Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2607.16869

Beyond Orbital Rotations: Correlation-Rank Limits and Clifford-Accessible Measurement, from Algebra and Global Optimization

Federico Zahariev, Vanda Glezakou

quant-phphysics.chem-phSubmitted Jul 18, 2026• Announced Jul 21, 2026• 0 citations• Revised Oct 7, 2026 (v2)
v1 · Jul 18, 2026v2 · Oct 7, 2026
arXivHTMLPDF

Abstract

One should not assume that reconstructing an electronic observable with a smaller set of measurement settings will cut the number of repetitions in proportion. Each setting is simply a basis change and qubit readout, but the savings are not guaranteed to be commensurate. Take two electrons of opposite spin in two spatial orbitals: the orbital degree of freedom of each electron acts as a logical qubit and rotations are independent in the spin sectors. From the observable's one- and two-qubit coefficients we derive an exact floor for the setting count and a decomposition to meet it in no more than three settings. With fewer settings, the minimum error in the two-qubit correlation coefficients is set by the singular values left out. A case in point is an exchange with equal nonzero couplings on the three Pauli axes. This demands three orbital settings yet only one joint measurement. We offer a Clifford circuit for four physical qubits to effect the necessary mapping of Pauli products under conjugation. In the case of a one-parameter family of logical singlet-triplet mixtures, the orbital-to-Clifford shot cost ratio is three at maximum mixing and trends toward 3/2 near the logical singlet. Where spin sectors may be mixed, isotropic exchange needs one orbital setting. For spin-conserving Hamiltonians with $N$ spatial orbitals encoded via Jordan-Wigner, the $X$-rank, the binary rank of the $X$ parts of its strings, of any commuting Pauli string family is no more than $2(N-1)$. Searches with the Robust Adaptive Nature-inspired Global Explorer (RANGE) reach this limit for CH$_4$, with the rank of the Hamiltonian support vouching for the maximality of the search values. And for all nonidentity Pauli strings on $n$ qubits, semidefinite witnesses from the Lovász theta number give a lower bound on the Clifford setting count that agrees with the integer cover and counting bounds at $2^n+1$.

Comments
Revised version. Heavy-element X-ranks recomputed from audited operators, text revised throughout, references updated. 21 pages, 2 figures

Related papers

Certified Bounds on the Measurement Cost of Molecular Hamiltonians

Federico Zahariev, Vanda Glezakou

Quantum estimates of molecular energy require repeated state preparation and measurement. We bound the minimum repetition count for given measurements and processing rules from above with a practical schedule and from below with independently verified inequalities. For an H$_4$ chain at H-H distances of 0.90 and 1.80 Åwith 2 per cent uniform noise in the ground state, entangling Clifford measurements, which map Pauli products onto one another, cut leading-order repetitions by at least 65.9 and 61.0 per cent, respectively, relative to any schedule using only single-qubit Pauli measurements and Hamiltonian terms. With 69 unaltered measurement settings, each a basis change followed by qubit readout, processing further observables saves at least 12.04 and 6.70 per cent. For four lanthanide and actinide Hamiltonians with depolarized Hartree-Fock covariances, reducing the sum of absolute Hamiltonian coefficients does not proportionally reduce sampling. The bounds cover linear unbiased estimates from independent batches at small error.

quant-pharXiv:2607.16866•Jul 21, 2026•0 citations

Hybrid Qubit-Rotor Quantum Systems: Clifford Structure, Universal Control, and Applications

Dengyao Luo, Arvin Kushwaha, Mastawal Tirfe +1 more

A $U(1)$ quantum rotor pairs a periodic angle with an integer-valued conjugate momentum, and occurs in molecular rotation, superconducting phase-charge circuits, and compact gauge fields. Coupling such a rotor coherently to qubits gives a hybrid register whose control structure is not inherited from either the oscillator-qubit or the qudit case. We develop a Clifford theory, together with a universal-control result, for registers of $n$ qubits and $r$ rotors. We classify all automorphisms of the hybrid phase space $\mathbb{F}_2^{2n}\times\mathbb{Z}^r\times\mathbb{T}^r$ that preserve the Weyl commutation relations, and give an explicit finite Clifford circuit for each one. The classification is directional: rotor momentum parity may control qubit Pauli operations within the Clifford group, while every nonzero qubit-controlled rotor momentum shift is non-Clifford. It also yields normal forms for the mixed qubit-rotor couplings and the exact minimum number of elementary mixed gates needed to synthesize them. Adding a rotor cosine potential and one fixed qubit-rotor conditional phase to the local Clifford operations gives universal control on the full Hilbert space in the strong operator topology. We then apply this structure in three settings: an exact controlled-shift realization of gauge-covariant matter hopping, which is necessarily non-Clifford; rotor phase estimation with direct angle readout and probe optimization under momentum-support and energy constraints; and finite Fourier transforms on rotor momentum codes, where the one-rotor transform for $d=2^s$ compiles into $O(s)$ momentum-selective and controlled-shift instructions and each cross-register Fourier factor is implemented by one quadratic rotor Clifford gate.

quant-pharXiv:2608.20227•Aug 21, 2026•1 citation

Exact Clifford Optimality for Two-Sided Locally Randomized Classical Shadows

Xiaotian Nie, Tao Zhang, Yadong Wu +1 more

Entangling measurements can reduce the cost of learning many-body correlations. We consider classical shadows in which a fixed unitary on a $k$-qubit region is sandwiched between two layers of random single-qubit Clifford gates, and ask which middle unitary best estimates Pauli strings with full support on the region. Combining symplectic support counting with a constraint from operator-size parity, we prove that within this architecture the minimum squared shadow norm over all Clifford middle unitaries is $C_k^{\rm opt}=4\cdot 9^k/(3\cdot 5^k+3^k-6)$ for odd $k$ and $C_k^{\rm opt}=4\cdot 9^k/(3\cdot 5^k-3^k+2)$ for even $k$. A dual-axis rotation circuit $U_*$ attains this minimum for every $k$. Asymptotically $C_k^{\rm opt}\simeq(4/3)\cdot(9/5)^k$: the exponential base of Wu et al.'s contractive unitary is optimal, and the prefactor improves from $2$ to $4/3$. For $1\leq k\leq7$, exact certificates show that $U_*$ is optimal over the entire unitary group $\mathrm U(2^k)$, and we conjecture that this holds for every $k$.

quant-pharXiv:2610.04621•Oct 6, 2026•0 citations

Efficient Block Encoding of Structured Hamiltonians by Separating Where and What

Alessandro Summer, François Jamet

In fault-tolerant block encodings of structured Hamiltonians, the support geometry can set how the non-Clifford cost of SELECT scales with the system size, however many operators act on each support. We construct permute-act-unpermute circuits: a CSWAP network fixed by this geometry brings a selected support to a fixed target register, a shared local circuit acts there, and the permutation is undone. The operators add only an $XZ$ mask on this register, at a cost independent of the system size, and one query can mix several support geometries. For two-site supports, we establish and attain the minimum CSWAP count within address-controlled networks of site transpositions. Compiled for Pauli terms, these circuits use $8\mathrm{S}N+o(N)$ $T$ gates with $O(\log N)$ permutation work qubits, where $N$ is the number of system qubits and $\mathrm{S}$ is determined by the support geometry, with $\mathrm{S}=1$ for nearest-neighbour interactions and $\mathrm{S}=3/2$ for all-to-all pairs; in both cases the correction is $O(\log N)$. We also introduce a bridged CSWAP, which remembers the logical value of its temporary AND across the target action and repairs it from the prepared labels, so that the inverse CSWAP undoes the exchange without recomputing the value. When the repair is Clifford, it halves the $T$ count of a matched forward-inverse CSWAP pair at the cost of one retained work qubit. For a Heisenberg ring, the complete block-encoding query reduces the $T$ count by a factor approaching $3$ as the system size grows, relative to flat and product-wise Pauli walks. For a five-orbital Anderson impurity model, which combines nearest-neighbour and all-to-all support geometries, the reduction is about $1.7$ at large bath sizes, relative to the cheapest of three comparators, each built from the methods of the references it follows.

quant-pharXiv:2610.03686•Oct 5, 2026•2 citations

Structurally Sparse Qubit Hamiltonians from Spectral Graph Theory

Josh Loutey

We show that qubit Hamiltonians derived from spectral graph theory lattices exhibit structurally favorable Pauli decompositions for quantum simulation. After Jordan–Wigner transformation, the GeoVac lattice Hamiltonian yields $O(Q^{3.15})$ Pauli string terms in the number of qubits $Q$, compared to $O(Q^{4+})$ for conventional Gaussian-basis Hamiltonians (our own log-log fit of the published Pauli counts of Trenev et al. [ref] gives $Q^{4.25}$ for LiH and $Q^{3.92}$ for H$_2$O, spanning $Q^{3.9}$–$Q^{4.3}$). The advantage originates in angular momentum selection rules—enforced by Gaunt integrals at the two-electron integral level—which cause the electron repulsion integral (ERI) tensor density to decay as $\sim 1/M$ (empirically $\approx M^{-0.85}$–$M^{-1}$) with the number of spatial orbitals $M$, while Gaussian ERI tensors remain near-dense at $O(1)$. Beyond term counts, we quantify two operationally relevant cost metrics: the number of qubitwise-commuting (QWC) measurement groups scales as $O(Q^{3.36})$, and the Pauli 1-norm $\lambda = \sum_i |c_i|$—which controls both Trotter step counts and qubitization query complexity—scales as $O(Q^{1.69})$ with $R^2 = 0.997$. The sub-quadratic 1-norm scaling is the key result for fault-tolerant simulation: it implies that Trotter circuit depth grows far more slowly with system size than term counts alone would suggest. All Gaussian baselines are validated against published integral values (Szabo and Ostlund for H$_2$; analytical STO for He). For composed multi-center systems, the pseudopotential approximation of the composed natural geometry framework [ref] yields block-diagonal electron repulsion integrals with a universal Pauli scaling of $O(Q^{2.5})$, independent of molecular topology (exponent spread $0.02$ across three molecules). At equal qubit counts, this represents two or more orders of magnitude fewer Pauli terms than Gaussian baselines (published counts of Trenev et al. [ref], $Q^{3.9}$–$Q^{4.3}$ span fitted by us)—an advantage that grows with system size due to the $\sim 1.5$ exponent gap. These comparisons are at matched qubit counts, not matched accuracy; the composed basis carries classical equilibrium-geometry errors of 5–26% (LiH/BeH$_2$/H$_2$O; Sec. (ref)), making the framework optimal for fixed-geometry quantum simulation rather than geometry optimisation. Across the composed molecular library (35 molecules spanning $Z = 1$–56, H through Ba, including the first transition series ScH–ZnH), the composed Pauli count is exactly linear in the qubit count: $N_{\mathrm{Pauli}} = 11.10 × Q$ (exact, non-identity terms; $d$-orbital blocks give 9.23). These counts are realized under a pair-diagonal angular approximation ($m_a=m_c$, $m_b=m_d$) adopted deliberately for sparsity; the exact global-$M_L$ Coulomb rule (used in the framework's precision-physics paths, and considered but left on the table here) gives $\approx 2.5×$ more terms with the same scaling, under which the $d$-block coefficient is in fact higher, not lower (Sec. (ref)). GeoVac Paper 14. Part of the GeoVac research corpus (quantum computing group). Code, tests, and the full paper series: https://github.com/jloutey-hash/geovac. Corpus archive DOI: 10.5281/zenodo.20482394. Corpus version at deposit: v4.75.0.

•Jul 9, 2026•0 citations

Data informed Ansatz Selection for Variational Quantum Circuits

Manav Seksaria, Advith Desu, A K Prabhakar

Choosing a variational quantum circuit (VQC) usually means searching over ansatze and depths, and training each candidate. We give three rules that fix these choices from the data and the gate set alone, before any training. First, a circuit without entangling gates has a rank-one Fourier coefficient tensor on every subset of qubits. A singular value decomposition of a dataset's Fourier coefficients therefore indicates whether entanglement is needed, which qubits it connects, and a lower bound on the error of omitting it. Second, the rank of the quantum Fisher information matrix grows by $p_0$ per layer, after a first-layer deficit $d_0$, until it reaches a ceiling set by the dynamical Lie algebra. This law fixes the depth. Third, reading out all $2^N - 1$ Pauli-Z products from one batch of shots resolves half of the state's directions, the most any single-basis readout can. A single-qubit readout resolves an exponentially small fraction. For the Madelon dataset, the rules prescribe 7 blocks of 3 qubits at depth 3: a 239-parameter classifier with 78.5% test accuracy, found without a grid search. Ablations confirm three predicted effects: shallower circuits, smaller blocks and dropping reuploading all lower accuracy.

arXiv:2611.05079•Nov 3, 2026•0 citations
One should not assume that reconstructing an electronic observable with a smaller set of measurement settings will cut the number of repetitions in proportion. Each setting is simply a basis change and qubit readout, but the savings are not guaranteed to be commensurate. Take two electrons of opposite spin in two spatial orbitals: the orbital degree of freedom of each electron acts as a logical qubit and rotations are independent in the spin sectors. From the observable's one- and two-qubit coefficients we derive an exact floor for the setting count and a decomposition to meet it in no more than three settings. With fewer settings, the minimum error in the two-qubit correlation coefficients is set by the singular values left out. A case in point is an exchange with equal nonzero couplings on the three Pauli axes. This demands three orbital settings yet only one joint measurement. We offer a Clifford circuit for four physical qubits to effect the necessary mapping of Pauli products under conjugation. In the case of a one-parameter family of logical singlet-triplet mixtures, the orbital-to-Clifford shot cost ratio is three at maximum mixing and trends toward 3/2 near the logical singlet. Where spin sectors may be mixed, isotropic exchange needs one orbital setting. For spin-conserving Hamiltonians with NNN spatial orbitals encoded via Jordan-Wigner, the XXX-rank, the binary rank of the XXX parts of its strings, of any commuting Pauli string family is no more than 2(N−1)2(N-1)2(N−1). Searches with the Robust Adaptive Nature-inspired Global Explorer (RANGE) reach this limit for CH4_44​, with the rank of the Hamiltonian support vouching for the maximality of the search values. And for all nonidentity Pauli strings on nnn qubits, semidefinite witnesses from the Lovász theta number give a lower bound on the Clifford setting count that agrees with the integer cover and counting bounds at 2n+12^n+12n+1.
Quantum estimates of molecular energy require repeated state preparation and measurement. We bound the minimum repetition count for given measurements and processing rules from above with a practical schedule and from below with independently verified inequalities. For an H4_44​ chain at H-H distances of 0.90 and 1.80 Åwith 2 per cent uniform noise in the ground state, entangling Clifford measurements, which map Pauli products onto one another, cut leading-order repetitions by at least 65.9 and 61.0 per cent, respectively, relative to any schedule using only single-qubit Pauli measurements and Hamiltonian terms. With 69 unaltered measurement settings, each a basis change followed by qubit readout, processing further observables saves at least 12.04 and 6.70 per cent. For four lanthanide and actinide Hamiltonians with depolarized Hartree-Fock covariances, reducing the sum of absolute Hamiltonian coefficients does not proportionally reduce sampling. The bounds cover linear unbiased estimates from independent batches at small error.
A U(1)U(1)U(1) quantum rotor pairs a periodic angle with an integer-valued conjugate momentum, and occurs in molecular rotation, superconducting phase-charge circuits, and compact gauge fields. Coupling such a rotor coherently to qubits gives a hybrid register whose control structure is not inherited from either the oscillator-qubit or the qudit case. We develop a Clifford theory, together with a universal-control result, for registers of nnn qubits and rrr rotors. We classify all automorphisms of the hybrid phase space F22n×Zr×Tr\mathbb{F}_2^{2n}\times\mathbb{Z}^r\times\mathbb{T}^rF22n​×Zr×Tr that preserve the Weyl commutation relations, and give an explicit finite Clifford circuit for each one. The classification is directional: rotor momentum parity may control qubit Pauli operations within the Clifford group, while every nonzero qubit-controlled rotor momentum shift is non-Clifford. It also yields normal forms for the mixed qubit-rotor couplings and the exact minimum number of elementary mixed gates needed to synthesize them. Adding a rotor cosine potential and one fixed qubit-rotor conditional phase to the local Clifford operations gives universal control on the full Hilbert space in the strong operator topology. We then apply this structure in three settings: an exact controlled-shift realization of gauge-covariant matter hopping, which is necessarily non-Clifford; rotor phase estimation with direct angle readout and probe optimization under momentum-support and energy constraints; and finite Fourier transforms on rotor momentum codes, where the one-rotor transform for d=2sd=2^sd=2s compiles into O(s)O(s)O(s) momentum-selective and controlled-shift instructions and each cross-register Fourier factor is implemented by one quadratic rotor Clifford gate.
Entangling measurements can reduce the cost of learning many-body correlations. We consider classical shadows in which a fixed unitary on a kkk-qubit region is sandwiched between two layers of random single-qubit Clifford gates, and ask which middle unitary best estimates Pauli strings with full support on the region. Combining symplectic support counting with a constraint from operator-size parity, we prove that within this architecture the minimum squared shadow norm over all Clifford middle unitaries is Ckopt=4⋅9k/(3⋅5k+3k−6)C_k^{\rm opt}=4\cdot 9^k/(3\cdot 5^k+3^k-6)Ckopt​=4⋅9k/(3⋅5k+3k−6) for odd kkk and Ckopt=4⋅9k/(3⋅5k−3k+2)C_k^{\rm opt}=4\cdot 9^k/(3\cdot 5^k-3^k+2)Ckopt​=4⋅9k/(3⋅5k−3k+2) for even kkk. A dual-axis rotation circuit U∗U_*U∗​ attains this minimum for every kkk. Asymptotically Ckopt≃(4/3)⋅(9/5)kC_k^{\rm opt}\simeq(4/3)\cdot(9/5)^kCkopt​≃(4/3)⋅(9/5)k: the exponential base of Wu et al.'s contractive unitary is optimal, and the prefactor improves from 222 to 4/34/34/3. For 1≤k≤71\leq k\leq71≤k≤7, exact certificates show that U∗U_*U∗​ is optimal over the entire unitary group U(2k)\mathrm U(2^k)U(2k), and we conjecture that this holds for every kkk.
In fault-tolerant block encodings of structured Hamiltonians, the support geometry can set how the non-Clifford cost of SELECT scales with the system size, however many operators act on each support. We construct permute-act-unpermute circuits: a CSWAP network fixed by this geometry brings a selected support to a fixed target register, a shared local circuit acts there, and the permutation is undone. The operators add only an XZXZXZ mask on this register, at a cost independent of the system size, and one query can mix several support geometries. For two-site supports, we establish and attain the minimum CSWAP count within address-controlled networks of site transpositions. Compiled for Pauli terms, these circuits use 8SN+o(N)8\mathrm{S}N+o(N)8SN+o(N) TTT gates with O(log⁡N)O(\log N)O(logN) permutation work qubits, where NNN is the number of system qubits and S\mathrm{S}S is determined by the support geometry, with S=1\mathrm{S}=1S=1 for nearest-neighbour interactions and S=3/2\mathrm{S}=3/2S=3/2 for all-to-all pairs; in both cases the correction is O(log⁡N)O(\log N)O(logN). We also introduce a bridged CSWAP, which remembers the logical value of its temporary AND across the target action and repairs it from the prepared labels, so that the inverse CSWAP undoes the exchange without recomputing the value. When the repair is Clifford, it halves the TTT count of a matched forward-inverse CSWAP pair at the cost of one retained work qubit. For a Heisenberg ring, the complete block-encoding query reduces the TTT count by a factor approaching 333 as the system size grows, relative to flat and product-wise Pauli walks. For a five-orbital Anderson impurity model, which combines nearest-neighbour and all-to-all support geometries, the reduction is about 1.71.71.7 at large bath sizes, relative to the cheapest of three comparators, each built from the methods of the references it follows.
We show that qubit Hamiltonians derived from spectral graph theory lattices exhibit structurally favorable Pauli decompositions for quantum simulation. After Jordan–Wigner transformation, the GeoVac lattice Hamiltonian yields O(Q3.15)O(Q^{3.15})O(Q3.15) Pauli string terms in the number of qubits QQQ, compared to O(Q4+)O(Q^{4+})O(Q4+) for conventional Gaussian-basis Hamiltonians (our own log-log fit of the published Pauli counts of Trenev et al. [ref] gives Q4.25Q^{4.25}Q4.25 for LiH and Q3.92Q^{3.92}Q3.92 for H2_22​O, spanning Q3.9Q^{3.9}Q3.9–Q4.3Q^{4.3}Q4.3). The advantage originates in angular momentum selection rules—enforced by Gaunt integrals at the two-electron integral level—which cause the electron repulsion integral (ERI) tensor density to decay as ∼1/M\sim 1/M∼1/M (empirically ≈M−0.85\approx M^{-0.85}≈M−0.85–M−1M^{-1}M−1) with the number of spatial orbitals MMM, while Gaussian ERI tensors remain near-dense at O(1)O(1)O(1). Beyond term counts, we quantify two operationally relevant cost metrics: the number of qubitwise-commuting (QWC) measurement groups scales as O(Q3.36)O(Q^{3.36})O(Q3.36), and the Pauli 1-norm λ=∑i∣ci∣\lambda = \sum_i |c_i|λ=∑i​∣ci​∣—which controls both Trotter step counts and qubitization query complexity—scales as O(Q1.69)O(Q^{1.69})O(Q1.69) with R2=0.997R^2 = 0.997R2=0.997. The sub-quadratic 1-norm scaling is the key result for fault-tolerant simulation: it implies that Trotter circuit depth grows far more slowly with system size than term counts alone would suggest. All Gaussian baselines are validated against published integral values (Szabo and Ostlund for H2_22​; analytical STO for He). For composed multi-center systems, the pseudopotential approximation of the composed natural geometry framework [ref] yields block-diagonal electron repulsion integrals with a universal Pauli scaling of O(Q2.5)O(Q^{2.5})O(Q2.5), independent of molecular topology (exponent spread 0.020.020.02 across three molecules). At equal qubit counts, this represents two or more orders of magnitude fewer Pauli terms than Gaussian baselines (published counts of Trenev et al. [ref], Q3.9Q^{3.9}Q3.9–Q4.3Q^{4.3}Q4.3 span fitted by us)—an advantage that grows with system size due to the ∼1.5\sim 1.5∼1.5 exponent gap. These comparisons are at matched qubit counts, not matched accuracy; the composed basis carries classical equilibrium-geometry errors of 5–26% (LiH/BeH2_22​/H2_22​O; Sec. (ref)), making the framework optimal for fixed-geometry quantum simulation rather than geometry optimisation. Across the composed molecular library (35 molecules spanning Z=1Z = 1Z=1–56, H through Ba, including the first transition series ScH–ZnH), the composed Pauli count is exactly linear in the qubit count: NPauli=11.10×QN_{\mathrm{Pauli}} = 11.10 × QNPauli​=11.10×Q (exact, non-identity terms; ddd-orbital blocks give 9.23). These counts are realized under a pair-diagonal angular approximation (ma=mcm_a=m_cma​=mc​, mb=mdm_b=m_dmb​=md​) adopted deliberately for sparsity; the exact global-MLM_LML​ Coulomb rule (used in the framework's precision-physics paths, and considered but left on the table here) gives ≈2.5×\approx 2.5×≈2.5× more terms with the same scaling, under which the ddd-block coefficient is in fact higher, not lower (Sec. (ref)). GeoVac Paper 14. Part of the GeoVac research corpus (quantum computing group). Code, tests, and the full paper series: https://github.com/jloutey-hash/geovac. Corpus archive DOI: 10.5281/zenodo.20482394. Corpus version at deposit: v4.75.0.
Choosing a variational quantum circuit (VQC) usually means searching over ansatze and depths, and training each candidate. We give three rules that fix these choices from the data and the gate set alone, before any training. First, a circuit without entangling gates has a rank-one Fourier coefficient tensor on every subset of qubits. A singular value decomposition of a dataset's Fourier coefficients therefore indicates whether entanglement is needed, which qubits it connects, and a lower bound on the error of omitting it. Second, the rank of the quantum Fisher information matrix grows by p0p_0p0​ per layer, after a first-layer deficit d0d_0d0​, until it reaches a ceiling set by the dynamical Lie algebra. This law fixes the depth. Third, reading out all 2N−12^N - 12N−1 Pauli-Z products from one batch of shots resolves half of the state's directions, the most any single-basis readout can. A single-qubit readout resolves an exponentially small fraction. For the Madelon dataset, the rules prescribe 7 blocks of 3 qubits at depth 3: a 239-parameter classifier with 78.5% test accuracy, found without a grid search. Ablations confirm three predicted effects: shallower circuits, smaller blocks and dropping reuploading all lower accuracy.