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

When a Quantum Compiler Looks Like a Better Learner: Confounds in Quantum Machine Learning Comparisons

TN.Holdings, TN.Holdings

Published Oct 23, 2026• 1 citation
Publisher

Abstract

PSF-Zero places a quantum classifier's entangling ring on hardware by formula, without a layout search and without reading the day's calibration; Qiskit's preset pass manager searches, using the calibration. We ask which of the two gives the better trained classifier, and report what it took to answer that question honestly. The model is a 10-qubit, 16-layer data re-uploading classifier trained by a finite-difference optimiser on a noise model of ibm_marrakesh (the twin) and tested on an independently perturbed noise model (the stand-in); one day of hardware runs is included. Four effects can each produce a winner without any difference in learning: (i) the task's answer sat at the untrained starting point: on the device day's task the untrained model is already 0.934 accurate (noiseless), and training lowered both compilers' models to 0.762 (PSF-Zero) and 0.730 (Qiskit); (ii) a readout offset makes a circuit answer one class on 84-94 % of balanced test points, which moves accuracy while leaving the model unchanged; (iii) arms with different output scales lose different amounts to a fixed shot count; (iv) the twin's scale was off by up to 2.3x on hardware. With the task moved far from the start and the rules fixed before each run, PSF-Zero's newest ring schedule (the zipper, 336 CZ and 11.7 us at 16 layers) and Qiskit (506 CZ, 48.2 us, 14 active qubits) gave classifiers that the declared tests could not tell apart: on stand-in accuracy the zipper won on one of three teachers at each budget, below the declared two of three, and on the trained model itself neither arm won on any teacher. Formula placement could not be told apart from the search while using a third fewer entangling gates and a quarter of the duration; it did not beat it. We give the vanity and the essential number side by side throughout, mark what was declared before the data and what was found after, and close with a checklist for compiler comparisons through trained quantum models. In one preliminary hardware job on ibm_kingston, the zipper's circuits followed a twin built from that day'scalibration (scale 1.00, correlation 0.98), while Qiskit's lost most of their signal and came out inverted on theread-out qubit; the cause is not known. Code and records: https://github.com/TN-Holdings-LLC/psf-zero (paper and figure script) and https://github.com/TN-Holdings-LLC/TOQB (runs, records, declarations).

DOI
10.5281/zenodo.23600811

Related papers

Training quantum neural networks for Tl-limited hardware: what the qg symmetry filter does and does not do

Vicente Humberto Monteverde

A quantum neural network (QNN) built from gates that conserve the Hamming weight keeps its data in one weight sector, so the shots that left the sector can be discarded at readout. This is the qg filter of the qang library [1, 2]. We show that, with equal amplitude damping (T1) on every qubit, the filtered readout of such a network is exactly the noiseless one in any weight sector. The kept fraction of shots is (1 − γ)k d for weight k and depth d, and training under T1 with the filter gives exactly the parameters of noiseless training. We then measure what this is worth on four small classification datasets. Every result is reported with the filter and without it, and every prediction was committed before its run, failures included. Over 60 runs per model, a network trained on a simulator and run under T1 gains +2.5 points from the filter at weight 1 (95% CI [0.8, 4.1]) and +16.5 points at weight 2 ([11.8, 21.2]), recovering its noiseless accuracy exactly. A network trained under the calibrated noise without the filter catches up (diflerences of 0.1 points), and the filter alone corrects neither dephasing nor unequal T1 across qubits, although it loses less than one point up to a ±80% spread of 1/T1. Unequal T1 is corrected by training UÏtl the filter under the calibrated decay rates: the kept state depends only on their ratios, so the model lands within 0.5 points of noiseless at a ±100% spread, ignores a 10% calibration error and, at weight 2, loses 0.7 points under a 1.5-fold drift of all T1 that costs the raw noise-aware model 10. With noise-aware training and a richer readout, the model without the filter was 1.2 points better in one study, but over 60 new runs that gap shrank to +0.5 points and was not significant. The weight-2 deficit comes from the pair-product encoding, not from the readout; with an encoding that fills the whole sector and a full-rank readout, weight 2 reaches the weight-1 accuracy. On noise models of IBM and IonQ devices the filter cuts the error of the measured qgZ by 2.7-3.6×; on the IonQ Forte-1 noise model over 189 inputs of four datasets the filtered model is 1.1 points more accurate than the raw one, and calibrating the errors the filter keeps with five echo circuits per model halves the remaining error of the decision value: decisions that difler from the noiseless model fall from 15 (raw) and 7 (filter) to 3 over 945 readings on five device noise models. With a faithful weight-2 encoding the advantage of the unfiltered model under noise-aware training disappears, and a classifier given both the filtered and the raw features gains nothing over the filter alone. Over 513 readout configurations, including kept fractions down to 2.6 ×10−7, unequal T1 and dephasing, the filter has the lower readout error whenever about 20 shots are kept, and a closed-form rule predicts the winner. The same holds beyond the first model: at 8 qubits a weight-2 network keeps its noiseless 0.965 with the filter and drops to 0.705 without it; a five-class classifier (one class per qubit) keeps its noiseless accuracy under T1 with the filter and loses 15-18 points without it on five-class digits; training from 100 or 1000 shots per step with the filter is 1.0-1.5 points ahead of training without it; and the filtered gradient equals the noiseless one, while the raw one is K times smaller at weight 1. On five device noise models the echo calibration turns class confusions back: flipped decisions of the five-class model fall from 51 (raw) and 24 (filter) to 15 over 500 readings. The echo helped only where the filtered error was well above shot noise, and it still removed a third of the filtered error with a 28 × 28 matrix at 8 qubits. These models are classically simulable and no quantum advantage is claimed. The filter is

•Oct 11, 2026•0 citations

Exact answers, no map, no mitigation: user-written circuits through a placement-only OS on IBM Heron r2 devices

TN.Holdings

A user writes, in plain Qiskit, a circuit whose answer is known exactly. A small operating layer (psf-os) places it on an IBM Heron r2 device by translation, without reading any calibration-quality number, adds its own controls to the same job, and runs it with no error mitigation. We report how close the raw answers come and how that is shown. Two points of the report are of use beyond our own numbers. First, an absolute criterion |M - M_exact| <= c is biased in both directions: on 5 November a problem that DIFFERS under c = 0.05 kept 0.82-0.86 of its signal, as much as a problem that AGREES (0.73-0.85), and a small-answer problem AGREES while keeping only 0.73 of its signal. We therefore print the declared verdict first and the retention R = M/M_exact beside it on every row. Second, every number can be checked: answers are sealed by hash before each run, raw counts and job identifiers are committed, and one script re-scores all 13 device results of this paper from the raw counts and finds every number the same. The results: 22 practice circuits on two devices (all 22 READ on both; 19 and 18 AGREE), a daily record from 5 to 7 November (7 runs and one refusal), and IBM's dual-unitary kicked-Ising chain (arXiv:2411.00765) at N = 51 and N = 71 sites, where the raw decay rates agree with cos 2h at all three h (largest |k - cos 2h|: 0.0045 at N = 51, 0.0017 at N = 71) and every depth up to t = 24 (1,715 CZ) is readable. Local failure regions, not problem size, set several outcomes. A correction built from the job's own controls, declared in advance, reduced the error on both devices but did not make the answers agree across problems. No quantum advantage is claimed: every answer here is computed exactly on a classical machine. We acknowledge the use of IBM Quantum services for this work.

•Nov 8, 2026•0 citations

Folding a Ring onto Heavy-Hex: Formula-Built Schedules with Provable Gate Counts, and What Their Signal Shows on Twins and Devices

TN.Holdings

A ring of n CZ gates per layer, the entangling pattern of many variational classifiers, doesnot fit IBM’s heavy-hex devices unless n is a cycle length of the lattice, and heavy-hex hascycles of 12,20,24,28,... qubits only. We compare three schedules that a compiler can buildby formula, with no layout search and without reading the day’s calibration: the zipper (thering folded onto a line; 5n/2 − 4 two-qubit gates per layer for even n, (5n − 7)/2 for odd n),holes (the ring on a longer cycle, its spare qubits moved as |0⟩; n + 2(m − n)), and pendants(the ring on a shorter cycle, its extra nodes hung off it; n + (n − m), on exactly n qubits). Weprove that the pendant schedule’s count is the least possible for layered schedules on the cycleand its leaves in a stated routing model (exhaustive search finds the same for interleaved onesin small cases), that the hole schedule’s count is least within its class but not in general, andrecall that no cheaper schedule than the zipper was found by exhaustive searches at n = 4–12and 14. A rule that takes the fewest extra gates picks pendants at n = 13–17 and holes atn = 8–11 and 18–19 on IBM’s Heron devices; the rule is in PSF-Zero’s AI front end a26. Onnoise models, in declared matches against the zipper at 16 layers over four random classifiers,the rule won at every n = 13–18 on two twins. Its lead was 1.5–8.2× the zipper’s kept signal onthe twin of a 13-day-old snapshot and only 1.07–1.28× (about even at n = 18) on the twin of areal day’s calibration, the kind that had predicted the zipper on hardware: the old snapshot waspessimistic for the zipper. On two devices (ibmkingston, ibmfez) the pendant schedule keptthe signal its day’s twin predicted (0.90±0.12 against 0.77; 0.50±0.10 against 0.59), in two jobsof 7s each; these are coarse tests and no device comparison with the zipper was made. AgainstQiskit’s search (level 2 on the same twin) the rule won on every classifier at every n simulated(n = 13–16, 18): Qiskit’s circuits kept −0.05–0.64 of the signal against 0.61–0.82. The bare ringkept 0.78–0.86 at n = 20 and 0.81–0.83 at n = 24. On one 20-qubit cycle, two pendants keptnearly what the bare ring kept (0.75–0.84 against 0.78–0.86) and clearly more than two holes(0.61–0.68); with the classifiers and qubits fixed and only the schedule changed, the bare ringkept 1.44–1.81 times the zipper’s signal and two holes 1.13–1.20 times. Error-product estimatesof the kept signal did not rank the schedules. We give the benchmark number and the deviceuser’s number side by side throughout

•Oct 26, 2026•1 citation

Training quantum neural networks for Tl-limited hardware: what the qg symmetry filter does and does not do

Vicente Humberto Monteverde

A quantum neural network (QNN) built from gates that conserve the Hamming weight keeps its data in one weight sector, so the shots that left the sector can be discarded at readout. This is the qg filter of the qang library [1, 2]. We show that, with equal amplitude damping (T1) on every qubit, the filtered readout of such a network is exactly the noiseless one in any weight sector. The kept fraction of shots is (1 − γ)k d for weight k and depth d, and training under T1 with the filter gives exactly the parameters of noiseless training. We then measure what this is worth on four small classification datasets. Every result is reported with the filter and without it, and every prediction was committed before its run, failures included. Over 60 runs per model, a network trained on a simulator and run under T1 gains +2.5 points from the filter at weight 1 (95% CI [0.8, 4.1]) and +16.5 points at weight 2 ([11.8, 21.2]), recovering its noiseless accuracy exactly. A network trained under the calibrated noise without the filter catches up (diflerences of 0.1 points), and the filter corrects neither dephasing nor unequal T1 across qubits, although alone it loses less than one point up to a ±80% spread of 1/T1. With noise-aware training and a richer readout, the model without the filter was 1.2 points better in one study, but over 60 new runs that gap shrank to +0.5 points and was not significant. The weight-2 deficit comes from the pair-product encoding, not from the readout; with an encoding that fills the whole sector and a full-rank readout, weight 2 reaches the weight-1 accuracy. On noise models of IBM and IonQ devices the filter cuts the error of the measured qgZ by 2.7-3.6×; on the IonQ Forte-1 noise model over 189 inputs of four datasets the filtered model is 1.1 points more accurate than the raw one, and calibrating the errors the filter keeps with five echo circuits per model halves the remaining error of the decision value: decisions that difler from the noiseless model fall from 15 (raw) and 7 (filter) to 3 over 945 readings on five device noise models. With a faithful weight-2 encoding the advantage of the unfiltered model under noise-aware training disappears, and a classifier given both the filtered and the raw features gains nothing over the filter alone. Over 513 readout configurations, including kept fractions down to 2.6×10−7, unequal T1 and dephasing, the filter has the lower readout error whenever about 20 shots are kept, and a closed-form rule predicts the winner. These models are classically simulable and no quantum advantage is claimed. The filter is a way to make noise-free training valid under relaxation, at a shot cost known in advance. The models are released as qang. qml (version 0.6.13, pip install qang).

•Oct 5, 2026•0 citations

Symmetry-filtered read-out on IBM quantum devices: the qg filter on twenty device noise models through IBM Runtime

Vicente Humberto Monteverde

Circuits built from gates that conserve the Hamming weight keep their state in one weight sector, so the shots that left the sector can be discarded at read-out: the qg filter of the open-source library qang [1], a form of symmetry verification [2, 3]. We make it usable on IBM devices through IBM Runtime and Qiskit, and measure it on the noise models of twenty IBM devices. The library adds a Qiskit transpiler pass that decides whether the filter applies, read-out correction from each device's published per-qubit errors, qg read-outs of IBM Runtime Sampler and Estimator results, zero-noise extrapolation by gate folding that stacks with the filter, a prediction of the filter's kept fraction from the calibration, qubit selection with that prediction, a shot planner, and the weight-conserving quantum neural network (QNN) of qang as one parameterized job and as a qiskit-machine-learning SamplerQNN. Every prediction was committed before its run and is reported as passed or failed. On ten five-qubit weight-1 circuits the filter lowers the error of the measured qgZ = ⟨Z⟩ from 0.065-0.104 to 0.018-0.023 on three devices (0.013-0.022 with read-out correction), and filter plus zero-noise extrapolation reaches 0.011-0.015. On seventeen further devices never used before, the filter cuts the error 2.4-8.9× on every one (median 3.1× with read-out correction). Choosing the qubits by the predicted kept fraction keeps 6-17 points more shots than Qiskit's default layout on chain circuits, avoids a qubit published as reading 1 always, and the shot planner met a 0.02 target precision on three devices. For classification the gain is small: the QNN recovers 5-6 of the 10 decisions noise costs it. Reading a known five-component principal-component spectrum out of the data's density matrix in the weight-1 sector, filter plus read-out correction lowers the error 4.6× (median over twenty-one devices) and orders the spectrum correctly on 18 of 21 (13 raw). In Trotter simulation of the XY chain the filter halves the best error and moves the optimal number of Trotter steps deeper (16 of 21 devices), and in a quantum walk it turns a raw spreading indistinguishable from a classical random walk (exponent 0.87) into ballistic spreading (1.84). With two or three interacting excitations (XXZ chain) the gain falls from 5.6× to 2.6×: larger sectors let more of the noise through. Two interoperability bugs were found, one in qiskit-machine-learning and one in the library. All results are on device noise models; no hardware run is reported.

•Oct 13, 2026•0 citations

Readable Quantum Dynamics Without Error Mitigation: Dual-Unitary Chains up to 91 Qubits and 3D Lattice Tests on IBM Heron Devices

TN.Holdings

Paper 7 of this series carried user-written two-dimensional circuits through an operating layer (psf-os) that reads no calibration-quality number, adds its own controls to each job and applies no error mitigation. Here the same policy is carried to a chain that no fixed shape can hold and to three-dimensional lattices that the heavy-hex coupling graph cannot embed. IBM's dual-unitary kicked-Ising chain (arXiv:2411.00765) at its largest width, N = 91, read raw on ibm_marrakesh: the decay rates agree with cos 2h at all three h (largest |k - cos 2h| 0.0037) and every depth up to t = 24 (2 205 CZ) is readable, on a chain placed by a fixed rule round the failed elements; a second run on ibm_fez, with the ruler h = 0 in both jobs, again found all three rates in agreement, while h = 0.15 sat +0.0037 above cos 2h on both devices. In three dimensions, a place rule that judges the place with the highest own mirror replaced a gate that had refused every place at 48 and 72 sites; kicked Ising on 4 x 4 x 3 read at every angle and depth, Heisenberg XXX on 4 x 2 x 2 read with a Neel retention of 0.72 to 0.51. Two new layouts cut the transport CZ of the Heisenberg step by 42%, and were compared with the old one on the same places of two devices; an Ising layout with 46% fewer CZ came closer to the exact answer at the judged 4 x 2 x 2 place but farther on 4 x 3 x 2. On the way we changed our main measure for such comparisons, from the retention of an order parameter to a normalised per-site error, because noise that slows the dynamics can raise the retention. Pair creation in the Schwinger model read in 16 of 18 circuits and beat both trivial guesses in all 18. A light-cone test failed as predicted: every front arrived late, while the shape of the spread was reproduced. Every device result re-scores to the same numbers from its committed raw counts (23 of 23). No quantum advantage is claimed: every answer is computed exactly on a classical machine.

•Nov 11, 2026•0 citations
PSF-Zero places a quantum classifier's entangling ring on hardware by formula, without a layout search and without reading the day's calibration; Qiskit's preset pass manager searches, using the calibration. We ask which of the two gives the better trained classifier, and report what it took to answer that question honestly. The model is a 10-qubit, 16-layer data re-uploading classifier trained by a finite-difference optimiser on a noise model of ibm_marrakesh (the twin) and tested on an independently perturbed noise model (the stand-in); one day of hardware runs is included. Four effects can each produce a winner without any difference in learning: (i) the task's answer sat at the untrained starting point: on the device day's task the untrained model is already 0.934 accurate (noiseless), and training lowered both compilers' models to 0.762 (PSF-Zero) and 0.730 (Qiskit); (ii) a readout offset makes a circuit answer one class on 84-94 % of balanced test points, which moves accuracy while leaving the model unchanged; (iii) arms with different output scales lose different amounts to a fixed shot count; (iv) the twin's scale was off by up to 2.3x on hardware. With the task moved far from the start and the rules fixed before each run, PSF-Zero's newest ring schedule (the zipper, 336 CZ and 11.7 us at 16 layers) and Qiskit (506 CZ, 48.2 us, 14 active qubits) gave classifiers that the declared tests could not tell apart: on stand-in accuracy the zipper won on one of three teachers at each budget, below the declared two of three, and on the trained model itself neither arm won on any teacher. Formula placement could not be told apart from the search while using a third fewer entangling gates and a quarter of the duration; it did not beat it. We give the vanity and the essential number side by side throughout, mark what was declared before the data and what was found after, and close with a checklist for compiler comparisons through trained quantum models. In one preliminary hardware job on ibm_kingston, the zipper's circuits followed a twin built from that day'scalibration (scale 1.00, correlation 0.98), while Qiskit's lost most of their signal and came out inverted on theread-out qubit; the cause is not known. Code and records: https://github.com/TN-Holdings-LLC/psf-zero (paper and figure script) and https://github.com/TN-Holdings-LLC/TOQB (runs, records, declarations).
A quantum neural network (QNN) built from gates that conserve the Hamming weight keeps its data in one weight sector, so the shots that left the sector can be discarded at readout. This is the qg filter of the qang library [1, 2]. We show that, with equal amplitude damping (T1) on every qubit, the filtered readout of such a network is exactly the noiseless one in any weight sector. The kept fraction of shots is (1 − γ)k d for weight k and depth d, and training under T1 with the filter gives exactly the parameters of noiseless training. We then measure what this is worth on four small classification datasets. Every result is reported with the filter and without it, and every prediction was committed before its run, failures included. Over 60 runs per model, a network trained on a simulator and run under T1 gains +2.5 points from the filter at weight 1 (95% CI [0.8, 4.1]) and +16.5 points at weight 2 ([11.8, 21.2]), recovering its noiseless accuracy exactly. A network trained under the calibrated noise without the filter catches up (diflerences of 0.1 points), and the filter alone corrects neither dephasing nor unequal T1 across qubits, although it loses less than one point up to a ±80% spread of 1/T1. Unequal T1 is corrected by training UÏtl the filter under the calibrated decay rates: the kept state depends only on their ratios, so the model lands within 0.5 points of noiseless at a ±100% spread, ignores a 10% calibration error and, at weight 2, loses 0.7 points under a 1.5-fold drift of all T1 that costs the raw noise-aware model 10. With noise-aware training and a richer readout, the model without the filter was 1.2 points better in one study, but over 60 new runs that gap shrank to +0.5 points and was not significant. The weight-2 deficit comes from the pair-product encoding, not from the readout; with an encoding that fills the whole sector and a full-rank readout, weight 2 reaches the weight-1 accuracy. On noise models of IBM and IonQ devices the filter cuts the error of the measured qgZ by 2.7-3.6×; on the IonQ Forte-1 noise model over 189 inputs of four datasets the filtered model is 1.1 points more accurate than the raw one, and calibrating the errors the filter keeps with five echo circuits per model halves the remaining error of the decision value: decisions that difler from the noiseless model fall from 15 (raw) and 7 (filter) to 3 over 945 readings on five device noise models. With a faithful weight-2 encoding the advantage of the unfiltered model under noise-aware training disappears, and a classifier given both the filtered and the raw features gains nothing over the filter alone. Over 513 readout configurations, including kept fractions down to 2.6 ×10−7, unequal T1 and dephasing, the filter has the lower readout error whenever about 20 shots are kept, and a closed-form rule predicts the winner. The same holds beyond the first model: at 8 qubits a weight-2 network keeps its noiseless 0.965 with the filter and drops to 0.705 without it; a five-class classifier (one class per qubit) keeps its noiseless accuracy under T1 with the filter and loses 15-18 points without it on five-class digits; training from 100 or 1000 shots per step with the filter is 1.0-1.5 points ahead of training without it; and the filtered gradient equals the noiseless one, while the raw one is K times smaller at weight 1. On five device noise models the echo calibration turns class confusions back: flipped decisions of the five-class model fall from 51 (raw) and 24 (filter) to 15 over 500 readings. The echo helped only where the filtered error was well above shot noise, and it still removed a third of the filtered error with a 28 × 28 matrix at 8 qubits. These models are classically simulable and no quantum advantage is claimed. The filter is
A user writes, in plain Qiskit, a circuit whose answer is known exactly. A small operating layer (psf-os) places it on an IBM Heron r2 device by translation, without reading any calibration-quality number, adds its own controls to the same job, and runs it with no error mitigation. We report how close the raw answers come and how that is shown. Two points of the report are of use beyond our own numbers. First, an absolute criterion |M - M_exact| <= c is biased in both directions: on 5 November a problem that DIFFERS under c = 0.05 kept 0.82-0.86 of its signal, as much as a problem that AGREES (0.73-0.85), and a small-answer problem AGREES while keeping only 0.73 of its signal. We therefore print the declared verdict first and the retention R = M/M_exact beside it on every row. Second, every number can be checked: answers are sealed by hash before each run, raw counts and job identifiers are committed, and one script re-scores all 13 device results of this paper from the raw counts and finds every number the same. The results: 22 practice circuits on two devices (all 22 READ on both; 19 and 18 AGREE), a daily record from 5 to 7 November (7 runs and one refusal), and IBM's dual-unitary kicked-Ising chain (arXiv:2411.00765) at N = 51 and N = 71 sites, where the raw decay rates agree with cos 2h at all three h (largest |k - cos 2h|: 0.0045 at N = 51, 0.0017 at N = 71) and every depth up to t = 24 (1,715 CZ) is readable. Local failure regions, not problem size, set several outcomes. A correction built from the job's own controls, declared in advance, reduced the error on both devices but did not make the answers agree across problems. No quantum advantage is claimed: every answer here is computed exactly on a classical machine. We acknowledge the use of IBM Quantum services for this work.
A ring of n CZ gates per layer, the entangling pattern of many variational classifiers, doesnot fit IBM’s heavy-hex devices unless n is a cycle length of the lattice, and heavy-hex hascycles of 12,20,24,28,... qubits only. We compare three schedules that a compiler can buildby formula, with no layout search and without reading the day’s calibration: the zipper (thering folded onto a line; 5n/2 − 4 two-qubit gates per layer for even n, (5n − 7)/2 for odd n),holes (the ring on a longer cycle, its spare qubits moved as |0⟩; n + 2(m − n)), and pendants(the ring on a shorter cycle, its extra nodes hung off it; n + (n − m), on exactly n qubits). Weprove that the pendant schedule’s count is the least possible for layered schedules on the cycleand its leaves in a stated routing model (exhaustive search finds the same for interleaved onesin small cases), that the hole schedule’s count is least within its class but not in general, andrecall that no cheaper schedule than the zipper was found by exhaustive searches at n = 4–12and 14. A rule that takes the fewest extra gates picks pendants at n = 13–17 and holes atn = 8–11 and 18–19 on IBM’s Heron devices; the rule is in PSF-Zero’s AI front end a26. Onnoise models, in declared matches against the zipper at 16 layers over four random classifiers,the rule won at every n = 13–18 on two twins. Its lead was 1.5–8.2× the zipper’s kept signal onthe twin of a 13-day-old snapshot and only 1.07–1.28× (about even at n = 18) on the twin of areal day’s calibration, the kind that had predicted the zipper on hardware: the old snapshot waspessimistic for the zipper. On two devices (ibmkingston, ibmfez) the pendant schedule keptthe signal its day’s twin predicted (0.90±0.12 against 0.77; 0.50±0.10 against 0.59), in two jobsof 7s each; these are coarse tests and no device comparison with the zipper was made. AgainstQiskit’s search (level 2 on the same twin) the rule won on every classifier at every n simulated(n = 13–16, 18): Qiskit’s circuits kept −0.05–0.64 of the signal against 0.61–0.82. The bare ringkept 0.78–0.86 at n = 20 and 0.81–0.83 at n = 24. On one 20-qubit cycle, two pendants keptnearly what the bare ring kept (0.75–0.84 against 0.78–0.86) and clearly more than two holes(0.61–0.68); with the classifiers and qubits fixed and only the schedule changed, the bare ringkept 1.44–1.81 times the zipper’s signal and two holes 1.13–1.20 times. Error-product estimatesof the kept signal did not rank the schedules. We give the benchmark number and the deviceuser’s number side by side throughout
A quantum neural network (QNN) built from gates that conserve the Hamming weight keeps its data in one weight sector, so the shots that left the sector can be discarded at readout. This is the qg filter of the qang library [1, 2]. We show that, with equal amplitude damping (T1) on every qubit, the filtered readout of such a network is exactly the noiseless one in any weight sector. The kept fraction of shots is (1 − γ)k d for weight k and depth d, and training under T1 with the filter gives exactly the parameters of noiseless training. We then measure what this is worth on four small classification datasets. Every result is reported with the filter and without it, and every prediction was committed before its run, failures included. Over 60 runs per model, a network trained on a simulator and run under T1 gains +2.5 points from the filter at weight 1 (95% CI [0.8, 4.1]) and +16.5 points at weight 2 ([11.8, 21.2]), recovering its noiseless accuracy exactly. A network trained under the calibrated noise without the filter catches up (diflerences of 0.1 points), and the filter corrects neither dephasing nor unequal T1 across qubits, although alone it loses less than one point up to a ±80% spread of 1/T1. With noise-aware training and a richer readout, the model without the filter was 1.2 points better in one study, but over 60 new runs that gap shrank to +0.5 points and was not significant. The weight-2 deficit comes from the pair-product encoding, not from the readout; with an encoding that fills the whole sector and a full-rank readout, weight 2 reaches the weight-1 accuracy. On noise models of IBM and IonQ devices the filter cuts the error of the measured qgZ by 2.7-3.6×; on the IonQ Forte-1 noise model over 189 inputs of four datasets the filtered model is 1.1 points more accurate than the raw one, and calibrating the errors the filter keeps with five echo circuits per model halves the remaining error of the decision value: decisions that difler from the noiseless model fall from 15 (raw) and 7 (filter) to 3 over 945 readings on five device noise models. With a faithful weight-2 encoding the advantage of the unfiltered model under noise-aware training disappears, and a classifier given both the filtered and the raw features gains nothing over the filter alone. Over 513 readout configurations, including kept fractions down to 2.6×10−7, unequal T1 and dephasing, the filter has the lower readout error whenever about 20 shots are kept, and a closed-form rule predicts the winner. These models are classically simulable and no quantum advantage is claimed. The filter is a way to make noise-free training valid under relaxation, at a shot cost known in advance. The models are released as qang. qml (version 0.6.13, pip install qang).
Circuits built from gates that conserve the Hamming weight keep their state in one weight sector, so the shots that left the sector can be discarded at read-out: the qg filter of the open-source library qang [1], a form of symmetry verification [2, 3]. We make it usable on IBM devices through IBM Runtime and Qiskit, and measure it on the noise models of twenty IBM devices. The library adds a Qiskit transpiler pass that decides whether the filter applies, read-out correction from each device's published per-qubit errors, qg read-outs of IBM Runtime Sampler and Estimator results, zero-noise extrapolation by gate folding that stacks with the filter, a prediction of the filter's kept fraction from the calibration, qubit selection with that prediction, a shot planner, and the weight-conserving quantum neural network (QNN) of qang as one parameterized job and as a qiskit-machine-learning SamplerQNN. Every prediction was committed before its run and is reported as passed or failed. On ten five-qubit weight-1 circuits the filter lowers the error of the measured qgZ = ⟨Z⟩ from 0.065-0.104 to 0.018-0.023 on three devices (0.013-0.022 with read-out correction), and filter plus zero-noise extrapolation reaches 0.011-0.015. On seventeen further devices never used before, the filter cuts the error 2.4-8.9× on every one (median 3.1× with read-out correction). Choosing the qubits by the predicted kept fraction keeps 6-17 points more shots than Qiskit's default layout on chain circuits, avoids a qubit published as reading 1 always, and the shot planner met a 0.02 target precision on three devices. For classification the gain is small: the QNN recovers 5-6 of the 10 decisions noise costs it. Reading a known five-component principal-component spectrum out of the data's density matrix in the weight-1 sector, filter plus read-out correction lowers the error 4.6× (median over twenty-one devices) and orders the spectrum correctly on 18 of 21 (13 raw). In Trotter simulation of the XY chain the filter halves the best error and moves the optimal number of Trotter steps deeper (16 of 21 devices), and in a quantum walk it turns a raw spreading indistinguishable from a classical random walk (exponent 0.87) into ballistic spreading (1.84). With two or three interacting excitations (XXZ chain) the gain falls from 5.6× to 2.6×: larger sectors let more of the noise through. Two interoperability bugs were found, one in qiskit-machine-learning and one in the library. All results are on device noise models; no hardware run is reported.
Paper 7 of this series carried user-written two-dimensional circuits through an operating layer (psf-os) that reads no calibration-quality number, adds its own controls to each job and applies no error mitigation. Here the same policy is carried to a chain that no fixed shape can hold and to three-dimensional lattices that the heavy-hex coupling graph cannot embed. IBM's dual-unitary kicked-Ising chain (arXiv:2411.00765) at its largest width, N = 91, read raw on ibm_marrakesh: the decay rates agree with cos 2h at all three h (largest |k - cos 2h| 0.0037) and every depth up to t = 24 (2 205 CZ) is readable, on a chain placed by a fixed rule round the failed elements; a second run on ibm_fez, with the ruler h = 0 in both jobs, again found all three rates in agreement, while h = 0.15 sat +0.0037 above cos 2h on both devices. In three dimensions, a place rule that judges the place with the highest own mirror replaced a gate that had refused every place at 48 and 72 sites; kicked Ising on 4 x 4 x 3 read at every angle and depth, Heisenberg XXX on 4 x 2 x 2 read with a Neel retention of 0.72 to 0.51. Two new layouts cut the transport CZ of the Heisenberg step by 42%, and were compared with the old one on the same places of two devices; an Ising layout with 46% fewer CZ came closer to the exact answer at the judged 4 x 2 x 2 place but farther on 4 x 3 x 2. On the way we changed our main measure for such comparisons, from the retention of an order parameter to a normalised per-site error, because noise that slows the dynamics can raise the retention. Pair creation in the Schwinger model read in 16 of 18 circuits and beat both trivial guesses in all 18. A light-cone test failed as predicted: every front arrived late, while the shape of the spread was reproduced. Every device result re-scores to the same numbers from its committed raw counts (23 of 23). No quantum advantage is claimed: every answer is computed exactly on a classical machine.