manufacturing information

Let h(x) be local discretization pitch and assume the functional error admits a bound integral s(x) h(x)^a dx, with a>0 and nonnegative sensitivity density s. Minimize component count proportional to integral h(x)^(-d) dx subject to that error budget. A Lagrange multiplier gives:

$$
h(x)=\left[\frac{d}{\lambda a s(x)}\right]^{1/(a+d)},
$$

clipped to minimum and maximum manufacturable pitches and constrained by connectivity, interface matching, and minimum feature sizes. The sensitivity is an adjoint or perturbation bound over operating scenarios, not merely a numerical derivative at one nominal design. Near fracture, contact, percolation, and optical resonance, a linear local model may fail and requires robust bounds.

Define fabrication-description complexity K_fab(X,epsilon) as the shortest binary program for a fixed declared fabrication machine, including nonstandard feedstock recipes, fixtures, tests, and schedules, that produces X within tolerance with a declared failure bound. This quantity is reference-machine dependent and not generally computable. The package provides explicit recipe lengths, not measured Kolmogorov complexity.

A repeating b-ary construction may need one reusable rule plus a counter of size O(log N), or O(log N) expanded stage records; execution still uses Omega(N) primitive material events. Incompressible heterogeneous patterns require Omega(N log K) specification bits for K distinguishable local states. Addresses, external fixtures, or temporal instructions can move those bits but cannot erase them. Distinguish specification length, number of physical components, number of operations, sequential depth, and number of observations.

## 14. Second attack: remove the address layer from the bulk material

The initial design assumes each permanent material volume carries a programmable shell. That assumption wastes precision resources and leaves interfaces everywhere. Replace it, where possible, with a **removable sparse assembly framework**. Program only functional inclusions, material boundaries, channels, local catalysts, and fiducials. Then fill or grow large homogeneous regions using unaddressed feedstock. The address layer is a manufacturing aid rather than a constituent required in every final cell.

**Conditional surface-count proposition.** Consider piecewise homogeneous bodies with total controlled interface area A, bounded curvature at resolution h, V_f volume of truly fine functional regions, and bulk volume V_b fabricated at coarse pitch H. A process that can form and maintain their boundaries, fill each compatible region, and avoid trapped voids needs an asymptotic precision-component budget:

$$
N_{\mathrm{precision}}=O(A/h^2+V_f/h^3+V_b/H^3).
$$

This replaces V/h^3 only for this bounded-complexity class. It requires boundary access, process-selective filling, mechanical support, and conversion that respects the specification. For a one-millimeter cube and h=50 nm, full filling requires 8 x 10^12 cells whereas its six external surfaces have roughly 2.4 x 10^9 h-sized patches: a geometric ratio near 3,300 before overhead. This does not prove those billions of patches can be assembled cheaply, nor that every internal boundary can be ignored.

The framework could be DNA/protein scaffolds in water, patterned microcarriers, or sacrificial polymer features. Inorganic conversion precedents motivate the mechanism, but the complete compiler and heterogeneous filling system are a hypothesis. The strategy fails for dense molecularly heterogeneous matter, arbitrary dopant maps, or closed regions that cannot be filled without damage. It is compatible with additive manufacturing and directed self-assembly; novelty, if any, lies in the explicit combined accounting and future physical implementation, not the general idea of templating.

This stronger architecture replaces uniform nanovoxel filling as the recommended route to macroscopic objects. A second critique then asks whether a geometry-correct fused object is necessarily functional. It is not. Functional tests and sensitivity budgets must therefore be part of each module's acceptance rule. This revision stops the conceptual iteration: further performance claims require measurements of real modules rather than additional names for architectural layers.

## 15. Compiler and verification contracts

The pipeline starts with desired function and operating conditions, then continuum optimization, reachable-material checking, adaptive voxelization, process partitioning, module decomposition, conflict-graph construction, code assignment, and schedule generation. It returns a design plus explicit assumptions, process windows, quantitative rejection conditions, and a test plan. It must reject an unsupported target rather than produce an attractive rendering.

The objective is a weighted sum of functional error, assembly difficulty, residual risk, cost, and time:

$$
\mathcal{L}=\lambda_f L_{\mathrm{function}}+\lambda_a L_{\mathrm{assembly}}+\lambda_e L_{\mathrm{error}}+\lambda_c L_{\mathrm{cost}}+\lambda_t L_{\mathrm{time}}.
$$

Hard constraints include ingredient availability, thermal/chemical compatibility, accessible repair paths, minimum code distance after allowed rotations, precursor mass balance, transport capacity, and final lifetime. The weights do not legalize physically impossible solutions.

For each module, a contract states its external geometry, allowable environmental history, response bounds at its ports, inspectable observables, latent-error bound, and transformation uncertainty. Contracts compose only when the same physical assumptions hold at the joint. An optical resonance test does not certify an electrical barrier; surface metrology cannot bound arbitrary internal contamination. The supplied reference compiler computes conflict-graph colors, generates code assignments, and estimates an adaptive count. It is a prototype for these accounting steps, not a continuum multiphysics optimizer or a DNA sequence designer.

## 16. Reproducible kinetic study

The model is a continuous-time Markov chain for an open socket. Empty sites bind a correct or wrong component with rates lambda_C and lambda_W. Bound components detach at rates d_C and d_W. They traverse zero or two driven checking stages followed by metastable capture, at rate mu. Rejection returns the socket to empty. These transitions model an ideal fueled protocol; no claim is made that a chosen DNA sequence implements the rates.

After capture, unsealed joints can be lost during later stages. The locking protocol instead pays a finite processing delay and a per-joint conversion error. Hierarchy reduces the active variant count but dilutes higher-level modules under fixed primary-material concentration. This is an ideal-supply local-socket model: it omits three-dimensional free-cluster diffusion, spontaneous aggregation, depletion fluctuations, steric occlusion, and failed child provisioning. Perfect-yield products are conditional benchmark quantities, not measured or forecast device yields.

Six protocols share the same dynamics: random recognition; coded recognition; coded hierarchy; coded proofreading; proofreading with hierarchy; and proofreading/hierarchy with locking. The default concentration is 100 nM in primary-component units, k_on=10^6 per M per second, correct off-rate 0.01 per second, wrong-correct gap 6 kBT, progress rate 0.1 per second, hold-loss rate 10^-6 per second, fusion error 10^-4, 60-second fusion delay per level, and total deadline 20,000 seconds. All are illustrative model inputs. They are not a calibrated experimental parameter set.

The release contains 30 baseline cases, a 720-case sweep over count, discrimination, progress rate, and concentration policy, a 25-case fusion sweep, and an analytic threshold-floor sweep. Three kinetic cases were independently sampled with 10,000 Gillespie trajectories each. Nine tests check conservation, analytic infinite-time absorption, stochastic agreement, code distance, hierarchy counts, and the fact that late fusion errors remain. No rare-event rate of 10^-14 or smaller is inferred from Monte Carlo.

| Protocol, N=64 | Correct fraction | Wrong fraction | Missing fraction | Conditional perfect yield |
|---|---|---|---|---|
| Random | 0.0156 | 0.9844 | 0 | about 10^-114 |
| Coded | 0.3737 | 0.6263 | 0 | about 10^-27 |
| Hierarchical | 0.9101 | 0.0726 | 0.0173 | 0.00264 |
| Proofreading | 0.9988 | 0.00119 | below 10^-9 | 0.928 |
| Proofreading plus hierarchy | 0.9827 | 0.0000555 | 0.0173 | 0.333 |
| Plus locking | 0.99978 | 0.000156 | 0.0000603 | 0.986 |

At N=1,024 and the same fixed deadline/solids policy, the combined locked protocol has correct fraction only about 0.450, with about 0.550 missing and a log10 conditional perfect yield near -364. The main loss is incomplete assembly, not merely wrong recognition. At small N, ordinary proofreading can outperform hierarchy because staged handling creates delay and retention loss. Locking reduces loss yet adds wrong final joins. These counterexamples are central outputs, not failed optimizations to hide.

![Kinetic model comparison](../figures/kinetics.png)

*Figure 1. Six idealized kinetic protocols at the same deadline. Curves quantify the supplied local-socket model and are not experimental manufacturing yields.*

![Sensitivity to concentration policy and conversion](../figures/scaling.png)

*Figure 2. Changing module-concentration policy changes scaling; a concentration held constant across levels requires additional material/volume management. Uncorrected conversion errors and common-mode floors prevent unlimited microscopic reliability.*

## 17. Minimum viable experiment: a fused plasmonic sensor tile

Build a 16-carrier, approximately 0.2-0.4 micrometer planar tile whose selected carriers hold 30-40 nm gold particles forming defined dimers and isolated references. Use four to six payload/geometry variants based on one 40-60 nm origami design, four recognition-domain pairs in four ordered slots, and two stage-gating systems. Complements and gate fuels are additional strands; the underlying origami contains hundreds of staple sequences. This is a small carrier alphabet, not a six-molecule recipe.

First assemble four tetramers in separately controlled reactions. Apply reversible annealing and two timed, driven rejection/gating operations; quantify accepted wrong joints instead of assuming these operations realize ideal proofreading. Purify the retained tetramers, cap spent ports, and expose outer ports using the same recognition alphabet. Assemble the final tile; then grow a thin silica coating that bridges intended structural contact regions. This is one final irreversible material-conversion step. Test joints for actual continuous silica necks. Mere surface coating without mechanically joining carriers does not satisfy the fusion criterion.

![Logical experiment layout](../figures/experiment.png)

*Figure 3. A conceptual carrier/payload layout. Final molecular geometry and particle spacing require separate design and qualification.*

The functional observable is the coupled-particle optical spectrum and its reversible response to a calibrated refractive-index change. The tile also carries fluorescence distance reporters near selected joints to separate assembly fidelity from optical payload failure. Reference-particle spectra, not an asserted absolute wavelength, define the signal. The experiment tests recognition reuse, hierarchical joining, error rejection, and preservation of a useful optical function through conversion. It does not demonstrate a fault-tolerance threshold or general-purpose manufacturing.

Starting conditions to screen are pH 7.5-8.3, approximately 5-15 mM MgCl2, 1-20 nM per decorated carrier variant, and 20-45 C assembly after separately qualified origami folding. These are exploratory ranges. Exact DNA sequences, melting temperatures, buffer/coating compatibility, and silica precursor dosing must be designed and measured before execution. No unverified stock recipe is presented as laboratory-ready. The supplement specifies a staged design-of-experiments matrix, measurements, controls, sample counts, and decision thresholds.

Success requires: at least 50% correctly connected 16-carrier objects among the recovered object population; a separately reported mass recovery of at least 10% of carrier input; at least a tenfold reduction of wrong retained joints against an otherwise matched no-rejection control; code reuse that introduces less than a twofold increase in wrong-joint rate; final conversion preserving median particle-gap displacement within a prespecified 5 nm budget; and a reversible optical response exceeding five times measurement repeatability noise. These are go/no-go targets, not predicted outcomes. Demonstrating all of them in three independent batches would justify a larger functional assembly experiment.

## 18. Falsification and staged development

| Stage | Deliverable | Missing capability | Go/no-go criterion |
|---|---|---|---|
| 1: 10-100 carriers | Functional sensor tile with reused codes and permanent joints | Compatibility of rejection, port retirement, and conversion | Meet the prespecified tile criteria; otherwise redesign the interface or abandon that process combination |
| 2: 10^3-10^5 carriers | Modules with two payload families and aggregate functional testing | Scalable sorting, finite stock supply, defects that can be located and replaced | Measured risk and material cost predict a larger object with useful yield; no hidden rare-error extrapolation |
| 3: 10^6-10^9 carriers | Hierarchical nanosystem with verified ports | Correlated-error control, transport, parallel repair | Fault-injection tests show error contraction including joining and conversion, not merely better average morphology |
| 4: multi-material microsystem | Packaged sensor/actuator/electrical assembly | Incompatible process sequencing and low-resistance interfaces | Whole-system response and lifetime meet a benchmark against conventional integration |
| 5: macroscopic object with nanoregions | Coarse bulk body plus localized nanofunctions | Sparse-framework filling and global dimensional control | Count/cost advantage survives full process and waste accounting |
| 6: general-purpose feedstock-to-machine system | Automated compiler and multiple reproducible process families | Broad reachable library and composable module qualification | Multiple unrelated object classes fabricated from common inventories without bespoke local intervention |

No calendar prediction is assigned. Later stages depend on the earlier measured capabilities, not time alone.

Abandon the **global, all-active, one-pot architecture** if orthogonality or target concentration fails its measured fidelity/time bound. Abandon a **particular conversion route** if the required gap or material interface cannot survive even after compensation. Abandon the **finite-library claim for a stated application class** if almost every site requires a new payload recipe or if feedstock preparation and purification erase the advantage over standard manufacture. Abandon a **fault-tolerance claim** if a persistent correlated or conversion-error floor exceeds the system risk budget. These are different falsifications; failure of one should not be described as disproving all modular manufacturing.

Early decisive tests include complete cross-talk matrices over allowed orientations, wrong-register decoys, impurity-spike experiments, fault-injected modules, repeated port reuse with leakage measurement, matched pre/post-conversion three-dimensional metrology, and growth studies at fixed total solids concentration. Require blind classification and report incomplete and discarded objects. A spectrum averaged only over selected good structures cannot establish manufacturing yield.

## 19. Prior art, novelty classification, and intellectual scope

The central nanotechnology ingredients are already established individually. DNA bricks implement addressable heterogeneous shapes [R2]. Origami provides programmable scaffolds [R1]; material voxels explicitly separate payload from coordination [R3]. Foldable origami voxel chains and combinatorial crisscross assemblies already cover key modular and addressing ideas [R4,R6]. Patchy colloids and DNA-linked particles provide directional assembly and superlattices [R20,R21]. Protein design supplies compact modular interfaces [R11,R15]. DNA crosslinking and silica lattices address permanence and mechanical function [R12-R14].

Staged tile assembly already moves information from tile diversity into mixing schedules [R7,R8]. Tile proofreading and self-healing predate this project [R18,R19]. Multifarious mixtures address reusable components and competing targets [R9]; recent assembly-factor work explicitly addresses kinetic and encoding bottlenecks [R10]. Therefore neither finite alphabets, hierarchy, combinatorial recognition, nor the idea of proofreading can be claimed as a new invention here. Modular programming of geometry and interaction on shared DNA scaffolds is also directly demonstrated [R30].

| Claim | Class | Appropriate interpretation |
|---|---|---|
| DNA/inorganic functional carriers, templated optical response, silica conversion | A | Published enabling components, not validation of this integrated platform |
| Separate recognition, repair, and permanent joining | B | Practical synthesis of existing approaches |
| Conflict-graph accounting with explicit retirement leakage | C | Formal extension/engineering accounting; historical priority unestablished |
| Conditional logical threshold including conversion gadgets | B/C | A conditional transfer of established fault-tolerance logic; physical gadget is absent |
| Sparse removable framework plus functional contracts | C/D | Derived count law and integrated research hypothesis; templating itself is old |
| Arbitrary chemistry, atom-perfect machines, autonomous billion-component fabrication today | E | Unsupported and explicitly not claimed |

The search covered primary articles and author repositories across DNA bricks/origami, protein design, patchy colloids, nanoparticle superlattices, algorithmic assembly, proofreading, metamaterials, digital materials, directed self-assembly, frameworks, synthetic compartments, and molecular machinery. The companion prior-art matrix records sources and limitations. This was a targeted literature search, not an exhaustive patent clearance or proof of novelty. Where a publisher page was unavailable, the bibliographic record or accessible author manuscript was used, and no inaccessible experimental methods were represented as inspected.

## 20. Limitations and open problems

No physical error-correcting gadget is demonstrated. No DNA sequence set, carrier CAD file, kinetic calibration, molecular-dynamics validation, continuum finite-element model, or experimental data is supplied. The codebook is abstract and must be translated into sequence and geometry constraints. The Markov simulation omits numerous couplings that can only worsen or qualitatively change assembly relative to independent sockets. It is valuable for ruling out simplistic scaling claims, not certifying a nanofabricator.

Open mathematical work includes an experimentally grounded noise model; reachable tensor sets with non-negligible joint impedances; self-assembly bounds under finite inventory and transport; formal composition of geometric and functional acceptance tests; and optimal partitioning across process-incompatible materials. Open physical work includes affordable payload qualification, true local removal/replacement, robust port retirement, low-distortion fusion, and continuous electrical/thermal paths across joints.

The most promising achievable improvement is a reusable, experimentally measurable interface-and-conversion contract whose fault rate decreases under a real correction cycle. A proposed architecture cannot substitute for that measurement.

<!-- PAGEBREAK -->

## 21. Final assessment

**Status: experimentally testable.** Mathematical consequences and numerical benchmarks are present; the complete physical mechanism remains unvalidated.

| Completeness area | Estimate | Basis |
|---|---|---|
| Mathematical theory | 70% | Conditional proofs cover the central counts and bounds; no physical sufficiency theorem |
| Physical architecture | 40% | Specific plausible platforms and interfaces; no integrated fabricated design |
| Simulation validation | 35% | Reproducible kinetic comparisons and cross-checks; no spatial/molecular calibration |
| Experimental design | 65% | Falsifiable assay, controls, metrics and gates; sequence/CAD/chemistry optimization remains |
| Scalability analysis | 60% | Information, transport, yield, correlation and fusion limits treated; empirical scaling unknown |
| Universal-fabrication relevance | 25% | Credible relevance to restricted hybrid manufacturing; no general feedstock-to-machine capability |

These percentages are coarse editorial estimates of unresolved work within this manuscript's stated scope, not measured probabilities, technology-readiness scores, or fractions of the entire research field solved.

**Breakthrough significance: potentially transformative**, conditional on experimentally establishing the missing composable correction-and-conversion primitive. The delivered contribution is a bounded theoretical extension and test plan, not an established transformative breakthrough.

**Most important unresolved obstacle:** a physical module that corrects realistic correlated joining and conversion faults while preserving its required function, with a residual error bound that composes across levels.

**Most decisive next experiment:** the preregistered 16-carrier plasmonic tile experiment, testing reuse of the same interface alphabet, driven wrong-joint rejection, and silica joining while measuring both defect contraction and retained optical function.

## References

See the complete numbered bibliography and source-access notes in REFERENCES.md, included below in the PDF edition. The technical supplement contains proofs, definitions, additional parameter tables, the complete experiment plan, and the prior-art matrix. All code, input parameters, outputs, figures, and metadata accompany this manuscript.