WorksPhysics and Astronomy
Quantized memory of measurement records near the light cone of quantum circuits
Abstract
Quantum processors now measure qubits in the middle of a circuit, and the stream of outcomes is the raw output of error correction, measurement-based state preparation and monitored dynamics. For a scrambling circuit, predicting the statistics of this stream is as hard as simulating the processor, so there is no exact reference to check it against once the device outgrows classical simulation. We show that near the light cone the problem becomes exactly solvable, for any two-qubit gate. In a brickwork circuit with one polarized qubit in a maximally mixed background, measurements along the edge of the light cone form a two-state Markov chain, and measurements sites inside the edge are generated by qubits that move with it, while the rest of the chain, however large, acts on them as a memoryless bath. The memory of the record is therefore quantized, one bit on the edge and at most one more for every site of displacement, a bound that generic gates attain. The same causal structure determines which hardware noise the record can see, removes that sensitivity at no gate cost, and yields a tomography-free witness of dual-unitarity. On a 156-qubit superconducting processor, a demonstration with its analysis specified in advance finds the predicted memory one and two sites inside the light cone, and none where the theory forbids it. This gives dynamic circuits an exact reference at any system size, and shows that the collision models of open-system theory arise exactly inside chaotic dynamics.
Introduction
A quantum processor that measures some of its qubits in the middle of a circuit produces a stream of classical bits. That stream is the raw output of quantum error correction, where syndromes are measured round after round, of measurement-based preparation of long-range entangled states, and of the monitored dynamics in which measurements compete with scrambling. In each case the record is what the device reports, and whether the device works is a question about the statistics of that record. For a scrambling circuit the question is hard to answer. Predicting the joint distribution of a sequence of mid-circuit outcomes is in general as hard as simulating the circuit that produced them, so the reference against which a record could be checked disappears precisely where it is most needed, once the device runs circuits beyond the reach of classical simulation [1]. This paper asks whether any part of that record can be predicted exactly, for an arbitrary gate and at any system size, and answers that near the light cone of the circuit all of it can.
The tools that exist address the problem only in part. Randomized and cycle benchmarking characterize gates, and dedicated protocols now measure the errors of mid-circuit measurements and the crosstalk they cause on spectator qubits [2, 3, 4, 5, 6], but none of them provides an exact reference for the dynamics of a scrambling, non-Clifford circuit as a whole. Exactly solvable circuits do, but only once the gate has been fine tuned. Dual-unitary circuits admit closed-form correlation functions at any depth [7, 8, 9] and have served as a benchmark for a 91-qubit processor [10]. Their generalizations are organized as hierarchies of conditions on the gate [11, 12], exact hidden-Markovian subsystem dynamics has been engineered for gates that satisfy a solvability condition [13], and the recently introduced influence-solvable circuits contain most of these families as special cases [14]. Some tractability, however, owes nothing to the gate. In a brickwork circuit at infinite temperature, the reduced state of a few sites that ride the light cone evolves under a channel whose dimension does not depend on the length of the chain, whatever the gates are [7, 15, 16], and two-point functions on the edge of the cone follow from unitarity alone. Measurements have entered this picture only in special settings. In monitored circuits of XYZ gates, post-selected light-cone correlations can be obtained from typical correlations of a different unitary circuit [17], and dual-unitary circuits measured in a unitary error basis remain solvable, although every outcome is then equally likely and the record carries no information [18]. The object a device actually produces, the joint distribution of a sequence of mid-circuit outcomes with the back-action of every earlier measurement included, had no exact description for an arbitrary gate.
Our answer rests on a change of frame. Viewed along its diagonals, a brickwork circuit is a cascade of left-moving lines in which information passes from each line only to the line behind it. An observer who reads the chain sites inside the edge of the light cone, moving with it, interacts with of these lines and with nothing else, and for this observer the remaining qubits of a scrambling chain form an exactly memoryless bath (Theorem 3). On the edge itself the record is a two-state Markov chain with transfer matrices, for any gate and any system size (Theorem 1 and Corollary 1). Protection against noise follows the same causal structure. Channels acting behind the measured ray are invisible, channels acting ahead of it are invisible only if they are unital, and a Pauli twirl applied to each site just before it joins the ray restores the exact law at no gate cost (Theorem 2). The frame also fixes how much the record can remember. Its memory, measured by the rank of its joint distribution across a cut in time, is at most bits (Corollary 2), and generic gates reach the cap, as we prove for displacements of up to six sites (Proposition 1). One site inside the edge the measured qubit is a passing ancilla, the co-moving qubit is never reset, and the record is not Markovian at any finite order, even though two bits generate it exactly. This is the structure of a collision model [19] or of a cascaded quantum system [20, 21], arising here as an exact property of many-body dynamics rather than as a modelling assumption. Readouts slower than the edge fall behind through more and more lines, and their memory grows with the length of the record. Dual-unitary gates are exceptional. Their records carry at most one bit at every displacement we examined and none at all one site inside the edge, which turns that record into a witness of dual-unitarity that single-qubit noise cannot fake.
Three consequences reach beyond the model. For the validation of quantum hardware, the record law is an exact reference for a scrambling circuit with mid-circuit measurements whose cost does not grow with the number of qubits, in the regime of dynamic circuits in which such references have been missing, together with a precise account of the noise it can see and a tomography-free test of the executed gate. For the theory of many-body dynamics, it identifies an exact structure in chaotic circuits that does not depend on the gate, and it turns the question of where the output of monitored dynamics becomes hard to predict into a count of light-cone lines. For open-system theory, it shows that the collision models used to describe system-bath dynamics arise exactly inside a generic scrambling chain, with the bath’s memory quantized by geometry.
We test these predictions on the superconducting processor ibm_fez, accessed through the cloud, in three steps. Twelve records taken before the present analysis obey the Markov structure once the transition kernel is allowed to vary from step to step, as it must on a device in which every step uses a different pair of physical qubits. A single interleaved demonstration, analysed against acceptance criteria set in advance, then shows that relaxation injected ahead of the ray corrupts the record only in the absence of the late twirl, and that the witness places the executed dual-unitary gate within 0.011 rad of the dual-unitary manifold. Finally, the records of that demonstration revealed the memory predicted one site inside the edge, and a second demonstration, whose analysis was specified before its data existed, confirms it for two scrambling gates one and two sites inside the edge, together with its absence wherever the theory predicts none. Repetitions of the registered demonstration on two further processors find the memory wherever it is predicted and its absence wherever it is not.
The records studied here are classical in a precise sense, because the record law is its own sampler. They belong to the effectively classical temporal correlations identified in the influence-matrix literature [22, 23, 24], and in the language of process tensors they have a memory of fixed size [25]. Dual-unitary circuits are known to act as perfectly Markovian baths for a subsystem at rest [22], and temporal entanglement along tilted paths has been analysed for solvable and chaotic circuits after quenches from pure states [26]. Our statement is complementary. At infinite temperature and for an arbitrary gate, the bath seen from a frame that moves with the edge is memoryless, and frames differ only in how many qubits the observer has to carry along. The value of the construction lies in an exact reference that does not degrade with system size, not in any computational advantage.
The paper is organized as follows. Section II derives the record law on the edge, Sec. III gives the causal account of noise, and Sec. IV develops the co-moving frame, the memory of the record and the dual-unitarity witness. § 1.4 summarizes the checks against exact simulation, Sec. VI reports the demonstrations on hardware, and Sec. VII discusses scope and open questions. Proofs, the documentation of the cloud platform and the details of the demonstrations are collected in the appendices.
Measurement records on the light-cone edge
Consider a chain of qubits prepared in the state
in which a single site , which we call the germ, carries the Bloch vector and every other site is maximally mixed. The chain evolves under a brickwork of two-qubit gates, with even bonds acting on odd half-steps and odd bonds on even half-steps. After half-step we measure site in the basis and obtain a record with . The measured sites trace the left edge of the light cone of the germ, one site per half-step, which is the largest speed the circuit allows. We call this sequence of sites a rail (Figure 1).

Figure 1. Measurement circuit for , and . The germ site is prepared by a rotation , and every other site is randomized from shot to shot (), which realizes the maximally mixed background. Gates act in brickwork layers, and after half-step the site is measured in . Shaded gates lie on the edge of the light cone. As drawn, every site is randomized at , as in the recorded rails of Sec. VI A. In the late-twirl protocol of Theorem 2(iii), each rail site is instead randomized immediately before its shaded gate.
Two matrices describe the action of a single gate on the edge. For a gate on the sites , with the first tensor factor, the Heisenberg-picture light-cone channel has the real matrix elements
The Schrödinger-picture edge propagator, which carries the state of the right site over to the left site while the left site enters maximally mixed, , has the Bloch matrix (Appendix A). The transpose is easy to overlook and matters in practice. A germ polarized along cannot tell from , so a numerical check of the law needs terms with transverse components (Table I).
| Test | error |
| (2), 5 Haar gates 4 germs | |
| (2), inhomogeneous in time | |
| (4) against (2) | |
| -independence, | |
| Thm. 2(i,ii), unital depolarizing off ray | |
| Thm. 2(i), non-unital , , behind ray | |
| Thm. 3, window form, ( states) | |
| Thm. 3, co-moving against window form, | |
| Thm. 3, , executed schedule | |
| Thm. 3, , everywhere | |
| Thm. 3, every gate of the chain different, | |
| Thm. 3, readout at and | |
| Prop. 1, rank at , Haar gates | exactly |
| Thm. 2(iii), executed schedule, , | |
| Thm. 2(iii), (5), everywhere | |
| Thm. 2(iii), kernel under injected idle, | |
| , , DU gate, noise and readout error | |
| Wrong sublattice | |
| (2) at readout inside cone | |
| Thm. 3 with applied to data | |
| Untransposed law, transverse germ | |
| Untransposed law, germ (fails silently) | |
| twirl schedule, all slices |
Table I. Verification against full density-matrix simulation of system plus readout, maxima over the stated ensembles. Executed-schedule checks include damping in every time slice. The controls in the lower block are designed to falsify the claims and must fail, one of them silently. DU denotes the dual-unitary gate.
Theorem 1 (Edge record). Suppose the germ sits on the sublattice paired with the first brickwork layer, so that and share a gate at . Then the record distribution is
where is built from the gate on the edge at half-step , selects the identity component, and each measurement updates the ray by
Nothing in Eq. (2) refers to the gates beyond their unitarity, and nothing refers to the length of the chain. A likelihood costs products of matrices, and the gates may differ from one half-step to the next. The reason is causal (Figure 2). A brickwork layer moves information by at most one site, so the edge is the only path of maximal speed from the germ to each measurement, and every other path arrives too late. The back-action of a measurement obeys the same bound. As a result, the site that enters each edge gate from the left lies strictly outside the forward light cone of the germ and of all earlier measurements, and it arrives maximally mixed and uncorrelated with the ray (Appendix § 1.5.2). In this setting the maximally mixed background does the work that dual-unitarity does elsewhere. The parity condition is essential. With the germ on the other sublattice, the germ

Figure 2. Causal structure of the circuit, with gates as nodes and light-cone dependence as arrows. The germ reaches each record bit along a single path, through the edge gates (red). Grey nodes lie behind the ray and have no directed path to any , so no channel acting there can change the record [Theorem 2(i)]. Orange nodes lie ahead of the ray and reach the record only through the incoming leg of an edge gate, which arrives maximally mixed. Unitary gates and unital channels there leave the record unchanged, whereas non-unital channels do not [Theorem 2(ii)]. The graph makes no reference to the content of the gates, which is why Theorem 1 holds for any unitary.
and the first measured site belong to different gates, and Eq. (2) fails at the percent level (Table I).
For expectation values this is standard. Unitarity contracts every leg outside the backward light cone, so two-point functions on the edge are products of for any brickwork circuit [7, 16]. A record is a different object. A projective measurement is not a contraction, is not unital, and every outcome carries the back-action of the measurements that preceded it. What Theorem 1 adds is that the joint distribution of all outcomes nevertheless remains an ordered product of matrices, for any gate.
The structure simplifies further after the first measurement. The update projects the ray onto and removes its and components, so only the first outcome sees the transverse part of the germ, and from then on the hidden state is simply the last outcome (Appendix § 1.5.3).
Corollary 1 (Monitored chain). Under the hypotheses of Theorem 1,
with , and .
For unitary gates the bias vanishes and is the monitored decay rate, so that .
For the dual-unitary gate used in our demonstrations , and the light-cone correlations at lags one to three are 0.607, 0.368 and 0.224. The gate-invariant content of lies in its singular values, with for this gate, whereas its eigenvalues, , change under local unitaries and carry no invariant meaning. A single-qubit channel acting on the rail changes only , and a readout assignment error turns the chain into a hidden two-state chain. Equation (4) with free, step-dependent parameters and assignment errors is therefore the most general record that the causal structure allows, and it is the null hypothesis we test on hardware in Sec. VI A.
Noise and causality
Real devices add noise to every gate, idle period and readout, and an exact law is only useful if one knows which of these errors it is blind to. The light-cone structure answers this question sharply.
Theorem 2 (Causal dichotomy and late twirl). Let the ray occupy site after half-step . (i) Arbitrary completely positive trace-preserving maps, unital or not, acting on sites behind the ray leave the record distribution exactly invariant. (ii) Maps acting on sites , which have yet to join the rail, leave it invariant if and only if they are unital. (iii) If each incoming site receives a uniformly random Pauli operator immediately before its edge gate, the noisy record law
holds exactly, whatever happened to the incoming sites before they joined the ray. Here is an arbitrary single-qubit channel on the rail site and accounts for a readout assignment error .
Parts (i) and (ii) can be read off Figure 2. A site behind the ray never re-enters the backward light cone of a later measurement, so nothing done to it can matter. A site ahead of the ray reaches the record only as the incoming leg of an edge gate, and Theorem 1 needs that leg to arrive maximally mixed, which a channel guarantees exactly when it is unital (Appendix D). Part (ii) is the one real vulnerability of the protocol, and it is not an academic one. On hardware the qubits that will later join the rail idle through every earlier measurement window, and amplitude damping during these idle periods is non-unital. In a full density-matrix simulation with damping on every site, the record deviates from Eq. (2) at first order in the damping rate, and no law of the form of Eq. (2) reproduces the deviation (Figure 3).
Figure 3. The causal dichotomy of Theorem 2 in full density-matrix simulation. Deviation of the record distribution from Eq. (2) as a function of the damping rate , for damping behind the ray, unital noise off the ray, damping ahead of the ray, and the two twirl schedules. Only non-unital noise ahead of the ray is visible, and the late twirl removes it exactly.
Part (iii) removes this vulnerability. Averaging over the four Pauli operators on a site maps that site, together with every correlation it has with the rest of the chain, to , which is precisely the hypothesis of Theorem 1. Applied just before the edge gate instead of at the start of the circuit, the twirl erases whatever the incoming site picked up while it waited. It costs no two-qubit gate and at most one single-qubit gate per rail site. It also makes the preparation of the rail sites irrelevant, so they may be left in , the most strongly polarized state there is.
Dual-unitary gates are nearly immune to the exposure even without the late twirl. Their stay channel is completely depolarizing, so a polarized incoming leg affects the record only through terms bilinear in its own polarization and that of the ray. In simulation, relaxation of strength on every site shifts the transition kernel of the dual-unitary gate by at most 0.011, and that of a Haar-random gate by up to 0.042. For this reason the hardware test of the late twirl in Sec. VI C uses a Haar-random gate.
The co-moving frame and the memory of the record
Now move the readout sites into the cone, so that after half-step it measures site . The germ then reaches each measurement along many paths, and the argument behind Theorem 1 no longer applies as it stands. One can still follow the reduced state of the sites between the edge and the readout, and this gives a correct law with hidden states (Appendix F). That description, however, hides the structure of the problem. The natural coordinates are the diagonals of the circuit.
Every wire segment of a brickwork circuit runs either from the right input of a gate to its left output or from its left input to its right output, and we call the first kind left moving. The left-moving segments join into diagonal lines , where is the edge that carries the germ (Figure 4). In layer the gate of acts on the bond . Its left output continues along , sites inside the edge. Its right output , sites inside, enters the gate of in the next layer as the left input, and we call it a connector. The left input of comes from outside the cone. Information thus flows along each line and from each line to the one behind it, but never forward. Seen along its diagonals, the brickwork is a cascade.

Figure 4. The brickwork in light-cone coordinates. The left-moving lines (red), the edge of the germ, and (blue) are joined by right-moving connectors (orange). Information passes from to and never back. A readout on the edge (, circles) projects the co-moving qubit of itself. A readout one site inside (, squares) projects the connector that has just collided with that qubit, and a readout two sites inside (, triangles) projects the qubit of . A right-moving cut (dashed) separates the past of a record from its future and crosses one wire per line, which bounds the memory of the record (Corollary 2).
Theorem 3 (Co-moving frame). Let the readout sit sites inside the edge, at , and let . Under the hypotheses of Theorem 1, the record is generated by the co-moving qubits alone. In every layer the gate of acts on , with a fresh site. The readout of Eq. (3) acts on if is even and on the outgoing connector if is odd, after which is discarded. Writing for the resulting linear map on the strip of co-moving qubits,
where places the germ on and on every other qubit. A likelihood costs operations, and the length of the chain appears nowhere. For the strip is the single qubit , and Eq. (6) reduces to Eq. (2).
The proof (Appendix § 1.5.5) has two halves. Lines ahead of lie outside the forward cone of the germ and of every readout, so they feed with maximally mixed qubits, just as in Theorem 1. At the other end, the connector leaves at maximal speed, while every later readout sits on or on a connector leaving it. The connector thus never enters the backward cone of a later readout, and trace preservation removes it together with everything it touches further on. In site coordinates the strip is simply the window of sites between the edge and the readout. For the propagation of the reduced state alone, the evolution of such a window is the channel of [16], and the strip is the thin limit of the light-cone contraction schemes of Refs. [27, 28]. What these constructions do not contain is the record, with the back-action of every measurement, and what the co-moving picture implies for its memory.
How much does the record remember? Split it at some time into a past and a future, and arrange the joint probabilities into a matrix whose rows are labelled by past words and whose columns are labelled by future words. Its rank is the dimension of the smallest linear model that generates the record [29]. We define the memory of the record as the base-two logarithm of the largest rank over all cuts. A sequence of independent bits has no memory, and a two-state Markov chain has one bit.
Corollary 2 (Memory ladder). For a readout sites inside the edge, for every cut. The memory of the record is therefore at most bits, whatever its length.
To see this, cut the network along a right-moving diagonal that passes between the last readout of the past and the first readout of the future (Figure 4). The cut crosses each of once. Every other wire it crosses carries only the identity, ahead of because those qubits are maximally mixed and behind because nothing there is ever read. Each crossed line can carry an operator from a four-dimensional space, which bounds the rank by for odd . For even the crossed wire of has just been projected by onto the span of and , and the bound becomes . The window description, with its states, is thus the square of what the record needs.
The bound is not a loose estimate. Almost every gate attains it.
Proposition 1 (Generic saturation). Fix . If one two-qubit gate attains for some cut of some record length, then so does every gate outside a set of Haar measure zero. Such a gate exists for every .
Every entry of is a polynomial in the entries of and their complex conjugates, so a minor of order is a real-analytic function on the connected group . Such a function either vanishes identically or vanishes only on a set of measure zero, and a single gate with a nonzero minor decides which. We find such gates for every by constructing the reachable states of the strip directly, without enumerating records, and the smallest retained singular value stays at least six orders of magnitude above rounding error (Appendix E). For these displacements the memory of the record is exactly bits for almost every gate, and the states the strip can reach span exactly dimensions (Table I). We expect the same for every , but we have no witness construction that is uniform in .
The ladder has a simple physical reading. On the edge the measurement acts on the co-moving qubit itself. The projection wipes out everything the qubit knew, the hidden state is the last outcome, and the record is the Markov chain of Corollary 1. One site inside, the qubit of is never measured. In every layer it collides with a fresh qubit arriving from outside the cone, and that qubit is measured as it leaves. This is a collision model with indirect measurement [19], realized here exactly inside a scrambling chain of arbitrary length. The unmeasured qubit keeps a coherent memory of its entire history, and the record is not a Markov chain of any finite order. For the Haar-random gate of our demonstrations, outcomes older than the last four still change the probability of the next outcome by up to 0.054, although two bits of memory generate the record exactly. We stress that is a property of the readout, for a fixed and arbitrary gate. It has nothing to do with the hierarchies of gate conditions that extend solvability into the cone for special gates [11, 12]. At every displacement we tested, the law agrees with full density-matrix simulation to machine precision (Table I and Appendix F).
Slower readouts. A readout that moves at velocity , at site , has a displacement that grows with time. The same strip with gives its exact law, and the cut argument shows that its memory can grow by at most bits per half-step. For Haar-random gates it does grow, at every velocity we computed, whereas on every ray of maximal speed it saturates at bits (Fig. 5). The edge, in other words, is not the boundary of solvability but the first rung of a ladder, and a record can become hard to compute only if its readout falls behind the edge without bound.
record
Record probabilities are linear in , so is the response of the record to the germ alone, and every contribution that does not depend on the germ cancels exactly. Local unitaries preserve dual-unitarity, so coherent single-qubit errors cannot produce a signal either. We have checked that vanishes to machine precision for a dual-unitary gate under depolarizing noise and amplitude damping on the rail, relaxation of every qubit in every time slice, and arbitrary readout assignment errors (Table I). A nonzero at therefore certifies that the executed two-qubit dynamics has left the dual-unitary manifold, and Theorem 3 predicts its value for any given gate.
A witness of dual-unitarity
At the germ and every readout lie strictly inside the light cone. Dual-unitary circuits carry correlations only along the light-cone edges [7], so for a dual-unitary gate the record is exactly uniform. In the co-moving picture, the qubit that leaves a dual-unitary collision carries no information about the qubit it met, and the record has no memory at all. On a device this null is spoiled by effects that have nothing to do with the gate, such as relaxation toward and asymmetric readout. We therefore work with the germ-antisymmetrized
Checks against exact simulation
Every statement above was checked against full density-matrix simulation of the system together with the readout, and each check was paired with a control designed to fail if the claim were wrong (Table I). The controls fail at the percept level, by amounts that depend on the gate, so we quote the values for the ensembles we used and not as universal constants. Two entries deserve a comment. The late-twirl checks simulate the executed schedule, with damping in every time slice including the readout windows, so the object that is verified is the object that runs. And one control fails silently. The untransposed law passes a test with a -polarized germ to machine precision and is exposed only by transverse germs.
Demonstrations on a cloud quantum processor
Unless stated otherwise, the demonstrations ran on ibm_fez, a 156-qubit superconducting processor of IBM Quantum accessed through the cloud, on linear chains of physical qubits with the germ on the sublattice required by Theorem 1. We call these runs demonstrations, since the device was not under our control. No error mitigation was used in any run, and the interleaved jobs used no dynamical decoupling and no twirling by the runtime. The maximally mixed background was realized by randomizing the background sites from shot to shot. In the recorded rails the dual-unitary gate was compiled from an exact three-CNOT template [30], and in the interleaved demonstrations every gate was compiled from its matrix by the transpiler. The processors, the dates, the job identifiers, the calibration data at the time of each run and the software settings are documented in Appendix § 1.
Recorded rails obey the Markov structure
Our first test uses twelve rails recorded before the present analysis, at chain lengths to and depths , and , with shots each. Eleven are driven by the dual-unitary gate and one by a Haar-random gate. For each rail we fit Eq. (4), extended by asymmetric readout assignment errors, to the full distribution of the outcomes, once with a kernel common to all steps and once with a separate kernel for each step (Table II). The common-kernel chain fails, with a pooled . The step-dependent chain fits, with a pooled . It is also the correct null hypothesis, since each step of a rail runs on a different pair of physical qubits with its own gate and readout errors. The deepest rail is the most demanding test, and it is described at on 206 degrees of freedom. Two of the twelve rails stay above 3.
| Rail | gate | common | per step | ||
| R1 | 12 | 4 | DU | 23.44 (10) | 0.75 (6) |
| R2 | 12 | 4 | DU | 4.60 (10) | 1.65 (6) |
| R3 | 12 | 4 | DU | 3.09 (10) | 1.52 (6) |
| R4 | 20 | 4 | DU | 7.56 (10) | 1.27 (6) |
| R5 | 20 | 4 | DU | 3.41 (10) | 1.38 (6) |
| R6 | 20 | 4 | DU | 7.89 (10) | 3.08 (6) |
| R7 | 20 | 4 | Haar | 8.25 (10) | 1.16 (6) |
| R8 | 20 | 6 | DU | 1.52 (57) | 0.61 (49) |
| R9 | 50 | 4 | DU | 3.09 (10) | 4.89 (6) |
| R10 | 50 | 8 | DU | 12.49 (213) | 1.44 (206) |
| R11 | 54 | 4 | DU | 7.67 (10) | 1.75 (6) |
| R12 | 54 | 4 | DU | 8.46 (10) | 0.27 (6) |
| pooled | 9.52 | 1.37 |
Table II. Markov-structure test of Corollary 1 on twelve rails recorded on ibm_fez with shots each. Reduced of Eq. (4), extended by asymmetric readout assignment errors, against the full distribution of outcomes, with one kernel common to all steps and with one kernel per step. DU denotes the dual-unitary gate. Degrees of freedom are given in brackets.
The test also tells noise on the rail apart from defective hardware. A single-qubit rotation planted deliberately on the rail changes the kernel but not the Markov structure (Sec. II), and indeed the planted records fit as well as the controls measured on the same chains, with between 0.6 and 3.9 against 1.6 to 3.8. Chains that contain a qubit with a stuck readout, recorded on ibm_marrakesh, fail even the step-dependent test, at between 110 and 201.
The rail means themselves are the same function of at , 20, 50 and 54. Random Pauli errors injected on every off-ray site at rates up to leave them unchanged within statistics, with weighted slopes of , , and per unit of at the four steps, as Theorem 2 predicts for unital noise. Pauli injection cannot probe the non-unital half of part (ii), and that is the purpose of the demonstration described in Sec. VI C.
| Configuration | common | per step |
| dual-unitary, twirl at | 1.89 | 1.29 |
| dual-unitary, late twirl | 1.63 | 1.05 |
| Haar, twirl at | 1.98 | 0.69 |
| Haar, late twirl | 2.55 | 1.28 |
| Haar, twirl at , idle injected | 22.76 | 0.73 |
| Haar, late twirl, idle injected | 1.60 | 1.21 |
Table 3. TABLE III. Markov-structure test of the interleaved demonstration at , with shots per configuration and both germ signs fitted jointly with a shared kernel. The common kernel has 24 degrees of freedom and the per-step kernel 20.
An interleaved demonstration of noise protection and of the witness
The remaining predictions were tested in a single job on one chain of twenty qubits, with the germ at chain position and the rail region on physical qubits 147 to 142, whose reported times lay between 102 and 155 s. Background sites were prepared by a Hadamard gate followed by a mid-circuit measurement, which produces exactly in each shot. We ran eight configurations, each with the germ at and at and 16 384 shots per germ sign, and interleaved the 128 resulting circuit blocks in time so that drift cannot distinguish one configuration from another. The late twirl was realized exactly, by enumerating all Pauli assignments of the incoming sites with equal numbers of shots. Two-qubit gates were compiled from their matrices by the transpiler, and we checked that no compiled circuit contained a SWAP and that every record bit was measured on the intended physical qubit.
The acceptance criteria were set before submission. With the late twirl, every component of the idle-induced kernel shift had to satisfy , and with the twirl at at least one component had to exceed . The dual-unitary witness had to satisfy at every step, and the Haar witness had to exceed at the first step. Every late-twirl configuration had to pass the step-dependent Markov test at . All criteria were met.
All six configurations at pass the step-dependent Markov test (Table III), and the agreement with the noiseless law is close. The first-step germ-odd mean of the dual-unitary gate is 0.6784, against a predicted 0.6803, and the measured kernel contractions , 0.570 and 0.547 lie within ten percent of the noiseless value 0.607. A fit of Eq. (5) to the dual-unitary data gives a rail depolarizing strength of 1.2%, a post-readout relaxation of 3.9% and readout assignment errors below 0.5%.
Injected idle relaxation
To test part (ii) of Theorem 2 and its remedy, we inserted an idle period of on every qubit of the chain after the background preparation. This relaxes each incoming site by before it reaches the ray. With the twirl at , this non-unital noise ahead of the ray shifts the kernel contractions by and , at and (Figure 6). With the late twirl, on the same chain and in the same job, no component of the kernel moves by more than , as part (iii) predicts. Without injection the two protocols differ little, by at most for the dual-unitary gate and by in one kernel component for the Haar gate, consistent with the near-immunity of dual-unitary gates discussed in Sec. III. The injected relaxation changes the kernel but not the Markov structure. The record with injected idle still fits a step-dependent chain at and fails a common-kernel chain at 22.8.
Figure 6. Part (iii) of Theorem 2 on hardware. Shift of the kernel contraction of the monitored chain, Eq. (4), caused by an idle period of injected on every qubit, for the Haar-random gate. Filled circles show the twirl at and open squares the late twirl. Both configurations ran on the same chain in one interleaved job, with shots each. Error bars are bootstrap standard errors.
Measuring the executed gate
At the dual-unitary gate gives with a standard error of 0.0055, null at every step to within (Figure 7). The Haar-random gate on the same qubits gives , at , against a noiseless prediction of 0.065. At steps 1, 3 and 4 the Haar signal agrees with the noiseless law of Theorem 3 to within , and at step 2 it differs by 3.1. The dual-unitary witness at the same step is null, so the deviation is not a readout effect.
Figure 7. The germ-antisymmetrized record of Eq. (7), measured one site inside the edge () on ibm_fez. Open squares show the dual-unitary gate, for which the theory predicts exactly zero. Filled circles show a Haar-random gate on the same qubits, with the noiseless prediction of Theorem 3 as a dashed line. We used shots per gate. Error bars are bootstrap standard errors.
The null is in fact a measurement of the executed gate. We displace the nonlocal coefficients of the gate away from the dual-unitary point, with the local dressing held fixed, propagate each candidate through Eq. (6) with the noise calibrated on the data of the same chain, and compare with the measured (Appendix H). At 95% confidence, rad for a common displacement and rad for a differential one, . The corresponding dual-unitarity defect is at most 0.042, compared with 0.797 for the Haar gate. The bound needs neither tomography nor a fitted baseline, and single-qubit noise cannot fake a violation.
Memory of the displaced record
Theorem 3 predicts more than the means of the record. It predicts that the record of a Haar-random gate remembers, and that the edge record and the record of the dual-unitary gate do not. The simplest observable that sees the difference is the order-two memory
which vanishes identically for a Markov chain and for a sequence of independent bits. As for the witness, we use its germ-antisymmetrized part . The reason is that a readout assignment error turns the chain of Corollary 1 into a hidden chain, which has a memory of its own, but that memory does not depend on the germ. For the assignment errors of this chain, up to 1.4%, exact simulation gives a germ-even memory of 0.019 and a germ-odd part below .
Discovery. We first looked for memory in the records of the interleaved job after it had run, outside its acceptance criteria. For the Haar gate at these records give , a 6.5 effect, against 0.052 from Theorem 3 with no free parameter, and against zero on four components. The dual-unitary gate at and both gates at give between 2.0 and 7.7 (Appendix K).
Registered test. An effect found after the fact needs independent confirmation. Before any further data existed, we specified a second demonstration, its observable, its error estimate, its noise model and five acceptance criteria, and recorded them together with a checksum of the analysis code (Appendix L). Nine configurations were interleaved in one job (daop430pqrnc739aancg), namely the dual-unitary gate, the Haar gate of the first job (gate A) and a second Haar-random gate (gate B), each at , 1 and 2, all with the late twirl and both germ signs, with shots per configuration and germ sign, 589 824 in all. Gate B was chosen in advance for its large predicted memory. The chain consists of the twenty qubits of the first job traversed in the opposite direction, which puts the germ on qubit 110 and the edge rail on qubits 109, 118, 129 and 130, none of which played these roles before. The prediction is the law of Theorem 3 with a single noise model for the chain, fitted jointly to the three records of the same job. No parameter is fitted to a displaced record.
With , the 99.9% point of the distribution with four degrees of freedom, the criteria read as follows. (K1) The dual-unitary records at and 2 give against zero, and the three edge records give against the law. (K2) Gate A at reproduces above 5, with against the law. (K3, K4) Gate B differs from zero at at both and , exceeds 5 in the component that the noiseless law predicts to be largest, and agrees with the law at , and gate A at agrees with the law as well. (K5) A step-dependent two-state Markov chain is rejected at for both Haar gates at , each with a more than 2.5 times the largest value among the edge records and the dual-unitary record at at 11.6 and 30.3 on 12 degrees of freedom, which corresponds to .
All five criteria are met (Figure 8 and Table IV). Gate A at gives , at 7.3, against a predicted 0.037. Gate B gives and at , against and , and 0.098 0.006 at , at 17.6, against 0.089. The five records predicted to carry no germ-odd memory give between 1.0 and 12.4. The Markov chain describes the edge records and the dual-unitary record at at between 0.93 and 1.39, and it is rejected for the two Haar records at at 11.6 and 30.3 on 12 degrees of freedom, which corresponds to . and .
| record | largest | Markov | ||
| DU, | 0.9 | 1.0 | 0.97 | |
| DU, | 3.5 | 3.5 | 0.93 | |
| DU, | 5.4 | 5.4 | ||
| Haar A, | 12.3 | 12.4 | 1.39 | |
| Haar A, | 71.6 | 7.3 | 11.63 | |
| Haar A, | 34.0 | 9.5 | ||
| Haar B, | 1.2 | 1.2 | 0.99 | |
| Haar B, | 251.8 | 5.9 | 30.30 | |
| Haar B, | 356.5 | 14.3 |
Table IV. The registered memory demonstration on ibm_fez. For each record, of the germ-antisymmetrized order-two memory on its four components against zero () and against the law of Theorem 3 with the chain noise (), the significance of the component that the noiseless law predicts to be largest, and /dof of a step-dependent two-state Markov chain fitted to each germ sign (12 degrees of freedom). The registered threshold is . DU denotes the dual-unitary gate.
Figure 8. Memory of the record in the registered demonstration on ibm_fez. The measured germ-antisymmetrized order-two memory of Eq. (8) is plotted against the law of Theorem 3, with the chain noise fitted to the records of the same job, for all 36 components of the nine configurations. Open triangles mark the five records predicted to carry no germ-odd memory, the dual-unitary gate at , 1 and 2 and both Haar gates at . We used shots per configuration and germ sign. Error bars are binomial standard errors.
The chain was noisier than in the first job, with a fitted depolarizing strength of 5.4% per layer, and the noise matters. Against the noiseless law, gate B gives at and 154 at , against 5.9 and 14.3 for the law with the chain noise. Over all 36 components the law with the chain noise gives , somewhat more than the number of components, which reflects the limits of a five-parameter noise model (Appendix L). One and two sites inside the edge, then, the record of a scrambling gate is no longer a Markov chain, and the co-moving qubits account for the size of the departure.
Repetitions. The registration provided for repetitions on other processors, each analysed on its own against the same criteria. The same nine configurations were submitted to ibm_marrakesh and ibm_kingston within fifteen minutes of the job on ibm_fez (Appendix L and Table XV). Every prediction that does not involve the noise model holds on all three processors. The fifteen records predicted to carry no germ-odd memory give , the twelve displaced Haar records differ from zero at between 30 and 623, gate A at exceeds in each time, and the step-dependent Markov chain is rejected for both Haar gates at at on every processor. What does not carry over unchanged is the agreement in size under one five-parameter noise model per chain. On ibm_kingston gate B at gives , just above , so that K3 fails while the other four criteria are met. On ibm_marrakesh the noise model fails already on the edge records to which it is fitted, at , although the step-dependent Markov chain describes the same records at 0.76 to 0.84. There K3 and K4 fail, and gate B at agrees instead with the noiseless law, at . Against the law with the chain noise, eight of the twelve displaced Haar records lie below . The memory is therefore reproduced on three
| ibm_fez | ibm_marrakesh | ibm_kingston | |
| germ qubit | 110 | 2 | 36 |
| noise model | 2.72 | 10.49 | 1.94 |
| nulls, largest | 12.4 | 3.7 | 6.6 |
| Haar A, | 71.6 / 7.3 | 94.7 / 11.0 | 46.9 / 12.1 |
| Haar A, | 34.0 / 9.5 | 63.2 / 20.9 | 30.2 / 0.6 |
| Haar B, | 251.8 / 5.9 | 623.1 / 115.8 | 231.3 / 21.1 |
| Haar B, | 356.5 / 14.3 | 273.8 / 48.5 | 606.8 / 8.0 |
| gate A, | |||
| gate B, | |||
| gate B, | |||
| Markov, nulls | 1.39 | 1.00 | 1.86 |
| Markov, Haar A, B | 11.6, 30.3 | 17.3, 53.2 | 10.0, 24.9 |
| criteria met | all | K1, K2, K5 | all but K3 |
Table XV. The registered memory demonstration on ibm_fez and its repetitions on ibm_marrakesh and ibm_kingston, each analysed on its own by the registered script against the registered criteria. Displaced Haar records give as in Table XIV. The nulls are the five records predicted to carry no germ-odd memory. Significances refer to for gate A and to the component largest in the noiseless law for gate B. The noise model row gives of its fit to the edge records. The Markov rows give of the step-dependent two-state chain, the largest value among the three edge records and the dual-unitary record at , and the values for both Haar gates at .
processors, while a quantitative prediction of its size is only as good as the model of the noise of the chain.
Locating a faulty gate
Causality also localizes faults. The correlator involves only the gates of steps , so the first lag at which it deviates points to a single gate. On ibm_marrakesh one correlator exceeded its noiseless value, the opposite of what decoherence normally does, and the analysis pointed to the gate of step 2. The corresponding qubit returned outcome 1 in 6.5% of shots, and , measured as 0.4505 against 0.4332, confirmed a nearly constant bit. The calibration record of the provider at the time listed this qubit with a readout assignment error of 2.6% and , inside the cuts used to select the chain, so the calibration did not reveal the fault. The two other runs on the same chain, submitted eight minutes earlier and 73 minutes later, single out the same qubit, and it later exceeded the device’s own readout-error threshold. Faults comparable to the baseline error of a chain require a paired reference measured in the same job. With planted rotations of 0.8 rad, the paired comparison identifies the planted step on both processors. The plant at step 2 moves by and , at about , while , which causality protects, stays within , and a control compared with itself flags nothing.
Discussion
The record law is a likelihood, and therefore also a sampler, so records on the edge carry no sampling advantage. Exactness and classicality have the same origin. Moving the readout into the cone does not destroy exactness. It raises the memory by one bit per site and the cost by a factor of sixteen for every two sites, in the displacement and not in . For generic gates a monitored record can be hard to compute only if the readout moves strictly slower than the edge, so that it falls behind through a number of lines that grows with time, adding at most bits of memory per half-step (Figure 5). The co-moving frame thus turns the question of where hardness lives into a count of lines, and it provides exact references along the way. A record at and depth involves four co-moving lines, and its law is computed exactly on eight qubits, however long the chain.
Figure 5. Memory of the record, the base-two logarithm of the largest rank of over all cuts, as a function of the length of the record, for a Haar-random gate. On rays of maximal speed the memory saturates at bits and then stays there as the record grows (grey lines). Below the maximal speed it grows with . The dotted line marks the largest memory a record of length can have. All values are exact and follow from Theorem 3.
For special gates, solvability reaches further into the cone, through hierarchies of gate conditions [11, 12] and through constructed circuits whose solvable influence matrices are non-Markovian [31]. Space-time duality maps sub-luminal monitored records onto unitary correlation functions of rotated circuits [32, 33], which is the natural setting in which to ask when a record becomes hard. As an instrument, the record law complements the benchmarking of mid-circuit measurements [2, 3]. Those protocols characterize the measurement operation itself, whereas the record law provides an exact reference for a scrambling circuit as a whole.
The scope of the results should be stated plainly. All exact statements require the sites that join the edge to arrive maximally mixed and uncorrelated. The infinite-temperature background provides this, and the late twirl enforces it whatever the background, but the twirl is an average over shots. It does not describe a chain quenched from a pure state, in which the bath ahead of the edge has a memory of its own. After such quenches the temporal entanglement of influence matrices in chaotic circuits follows a volume law in time [26], and we see no reason to expect a ladder of fixed height there. On the hardware side, the Markov tests reach depth , while the memory demonstrations use , which is too short for a direct test of the rank and leaves the order-two memory as the natural probe. The noise model behind the displaced predictions has five parameters and treats the chain as homogeneous. Fitted to the edge records it gives on ibm-fez, 1.94 on ibm-kingston and 10.5 on ibm-marrakesh, and the predicted size of the memory meets the registered threshold in all four displaced Haar records on the first processor, in three on the second and in one on the third. The memory itself appears on all three, but a prediction of its size requires a model of the noise of the chain, which the present one only approximates. Finally, Proposition 1 is established for only.
Two extensions follow directly from the construction. The cut argument does not use the identity of the gates, and with the late twirl in place local noise on the strip can at most double the bound at even . The rank of , or the decay of its singular values, would therefore test the memory ladder on hardware without any noise model, given records long enough to estimate it. The ladder also indicates where to look for records that are hard to predict, namely behind the edge at sub-maximal speed, where the number of co-moving lines grows with time.
Conclusion
The output of a quantum processor that measures qubits mid-circuit is a classical record, and for a scrambling circuit that record is in general as hard to predict as the circuit itself. We have shown that near the light cone this is not so. Causality alone makes the record on the edge exactly Markovian, for arbitrary gates and at any system size, and seen from a frame that moves with the edge, a scrambling chain is a memoryless bath for the few qubits the observer carries. The memory of the record is quantized, one bit on the edge and at most one more for every site of displacement into the cone, and generic gates saturate the bound.
For the validation of quantum hardware this provides what was missing in the regime of dynamic circuits, an exact reference for a scrambling circuit with mid-circuit measurements whose cost does not grow with the device, together with a precise account of the noise it is blind to, a remedy at no gate cost for the one channel it is not blind to, and a witness of the dual-unitarity of the executed gate that single-qubit noise cannot fake. For the theory of monitored many-body dynamics it locates where the output can become hard to predict, namely behind the edge at sub-maximal speed, and it shows that the collision models of open-system theory arise exactly inside generic scrambling circuits.
On a 156-qubit superconducting processor accessed through the cloud, twelve independently recorded rails obey the predicted Markov structure, and non-unital noise injected ahead of the ray corrupts the record only when the late twirl is absent. The witness bounds the nonlocal coefficients of the compiled dual-unitary gate to 0.011 rad, and a registered demonstration finds the memory of the co-moving qubits one and two sites inside the edge, for two scrambling gates, at the predicted values, and none where none is predicted. Registered repetitions on two further processors find the memory and its absence in the same places, and they show that its predicted size is only as reliable as the model of the noise of the chain.
Acknowledgments
The author acknowledges support from the Helmholtz Association HGF (Germany), the Hamburgische Investitions- und Förderbank (IFB) (Germany), and the Ministry of Science, Research and Culture of the State of Brandenburg within the Center for Quantum Technologies and Applications (CQTA) (Germany). The hardware runs used the IBM Quantum processors, accessed through DESY.
Data and code availability. The hardware records of every run in Table IX, the calibration records and job identifiers of Appendix § 1, the circuits and results of every identified job, the protocol of the registered memory demonstration, the exact-simulation code and the analysis code that regenerate every number and figure in this work are available from the author and will be deposited in a public repository upon publication.
TABLE IX. The hardware runs analysed in this work, all in 2026, with submission times in UTC. The runbook of August did not store job identifiers, and those given for August were identified afterwards as described in the text.
| run | data file | processor | job | submitted | qubits, shots |
| Recorded rails (Sec. VI A and Table II) | |||||
| R1 | r12_clean | ibm_fez | d9ph073bvhrs73a2k7d0 | 5 Aug 10:27 | 12, |
| R2 | r12_v2 | ibm_fez | d9phbjd7skus73ckvhqg | 5 Aug 10:51 | 12, |
| R3 | results_n12 | ibm_fez | d9pentu28h6s739r7fvg | 5 Aug 07:53 | 12, |
| R4 | r20_clean | ibm_fez | d9ph0espcc1s73a7lhag | 5 Aug 10:28 | 20, |
| R5 | r20_generic | ibm_fez | d9pj5rm28h6s739rd330 | 5 Aug 12:56 | 20, |
| R6 | inj_0_0 | ibm_fez | d9phfmcpcc1s73a7m6r0 | 5 Aug 11:00 | 20, |
| R7 | r20_haar2 | ibm_fez | d9pj9jd7skus73cl21i0 | 5 Aug 13:04 | 20, |
| R8 | results_n20 | ibm_fez | d9peoh57skus73cks790 | 5 Aug 07:54 | 20, |
| R9 | r50_clean | ibm_fez | d9pgvbspcc1s73a7lfhg | 5 Aug 10:25 | 50, |
| R10 | results_n50 | ibm_fez | d9pesud7skus73cksd10 | 5 Aug 08:04 | 50, |
| R11 | r_big | ibm_fez | d9ph79bbvhrs73a2kkdg | 5 Aug 10:42 | 54, |
| R12 | r_100 | ibm_fez | d9pk84bbvhrs73a2off0 | 5 Aug 14:09 | 54, |
| inj_0_1 | ibm_fez | d9phhpm28h6s739rb0h0 | 5 Aug 11:05 | 20, | |
| inj_0_2 | ibm_fez | d9phi4628h6s739rb100 | 5 Aug 11:05 | 20, | |
| inj_0_3 | ibm_fez | d9phicspcc1s73a7m9tg | 5 Aug 11:06 | 20, | |
| inj_0_5 | ibm_fez | d9phikd7skus73ckvpl0 | 5 Aug 11:06 | 20, | |
| mirrored Haar gate | r20_haar | ibm_fez | d9pj7dspcc1s73a7odhg | 5 Aug 12:59 | 20, |
| Planted faults (Secs. VI A and VI F, Appendix M) | |||||
| step 3, no control | r20_plant3_marrakesh | ibm_marrakesh | d9pl28l7skus73cl4r9g | 5 Aug 15:05 | 20, |
| control | plant_control | ibm_marrakesh | d9pl5ku28h6s739rg6p0 | 5 Aug 15:12 | 20, |
| step 3 | plant_step3_v3 | ibm_marrakesh | d9plbgjbvhrs73a2qae0 | 5 Aug 15:24 | 20, |
| step 3, repeated | plant_step3_v4 | ibm_marrakesh | d9plcnl7skus73cl5a8g | 5 Aug 15:27 | 20, |
| step 2 | plant_step2 | ibm_marrakesh | d9plkpe28h6s739rgp9g | 5 Aug 15:44 | 20, |
| step 2, paired | paired_step2 | ibm_marrakesh | d9plmukpcc1s73a7s3pg | 5 Aug 15:49 | 20, |
| step 2, paired | paired_step2_fez | ibm_fez | d9plo757skus73cl5of0 | 5 Aug 15:51 | 20, |
| Stuck readout, one chain (Secs. VI A and VI F) | |||||
| stuck readout | r50_S4 | ibm_marrakesh | d9pfd257skus73cksv70 | 5 Aug 08:38 | 50, |
| stuck readout | r50_nodd | ibm_marrakesh | d9pfgv628h6s739r8c90 | 5 Aug 08:46 | 50, |
| stuck readout | r50_b | ibm_marrakesh | d9pgj8rbvhrs73a2jh90 | 5 Aug 10:00 | 50, |
| Interleaved and registered demonstrations (Secs. VI B to VI E, Appendices § 2.3 and L) | |||||
| interleaved | fez_lightcone | ibm_fez | dao fjddr85ps73ff7970 | 21 Sep 09:32 | 20, 262 144 |
| registered | memory_fez | ibm_fez | daop43opqnr c739aacng | 21 Sep 20:22 | 20, 589 824 |
| repetition | memory_marrakesh | ibm_marrakesh | daop93opqrnc739ab0og | 21 Sep 20:33 | 20, 589 824 |
| repetition | memory_kingston | ibm_kingston | daopb1n8gn2s739o77g0 | 21 Sep 20:37 | 20, 589 824 |
Table IX.
The two channels and why the transpose matters
Write the single-gate Pauli channel of as , a real matrix in the basis . For any single-qubit operator and any state ,
Hence is the Heisenberg adjoint of the Schrödinger map , which carries the state of the right site to the left site with a maximally mixed partner, and in an orthonormal Pauli basis . The record propagates states, not observables, so and not is the transfer matrix. The identity is elementary, but confusing the two matrices is an easy mistake to make and a hard one to detect. The block probed by a -polarized germ is symmetric, so a test with such a germ cannot tell the two apart. Any reimplementation should therefore be validated with transverse germs and over an ensemble of gates. Over 12 Haar-random gates the untransposed law fails by a median of , but by as little as for individual gates, so a single test with a transverse germ is not a reliable discriminator either. The same caution applies to every control in Table VII. Their failure magnitudes span more than two decades across Haar gates and should not be quoted as constants.
| Ensemble | error |
| 5 Haar gates 4 germ vectors, | |
| inhomogeneous, different at every step | |
| -independence, at fixed geometry | |
| unital depolarizing off ray | |
| arbitrary , , behind the ray | |
| Thm. 3, (16, 64, 256 states) | |
| executed schedule, late twirl, no noise | |
| executed schedule, late twirl, , | |
| late twirl, / 0.2 everywhere | |
| ahead of the ray | |
| ahead of the ray | |
| twirl schedule, every slice | |
| twirl, best model, | |
| untransposed law, transverse germ | |
| readout displaced , (2) | |
| readout displaced , (2) |
Table VII. Verification ensembles for the edge law and the late twirl, regenerated by verify_all.py. Executed-schedule checks include damping in every time slice. The controls in the lower block must fail.
Proof of Theorem 1
The theorem follows from a statement about the state just before each edge gate. The edge gate of half-step acts on the pair , where is the rail site read at half-step and is the incoming site. We claim that the reduced state of this pair, conditioned on the outcomes , is exactly , with the incoming site as the first factor and the conditional state of the rail site.
Proof. Write the global state after the readout of half-step , conditioned on the outcomes so far and not normalized, as , with the probability of these outcomes and a traceless operator. Initially and is supported on the germ. A brickwork layer conjugates by a unitary and enlarges its support by at most one site in each direction. A readout of site with outcome replaces by , with , and adds support only on . Every readout so far took place on a site with , inside the forward cone of the germ, so after half-step the support of is contained in . Now take the partial trace onto . Every Pauli string in with a nontrivial factor outside traces to zero, and no string has a factor on , which lies outside the cone. What survives is plus strings supported on alone, which is . The edge gate is the only gate of layer that touches , so the conditional state of the next rail site is , whose Bloch vector is obtained by applying to that of (Appendix A). The right output of the gate falls behind the ray and never influences a later readout. The readout then applies Eq. (3), and composing the steps gives Eq. (2). The parity hypothesis is what allows a single gate per half-step to carry the ray. If and do not share a gate at , the measured sites no longer trace the edge and the law fails (Table VIII).
| error | ||
| 9 | 7 (paired) | |
| 9 | 6 (unpaired) | [] |
| 10 | 7 (paired) | |
| 10 | 8 (unpaired) | [] |
Table VIII. Dependence on the sublattice of the germ. The law holds when the germ occupies the sublattice paired with the first brickwork layer and fails otherwise, independently of the length of the chain. The size of the failure depends on the gate, and we quote medians over 12 Haar-random gates with the full range in brackets.
Proof of Corollary 1
Write the vector of the ray after the gate of half-step as . The readout applies , and for
The prefactor is the probability of the outcome and is the Bloch vector of , so after the first readout the hidden state is the last outcome itself. The next gate maps to a vector whose component is , with and , so that occurs with probability . Only the first readout sees the full germ vector, through , and multiplying the factors gives Eq. (4) (Figure 9). For a unitary gate with a maximally mixed incoming site the edge propagator is unital and . A single-qubit channel acting on the rail between readouts replaces by its composition with that channel, which changes but not the form of the law. A readout assignment error acts on the recorded bit after the state update and turns the chain into a hidden two-state chain. The identity holds numerically to (Table I).

Figure 9. The path of Figure 2 as a walk on the four Pauli components of the ray. Each edge gate applies (grey), which fixes and mixes , and , and each readout applies (red), which keeps only and . After the first readout every walk stays on , which is the content of Corollary 1. Non-unital noise on the rail adds an edge from to , the bias . The probability of a record is the total weight of the walk, a product of matrices however many qubits are present.
Proof of Theorem 2
Let a channel act on site at half-step $h.
Suppose . Information propagates by at most one site per half-step, while the readouts move left by exactly one site per half-step. A perturbation at can reach site at half-step , and for all . It never meets the ray, and the record is invariant for arbitrary , unital or not.
Suppose . This site joins the rail at half-step as the incoming site of an edge gate, and the proof of Appendix § 1.5.2 requires it to arrive as . The region ahead of the ray is in the maximally mixed state, which unitaries and unital channels leave unchanged, so unital channels are invisible. For necessity, note that a non-unital produces an incoming Bloch vector , on which the edge propagator depends linearly, , with for generic .
For any state of the chain, . A uniformly random Pauli operator applied to site therefore erases its polarization and all of its correlations, exactly, on average over shots. Applied to each incoming site just before its edge gate, the twirl restores the hypothesis of Appendix § 1.5.2 regardless of what happened to the site before. Single-qubit channels on the rail compose with as in Appendix § 1.5.3, and an assignment error replaces by , which gives Eq. (5). A further channel between a readout and the next edge gate is absorbed into the next propagator, so the law also covers relaxation during the mid-circuit measurement itself, which lasts about . □
Consequences for hardware. Amplitude damping is non-unital, so part (ii) applies precisely to the qubits of the rail that the ray has not yet reached. These qubits idle from through several mid-circuit measurements and arrive polarized. Full density-matrix simulation gives a record deviation of at and at , first order in (Table VII). The naive repair, feeding a damped state into , is itself wrong at order , because the incoming sites are also correlated with one another through the gates they passed through while idle. Its residual is at . The late twirl removes the problem at the cost of moving single-qubit gates later in the schedule. Simulating the executed schedule with damping on every qubit in every time slice, layers and readout windows alike, gives at and at , whereas the same schedule with the twirl left at deviates from any such law by and . The verification runs on an intermediate representation of the schedule that is also what is emitted to Qiskit, so the object that is checked is the object that runs (runbook_patch.py).
Proofs of Theorem 3, Corollary 2 and Proposition 1
Light-cone coordinates. Label the gate on the bond in layer as . Under the parity hypothesis these are all the gates of layer , with running over the integers allowed by the chain. The left output of is site , which in layer is the right member of the bond and hence the right input of . The right output of is site , the left member of the bond of . The gates thus form a left-moving line threaded by one qubit , and the only coupling between lines is the connector from to . An operator inserted on can spread only to lines of index and higher, and one inserted on can only to lines of index and higher. The germ is the right input of and lies on .
Theorem 3. Let the readouts sit at displacement , with . For even they act on a after each gate of , and for odd on . (a) Lines of negative index receive only maximally mixed inputs, because neither the germ nor any readout can spread to them, and a unitary maps to itself. The left input of every gate of is therefore and uncorrelated with the strip. (b) Lines of index larger than are never read and never couple back, so by trace preservation the trace over them can be taken as soon as leaves . What remains is the evolution of and of the connectors in flight between them, which is Eq. (6). The argument holds gate by gate for inhomogeneous circuits and for arbitrary single-qubit channels on the strip. With the late twirl it holds for arbitrary noise ahead of , and noise behind is invisible by (b). For a readout trajectory whose displacement never exceeds , the argument is unchanged with , the readout acting on whichever wire of the strip sits at displacement .
Corollary 2. The right outputs of the gates lie on one right-moving diagonal
. Consider the cut that runs between and . It crosses each line once, on the segment between two consecutive gates, and it crosses no connector. For even the readout sits on the crossed segment of , and the cut is placed after it. The readouts then lie on one side and on the other. By (a) the wires of lines of negative index carry from the past side, and by (b) the wires of lines of index above are contracted with the identity on the future side, so both contribute rank one. The record distribution is a contraction of a past tensor and a future tensor over the operator spaces of the crossed wires, which bounds the rank of by . For even the crossed wire of has just been acted on by , whose range is spanned by and , and the bound improves to . In both cases it equals . For a trajectory with the same cut gives . Table V lists the ranks found numerically. □
| ray | bound | Haar | Haar | dual-unitary |
| , | 1 | 1 | 1 | 1 |
| , | 2 | 2 | 2 | 0 |
| , | 3 | 3 | 3 | 1 |
| , | 4 | 4 | 4 | 0 |
| , | 5 | 5 | 5 | 1 |
| 4 | 4 | 4 | 1 | |
| 6 | 5 | 5 | 1 | |
| 8 | 6 | 6 | 1 |
Table V. Memory of the record in bits, the base-two logarithm of the largest rank of over all cuts of a record of length , which can show at most 6 bits. Two Haar-random gates and a dual-unitary gate, computed from Theorem 3 by lightcone_comoving.py.
Proposition 1. Each record probability is a contraction of copies of and of with fixed tensors, and hence a polynomial in the real and imaginary parts of the entries of . A minor of order of is a real-analytic function on , which is a connected real-analytic manifold. If does not vanish identically, its zero set has Haar measure zero. The hypothesis provides a gate with for a suitable minor, and Corollary 2 bounds the rank from above. □
The witnesses are computed by ladder_saturation.py, which needs no enumeration of records. It builds the states that the strip can reach after a past of length , discarding at each step every state that is a linear combination of states already kept. Nothing is lost by this pruning, because the maps are linear. For the Haar gates tested the reachable states span exactly dimensions. The script then evaluates all future words of length on each kept state. Every kept state is the state after one definite past word, so the resulting matrix is, up to a rescaling of its rows, a submatrix of for records of length , and its rank bounds that of the full matrix from below. Table VI gives the smallest of its first singular values relative to the largest. It stays far above the rounding error of , while a singular value number does not exist because the matrix has only rows. For the memory of the record is therefore exactly bits for almost every gate.
| reachable | rank | (seed 4, 11) | ||
| 0 | 2 | 2 | 2 | , |
| 1 | 4 | 4 | 4 | , |
| 2 | 8 | 8 | 8 | , |
| 3 | 16 | 16 | 16 | , |
| 4 | 32 | 32 | 32 | , |
| 5 | 64 | 64 | 64 | , |
| 6 | 128 | 128 | 128 | (seed 4 only) |
Table VI. Witnesses for Proposition 1. Number of linearly independent reachable states and rank of the matrix of future-word probabilities evaluated on them, with its smallest singular value relative to the largest. Haar-random gates with seeds 4 and 11 of lightcone.haar_u.
Markov order. A finite rank does not make the record a Markov chain of finite order. At the hidden state is the unmeasured qubit of , which no finite string of outcomes determines. For the Haar gate of the interleaved demonstration, the largest change in the probability of the next outcome caused by the history before the last outcomes is 0.081, 0.081, 0.056, 0.054 and 0.039 for , at in records of length , and zero to machine precision at .
The window form of the displaced-readout law
The same record can be described in site coordinates. This is how it was first derived, and it provides an independent check of Theorem 3. Let the readout sit at . A site at influences the readout at half-step only if . Since is constant along a ray of maximal speed, the backward cone of the whole record is bounded on the right by at half-step . Intersecting it with the forward cone of the germ leaves the window of sites. Everything to the left of the window is maximally mixed and uncorrelated with it, and everything to the right is causally disconnected from the record. For the propagation of the reduced state alone, the evolution of such a window is the channel of Ref. [16], independent of . The new ingredient is its interleaving with the non-unitary readout update.
The transfer from to takes three steps. Adjoin a maximally mixed site on the left, apply the gates of layer whose support lies inside , and trace out the rightmost site. Under the parity hypothesis and for homogeneous gates the set of window gates is the same at every half-step, so a single matrix suffices. There is one window gate for , there are two for , and in general there are . The readout acts on the rightmost window site, which is by construction. The initial window state is obtained by applying the layer-1 gates inside to the product initial state and tracing out the rightmost site. Against full density-matrix simulation this form agrees to , and at and 2, while applying the law to data, or the reverse, fails by . Figure 10 shows the agreement for each displacement and Figure 11 the means themselves. Table VII lists the verification ensembles for the edge law and the late twirl, and Table VIII the dependence on the sublattice of the germ.
Figure 10. Theorem 3 against full density-matrix simulation. Largest deviation between the displaced-readout law and exact simulation of system plus readout, for a generic gate at and 2, with hidden states in the window form.

Figure 11. Displaced rail means for a generic gate at , 1 and 2. Lines show the law of Theorem 3 and crosses show full density-matrix simulation. The dual-unitary null at is included for comparison. The law agrees with the simulation to machine precision (Table I and Figure 10).
The dual-unitary null at
The null follows in one line from the co-moving frame. At the measured qubit is the right output of a gate whose left input is a fresh . For a dual-unitary gate the map from the right input to the right output with a maximally mixed left input, , is completely depolarizing, so it sends any unnormalized state of the right-going qubit to . Each outcome then occurs with probability whatever the past, and the record is exactly a sequence of fair coins, . In the language of correlation functions, the readouts lie on neither edge of the cone of the germ and on neither edge of each other’s cones, where dual-unitary circuits have no correlations [7]. All means, all correlators and the total-variation distance to the uniform distribution on vanish identically, and the computed distance at is 0.0000. At the means still vanish, while reproduces the values 0.6066, 0.3679 and 0.2232 exactly, because consecutive readouts are again connected by a light-cone edge. For a Haar-random gate with dual-unitarity defect 1.91 nothing vanishes at any .
The practical consequence is that a rail measures the dual-unitarity defect of the executed gate against an exact zero, with no fitted baseline and no simulator, and Theorem 3 supplies the exact law for whatever nonzero signal appears. Coherent miscalibration of the compiled three-CNOT template shows up directly. Unital noise outside the strip does not (Appendix § 1.5.5).
The germ-antisymmetrized witness
Record probabilities are linear in the initial state. Since , every record probability, for any noisy circuit and any readout error, has the form . The antisymmetrized record removes exactly, and with it every effect that does not depend on the germ, including relaxation toward and asymmetric readout error. What remains is the linear response of the record to the germ. For a dual-unitary gate at this response vanishes identically in the noiseless circuit (Appendix G), and it remains zero when the gates are dressed by single-qubit unitaries, which preserve dual-unitarity. We verified to machine precision that it also remains zero under depolarizing noise and amplitude damping on the rail window after every gate, amplitude damping on the window after every readout, amplitude damping on every qubit in every time slice, and asymmetric readout assignment errors.
Bound on the executed gate. Candidate gates are parametrized as
with the dressing of the deployed gate. The rail noise is calibrated by fitting the noisy law, with a depolarizing strength and amplitude damping after each gate, amplitude damping after each readout and asymmetric assignment errors, to the dual-unitary data of the same job. The fit gives a depolarizing strength of 0.012, damping of 0.000 after the gate and 0.039 after the readout, and assignment errors of 0.0045 and 0.0035. For each candidate the record is computed from Theorem 3 with this noise, and is compared with its minimum. The 95% interval, on a grid of spacing rad, is rad for a common displacement and rad for a differential one, . The dual-unitarity defect at the edges of these intervals is 0.042 and 0.039. At rad the predicted witness is for the common direction and for the differential one, several times the statistical error of 0.0055.
Cloud platform and methods
All hardware runs used processors of IBM Quantum, accessed through the cloud with Qiskit Runtime. Since the devices were not under our control and their characteristics change over time, we document here the processors, the dates, the job identifiers, the calibration data reported by the provider at the time of each run and the software settings, following the guidelines of the Physical Review journals for cloud quantum computing demonstrations. The calibration records and job identifiers are stored with the data.
Processors. ibm_fez, ibm_marrakesh and ibm_kingston are 156-qubit superconducting processors of IBM’s Heron r2 generation, with transmon qubits on a heavy-hexagonal lattice, tunable couplers, the native gate set CZ, , and , and support for mid-circuit measurement and reset. The configuration records of the provider give the processor family Heron, revision 2, for all three, and its calibration records give the backend versions 1.3.37 for ibm_fez, 1.0.21 for ibm_marrakesh and 1.0.0 for ibm_kingston at the times of the runs. The device clock had the period on all three. On the chain of the two jobs of September on ibm_fez a CZ gate took 68 ns, except on the bond between qubits 146 and 147, where it took 88 ns, and a readout took 1.66 s. Every circuit of this work was compiled to the native gate set by the Qiskit preset pass manager at optimization level 1, with the chain of physical qubits pinned as the initial layout, and was executed with the Sampler V2 primitive of Qiskit Runtime in job mode. The job records of the provider name the client software, Qiskit 2.5.1 with qiskit-ibm-runtime 0.48.0 for the runs of August 2026 and Qiskit 2.4.2 with qiskit-ibm-runtime 0.47.0 for the jobs of September. Table IX lists the runs.
Job identifiers. The runbook used in August 2026 stored the records but not the job identifiers. All twenty-seven runs of August analysed here were identified afterwards in the records of the provider. For each of them the outcome histogram of every circuit of the job, of 1250 or 2500 shots, coincides with the corresponding block of the stored records. Whatever the outcome probabilities, the chance that a run reproduces a given histogram exactly is largest when they equal the observed frequencies, and even then the chance that another run reproduces all histograms of one of these jobs is below . The original times of the data files, kept with the data, agree, since every one of these jobs finished before the runbook wrote its file, less than 18 minutes earlier. The four jobs of September 2026 recorded their identifiers at submission, and for them too the histogram of every circuit returned by the provider coincides with the corresponding block of the stored records. Every job started to run within 90 s of submission, except R12, which waited 63 minutes in the queue, the paired run on ibm_fez, which waited 11 minutes, and R11, R6 and R2, which waited 9, 4 and 3 minutes.
Calibration data. For every job we retrieved from the provider the calibration record in force when the job started to run, for all qubits and bonds of its chain, and these records are stored with the data. Table XI summarizes those of the runs analysed here, and Table X lists in full the records for the two jobs of September on ibm_fez, last updated at 09:03 and 19:08 UTC on the day of the jobs. The records contain no qubit frequencies. In the provider’s history a record can take effect minutes after its last update, at least 13 minutes in one case here. The records retrieved show six runs, R1, R4, r50-S4 and three planted-fault runs, that started under a record although a newer one had been last updated 71 to 785 s before. For four of the six that newer record is stored, with later runs, for the whole chain, and it differs from the one in force at most in the single-qubit gate errors, which Table XI does not list. For the four jobs of September the runbook had also stored, at submission, the readout assignment error and the relaxation time of every rail qubit, and the records agree with these values exactly. The chain of the interleaved job was selected with cuts of 3% on the readout error and 60 on over the rail region, chain positions 13 to 18, and ranked by the summed readout and two-qubit gate errors there, and the registered job used the same qubits in reverse order. The chains of August were selected at submission, against the calibration then published, with cuts of 5% on the readout error and on over the rail sites, and the device excluded 16 qubits by those cuts on the day of the 50-qubit runs. In the records in force at execution every rail site of the runs of August passes these cuts, except qubit 65, the last rail site of R11 and R12 and the germ of R10, at readout errors of 5.4% to 5.7%, qubit 43, a rail site of R10, at 8.1%, and qubit 11, the germ of R3, at 7.8%.
| qubit | (µs) | (µs) | (%) | () | () | |
| 98 | 0, 19 | 136 / 124 | 24 / 22 | 1.25 | 3.4 / 2.4 | 3.0 / 2.7 |
| 111 | 1, 18 | 182 / 198 | 113 / 123 | 0.63 | 2.5 / 2.6 | 2.6 / 2.3 |
| 110 | 2, 17 | 121 / 144 | 116 / 116 | 0.29 | 2.9 / 2.0 | 2.3 / 2.2 |
| 109 | 3, 16 | 166 / 162 | 69 / 71 | 0.61 | 3.6 / 3.4 | 2.4 / 2.9 |
| 118 | 4, 15 | 141 / 82 | 58 / 48 | 0.54 | 2.1 / 4.0 | 6.3 / 2.6 |
| 129 | 5, 14 | 132 / 103 | 109 / 91 | 0.85 | 2.4 / 2.1 | 8.1 / 2.9 |
| 130 | 6, 13 | 80 / 152 | 110 / 162 | 0.51 | 4.2 / 4.3 | 2.2 / 2.1 |
| 131 | 7, 12 | 206 / 159 | 245 / 160 | 7.03 | 1.3 / 1.7 | 2.3 / 2.4 |
| 138 | 8, 11 | 82 / 81 | 100 / 86 | 0.61 | 4.0 / 3.1 | 2.4 / 4.6 |
| 151 | 9, 10 | 87 / 71 | 116 / 101 | 1.05 | 1.8 / 3.0 | 7.6 / 7.0 |
| 150 | 10, 9 | 83 / 141 | 5.4 / 5.3 | 0.49 | 7.2 / 7.9 | 27.2 / 45.3 |
| 149 | 11, 8 | 49 / 67 | 5.1 / 4.8 | 5.10 | 162 / 162 | 26.4 / 26.4 |
| 148 | 12, 7 | 108 / 126 | 145 / 137 | 0.83 | 3.9 / 3.1 | 3.9 / 2.5 |
| 147 | 13, 6 | 102 / 177 | 105 / 139 | 0.42 | 5.5 / 3.0 | 3.6 / 2.2† |
| 146 | 14, 5 | 112 / 106 | 44 / 16 | 0.37 | 1.7 / 2.2 | 1.6 / 1.8 |
| 145 | 15, 4 | 155 / 196 | 124 / 135 | 1.39 | 1.9 / 1.6 | 1.8 / 1.7 |
| 144 | 16, 3 | 111 / 139 | 158 / 193 | 0.68 | 2.9 / 2.8 | 2.1 / 2.1 |
| 143 | 17, 2 | 131 / 158 | 154 / 156 | 0.59 | 2.0 / 2.0 | 1.9 / 1.8 |
| 142 | 18, 1 | 132 / 200 | 187 / 234 | 0.37 | 2.7 / 4.4 | 1.8 / 1.9 |
| 141 | 19, 0 | 120 / 127 | 83 / 85 | 0.90 | 1.7 / 2.0 | |
| median | 120 / 140 | 109 / 108 | 0.62 | 2.8 / 2.9 | 2.4 / 2.4 |
Table X. Calibration record of the provider for the twenty qubits of the chain of the two jobs on ibm_fez, in force when the first job ran (last update 09:03 UTC) and when the second ran (19:08 UTC) on 21 September 2026, within two seconds of their submission. Pairs of values refer to the first and the second job, and gives the chain position in each. The readout assignment error was the same in both records. is the error of the gate and that of the CZ gate on the bond to the qubit of the next row, which took 68 ns, or 88 ns where marked by †. In both jobs the germ sits at and the edge rail at to 16.
| run | processor | ran | calibrated | (s) | (s) | (%) | () | rail , | |
| 5 August 2026 | |||||||||
| R1 | ibm_fez | 12 | 10:27 | 09:11 | 112 [52] | 67 [15] | 1.38 [8.2] | 2.6 [3.3] | 3.30, 52 |
| R2 | ibm_fez | 12 | 10:54 | 10:26 | 150 [52] | 94 [18] | 0.99 [4.9] | 2.4 [8.4] | 4.91, 52 |
| R3 | ibm_fez | 12 | 07:53 | 07:38 | 153 [53] | 86 [21] | 0.98 [7.8] | 2.7 [10.3] | 7.85, 91 |
| R4 | ibm_fez | 20 | 10:28 | 09:11 | 110 [52] | 90 [18] | 0.70 [5.0] | 2.9 [5.5] | 4.91, 52 |
| R5 | ibm_fez | 20 | 12:56 | 12:38 | 108 [67] | 109 [24] | 1.06 [13.0] | 2.8 [8.7] | 2.10, 86 |
| R6 | ibm_fez | 20 | 11:04 | 10:26 | 110 [52] | 90 [18] | 0.70 [5.0] | 2.9 [5.5] | 4.91, 52 |
| R7 | ibm_fez | 20 | 13:04 | 12:38 | 108 [67] | 109 [24] | 1.06 [13.0] | 2.8 [8.7] | 2.10, 86 |
| R8 | ibm_fez | 20 | 07:54 | 07:38 | 142 [33] | 75 [21] | 1.04 [7.8] | 2.7 [10.3] | 1.81, 54 |
| R9 | ibm_fez | 50 | 10:25 | 09:11 | 121 [6.2] | 102 [4.7] | 0.84 [32.1] | 2.8 [1000] | 1.73, 48 |
| R10 | ibm_fez | 50 | 08:04 | 07:38 | 134 [33] | 92 [15] | 0.99 [9.2] | 2.7 [39.7] | 8.13, 85 |
| R11 | ibm_fez | 54 | 10:51 | 10:26 | 121 [29] | 99 [4.7] | 0.98 [10.0] | 2.8 [39.7] | 5.42, 74 |
| R12 | ibm_fez | 54 | 15:12 | 13:28 | 121 [29] | 99 [4.7] | 0.96 [48.4] | 2.8 [66.0] | 5.65, 74 |
| injection, | ibm_fez | 20 | 11:05 | 10:26 | 110 [52] | 90 [18] | 0.70 [5.0] | 2.9 [5.5] | 4.91, 52 |
| injection, | ibm_fez | 20 | 11:05 | 10:26 | 110 [52] | 90 [18] | 0.70 [5.0] | 2.9 [5.5] | 4.91, 52 |
| injection, | ibm_fez | 20 | 11:06 | 10:26 | 110 [52] | 90 [18] | 0.70 [5.0] | 2.9 [5.5] | 4.91, 52 |
| injection, | ibm_fez | 20 | 11:06 | 10:26 | 110 [52] | 90 [18] | 0.70 [5.0] | 2.9 [5.5] | 4.91, 52 |
| mirrored Haar gate | ibm_fez | 20 | 12:59 | 12:38 | 108 [67] | 109 [24] | 1.06 [13.0] | 2.8 [8.7] | 2.10, 86 |
| plant at step 3, no control | ibm_marrakesh | 20 | 15:06 | 12:04 | 186 [44] | 99 [13] | 0.77 [3.9] | 2.0 [8.8] | 0.85, 84 |
| plant control | ibm_marrakesh | 20 | 15:12 | 12:04 | 167 [44] | 108 [13] | 1.02 [7.0] | 2.2 [8.8] | 3.08, 82 |
| plant at step 3 | ibm_marrakesh | 20 | 15:24 | 15:04 | 167 [44] | 108 [13] | 1.02 [7.0] | 2.2 [8.8] | 3.08, 82 |
| plant at step 3, repeated | ibm_marrakesh | 20 | 15:27 | 15:04 | 167 [44] | 108 [13] | 1.02 [7.0] | 2.2 [8.8] | 3.08, 82 |
| plant at step 2 | ibm_marrakesh | 20 | 15:44 | 15:04 | 167 [44] | 108 [13] | 1.02 [7.0] | 2.2 [8.8] | 3.08, 82 |
| plant at step 2, paired | ibm_marrakesh | 20 | 15:49 | 15:31 | 167 [44] | 108 [13] | 1.02 [7.0] | 2.2 [8.8] | 3.08, 82 |
| plant at step 2, paired | ibm_fez | 20 | 16:03 | 15:40 | 111 [54] | 93 [18] | 1.10 [5.6] | 2.9 [6.4] | 4.08, 82 |
| stuck readout, r50_S4 | ibm_marrakesh | 50 | 08:39 | 07:56 | 161 [7.0] | 78 [7.0] | 1.03 [43.0] | 3.2 [1000] | 2.58, 141 |
| stuck readout, r50_nodd | ibm_marrakesh | 50 | 08:46 | 08:27 | 161 [7.0] | 78 [7.0] | 1.03 [43.0] | 3.2 [1000] | 2.58, 141 |
| stuck readout, r50_b | ibm_marrakesh | 50 | 10:00 | 09:32 | 161 [7.0] | 78 [7.0] | 1.03 [43.0] | 3.2 [1000] | 2.58, 141 |
| 21 September 2026 | |||||||||
| interleaved | ibm_fez | 20 | 09:32 | 09:03 | 120 [49] | 109 [5.1] | 0.62 [7.0] | 2.4 [27.2] | 1.39, 102 |
| registered | ibm_fez | 20 | 20:22 | 19:08 | 140 [67] | 108 [4.8] | 0.62 [7.0] | 2.4 [45.3] | 0.85, 82 |
| repetition | ibm_marrakesh | 20 | 20:33 | 01:28 | 196 [68] | 113 [31] | 0.69 [9.9] | 1.9 [12.8] | 0.81, 102 |
| repetition | ibm_kingston | 20 | 20:37 | 20:15 | 262 [125] | 167 [30] | 0.60 [3.8] | 2.0 [25.1] | 0.61, 208 |
Table XI. Calibration records of the provider for every run analysed in this work, in force when the run started. The columns ran and calibrated give the time the run started and that of the last update of the record, in UTC. Each record covers all qubits and bonds of the chain of qubits, and the entries give the median over the chain and, in brackets, the smallest and and the largest readout assignment error and CZ error . A CZ error of is the value 1 reported for two bonds in each of the 50-qubit chains of the stuck-readout runs and of R9, and no is reported for qubit 72 of the chain of R9. The last column gives the largest (in %) and the smallest (in s) over the rail sites of the runs of August and over the rail region, positions 13 to 18, of the jobs of September.
Poorly calibrated qubits. The provider’s records show three poorly calibrated qubits on the chain of the two jobs of September on ibm_fez. Qubit 149 had of 49 and 67 , , a readout assignment error of 5.1% and a error of 1.6%, qubit 150 had , and qubit 131 a readout assignment error of 7.0%. The CZ gates between 148 and 149 and between 149 and 150 had errors of 2.6% to 4.5%. These qubits lie at chain positions 7 to 12, ahead of the last rail site at position 13, and none of them contributes a record bit. In every configuration with the late twirl they are therefore exactly outside the support of the record law, by Theorem 2(iii), and this holds for the displaced records as well, since only the line is fed from outside the cone. Only in the configurations of the first job with the twirl at can 149 and 150, at positions 11 and 10, reach the record at all, through non-unital noise passed on by two or three gates to the site that joins the ray at . These are the configurations built to expose such noise (Sec. VI C). Qubit 98, with in the second job, enters only its records at , as a maximally mixed input of the first gate of , on which dephasing has no effect.
The chains of August pass through poorly calibrated qubits as well. The 50-qubit chain of the stuck-readout runs contains, at positions 8 to 16, qubits with readout errors of up to 43% and two bonds with a reported CZ error of 1, and in the 50-qubit chain of R9 qubit 72, at position 28, had a readout error of 32%, and no reported , with a CZ error of 1 on both of its bonds. R12 contains a qubit with a readout error of 48% at position 33, and the chain shared by R5, R7 and the mirrored-gate run one with 13% at position 7. All of them lie at least eight sites ahead of the last rail site, beyond the reach of the record. In these rails the background is randomized only at , so relaxation on the sites ahead of the last rail site, three for , can still reach the record at the last step (Theorem 2(ii)). On those sites the records list below for qubit 64 of R11 and R12, at 32 , and for qubit 12 of R8, at 33 , and these rails nevertheless fit the step-dependent chain at , 0.27 and 0.61 (Table II). Qubit 63 of R11, two sites ahead of its last rail site, had a readout error of 10%, which does not enter the record, since that qubit is not read. Nor do the records single out R6 and R9, the two rails above 3 in Table II, on their rails or within reach of them, and R6 ran on the chain of R4, which fits at 1.27.
Error mitigation, decoupling and twirling. No error mitigation of any kind was applied in any run, and no measurement mitigation. In the four jobs of September, dynamical decoupling, gate twirling and measurement twirling were switched off explicitly in the Sampler options, as the job records confirm. The runbook that recorded the runs of August 2026 could pad idle qubits with a CPMG sequence of two gates, and the circuits held by the provider show where it did. Among the twenty-seven runs of August analysed here only r50.S4 carries the sequence, with delays and gates on idle qubits. The other twenty-six, among them every recorded rail, r50.nodd and every planted-fault run, contain neither delays nor gates. The circuits of the September jobs contain no decoupling sequence, and their only delays are the injected idle periods of the interleaved job, one on each chain qubit in each of the 32 circuits with injection. A decoupling sequence consists of single-qubit unitaries. By Theorem 2 it cannot change the record when it acts off the ray, and on the rail it is absorbed into the step-dependent kernel, so the Markov-structure test of Sec. VI A does not depend on it.
Shots and background. The recorded rails used shots each, split over 16 random patterns of the background randomization. The current runbook prepares the background by a Hadamard gate and a mid-circuit measurement on every background site, which is exact in each shot. The dataset files do not record the preparation method, and the earliest rails were taken with a runbook that randomized the background with random patterns. The Markov test is applied to all of them in the same way. In the interleaved jobs the late twirl was realized by enumerating the Pauli assignments as a parameter sweep with equal shots, and the injected idle period was a delay instruction of 12 496 samples of the device clock (), that is 49.98 .
Chain selection. Chains of qubits were found by randomized depth-first search with a Wansdorff heuristic on the coupling map of the device. Calibration quality was required only on the rail sites, since noise behind the ray is irrelevant by Theorem 2(i), and the remaining qubits were left unconstrained. This is what made a 50-qubit chain schedulable on a device on which 16 qubits failed the cuts.
Transpilation. The two-qubit gate count of every transpiled circuit was compared with the analytic value for the intended layout, and jobs were refused whenever the transpiler inserted SWAP gates, since a rerouted rail voids the exact reference.
Qubit ordering. Qiskit orders qubits little-endian, so a raw appended to wires realizes , the mirror image of the matrix written with wire as the first tensor factor. The mirror is not harmless. For the deployed dressing angles , an systematic at shots. In the recorded dual-unitary rails the gate was compiled from an explicit three-CNOT template on named wires, and rebuilding that template in exact arithmetic gives on the same physical chain, and for the compilation flag used (endian_resolve.py). The Haar-random rails discriminate between the two hypotheses by a factor above 20 in and confirm the assignment independently. In the interleaved demonstrations every gate was appended as unitary(U,[a+1,a]), and the convention was checked against the matrix before submission. Every dataset written by the current runbook records the convention.
Statistics. The primary test is the Markov structure of Corollary 1, fitted to the full distribution of outcomes (Appendix M). It assumes nothing about the gate or the form of the noise. Kernel shifts are estimated from conditional means of the recorded bits, whose readout-induced bias is common to configurations measured on the same qubits and cancels in paired differences. Standard errors are analytic and agree with bootstrap estimates to within a few percent.
Fault localization. Because involves only the gates of steps , the first lag that deviates points to one gate, and shorter lags are probably unaffected. A uniform noise floor deviates from the first lag onward, and it is distinguished from a single fault by the absence of a clean prefix. On chains whose baseline error already breaks the first lag, this absolute localization is uninformative. We then calibrate on the causal prefix and predict the remainder, or take the difference of two runs on the same physical chain.
The interleaved demonstration
Chain and calibration. We used twenty physical qubits of ibm_fez in the order 98, 111, 110, 109, 118, 129, 130, 131, 138, 151, 150, 149, 148, 147, 146, 145, 144, 143, 142, 141, with the germ at chain position , physical qubit 143. The rail region, chain positions 13 to 18, consists of physical qubits 147 to 142, whose reported readout assignment errors at submission were 0.37% to 1.39% and whose times were 102 to 155 µs (Table X). The chain was selected by a randomized search that required readout errors of at most 3% and of at least 60 µs on the rail region, and it was ranked by the summed readout and two-qubit gate errors there.
Configurations. The eight configurations comprise the dual-unitary gate at with either twirl schedule, the Haar-random gate at with either schedule, both with and without the injected idle, and both gates at with the late twirl. Every configuration was run with the germ at and , , prepared by and .
Circuits. Background sites are prepared by a Hadamard gate and a mid-circuit measurement, on every non-germ site for the twirl at and on every site outside the rail for the late twirl. The late twirl is on site immediately before layer , with all assignments of enumerated as a parameter sweep with 64 shots each, which realizes the uniform Pauli average exactly. The injected idle is a delay of 49.98 s, an integer multiple of 16 samples, on every chain qubit after the background preparation. Gates were compiled by the transpiler at optimization level 1 with the chain pinned as the initial layout.
Checks before submission. The emitted gate was compared with its matrix, every circuit was simulated noiselessly at and compared with the exact law, and every transpiled circuit was checked for the absence of SWAP gates, an unpermuted final layout, two-qubit gates only between neighbouring chain qubits, each record bit measured on the intended physical qubit, and intact sweep parameters.
Execution. Each of the 16 circuits was split into 8 blocks, with 32 of the 256 twirl assignments per block for the late twirl and 2048 shots per block for the twirl at . The 128 blocks were interleaved in the order block, configuration, germ sign within one Sampler job (daofjddr85ps73fi7970). No dynamical decoupling, no runtime twirling and no error mitigation were used. The job returned 16 384 shots per configuration and germ sign, 262 144 in total, with all blocks complete.
Acceptance criteria. The criteria, written down before submission, are those stated in Sec. VI B, and all were met. Table XII lists the idle-induced kernel shifts.
| twirl | |||||
| 2 | |||||
| 3 | |||||
| 4 | |||||
| late | 2 | ||||
| late | 3 | ||||
| late | 4 |
Table XII. Idle-induced kernel shifts, injected minus not injected, for the Haar-random gate with both germ signs pooled.
Memory in the first interleaved job
The order-two memory of Eq. (8) is estimated from the recorded bits of each germ sign, and . Standard errors are bootstrap estimates over shots with 400 resamples and agree with the binomial estimate. Table XIII lists every late-twirl configuration of the interleaved job. The prediction is the co-moving law of Theorem 3 without noise. Including the noise calibrated on the dual-unitary record of the same job, a depolarizing strength of 0.012 per layer, relaxation of 0.039 after each readout and assignment errors of 0.0045 and 0.0035, moves the prediction for from 0.052 to 0.045 and gives on four components, against 2.1 without noise. The data do not distinguish between the two.
| Haar, | 50.6 | 2.1 | ||||
| law | ||||||
| DU, | 3.7 | 3.7 | ||||
| Haar, | 2.0 | 2.0 | ||||
| DU, | 7.7 | 7.7 |
Table XIII. Germ-antisymmetrized order-two memory in the first interleaved job on ibm_fez, late twirl, shots per configuration. The last two columns give on four components against zero and against the parameter-free law. Standard errors are 0.008 throughout, and 0.010 for the dual-unitary gate at , . Regenerated by memory_hw.py.
The germ-even part measures something else. On the edge the recorded bit differs from the projected one by the assignment error of its qubit, which makes the recorded chain a hidden chain. Exact simulation of the Haar gate gives a germ-even memory of 0.007, 0.019 and 0.037 for assignment errors of 0.5%, 1.4% and 3%, and a germ-odd part of , and . Pooled over the six configurations, the measured germ-even memory at is , the size expected from the reported assignment error of 1.39% of the qubit read at step 2. The germ-odd part that this error induces is more than an order of magnitude below the statistical error. The analysis was performed after the job and was not among its acceptance criteria.
The registered memory demonstration
Registration. The configurations, gates, seeds, shot numbers, observable, error estimate, noise model, thresholds and criteria K1 to K5 were written down in a protocol before the job was submitted, together with the SHA-256 checksum of the analysis script memory_experiment.py (06e1803b...97a1ea). The script that analysed the data has that checksum. The protocol was not deposited publicly before the job, so its date rests on the author’s records, which give 20:19 UTC on 21 September 2026, three minutes before the job was created. It is available with the code. Before submission we estimated its power on synthetic device data whose noise lies outside the fitted family and is harsher than that of the first job. A correct theory meets K1 to K4 jointly in 96% of 100 synthetic data sets, and the false-failure rate of a null at the threshold is 0.03% to 0.18% in 4000 synthetic data sets, against a nominal 0.1%.
Design. Job daop43opqrnc739aancg on ibm_fez was submitted on 2026-09-21 at 20:22:36 UTC. The chain is 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 138, 131, 130, 129, 118, 109, 110, 111, 98, the qubits of the first job in reverse order, with the germ at chain position , physical qubit 110. The reported readout errors on the rail region were 0.29% to 0.85% and the times 82 to 198 s. The record length is , the late twirl enumerates the 256 Pauli assignments with 128 shots each, the background is measured, and the circuits are interleaved in eight blocks, which gives 32 768 shots per configuration and germ sign. Gate A is the Haar gate of the first job (seed 17 of lightcone_records.haar_unitary). Gate B (seed 97) was selected before any data were taken, as the gate with the largest noiseless memory at both and among seeds 1 to 119 with dual-unitarity defect above 0.5, namely 0.120 and 0.165 against 0.052 and 0.031 for gate A. The circuits were emitted, transpiled and checked by the code of the first job. The column labelled largest in Table IV refers, as registered, to the component that is largest in the noiseless prediction.
Noise. A single model describes the chain, with a depolarizing strength and relaxation on the wires of the strip after every layer, relaxation after every readout, and assignment errors and , fitted jointly to the full distributions of the three records. The fit gives , , , and , at . These five numbers are then inserted into the co-moving law at and . The model is not perfect, as its own shows, and
the total on the 36 components of Table XIV reflects that. Every registered threshold is nevertheless met.
| record | (3, +) | (3, −) | (4, +) | (4, −) | |
| DU, | measured | +0.0065(70) | −0.0001(71) | −0.0009(58) | +0.0003(61) |
| DU, | measured | −0.0028(55) | −0.0041(55) | −0.0079(55) | +0.0044(55) |
| DU, | measured | +0.0001(52) | −0.0065(58) | +0.0065(55) | −0.0093(57) |
| Haar A, | measured | +0.0150(57) | −0.0127(58) | +0.0040(55) | +0.0007(56) |
| Haar B, | measured | +0.0002(55) | +0.0025(55) | −0.0055(55) | +0.0012(55) |
| Haar A, | measured | +0.0404(55) | −0.0030(56) | +0.0216(55) | −0.0077(55) |
| law, chain noise | +0.0370 | +0.0050 | +0.0106 | −0.0023 | |
| law, noiseless | +0.0522 | +0.0048 | +0.0160 | −0.0037 | |
| Haar A, | measured | −0.0217(57) | −0.0184(54) | +0.0095(56) | −0.0124(55) |
| law, chain noise | −0.0148 | −0.0192 | −0.0048 | −0.0055 | |
| law, noiseless | −0.0230 | −0.0314 | −0.0080 | −0.0091 | |
| Haar B, | measured | −0.0736(75) | +0.0911(73) | −0.0033(55) | +0.0054(55) |
| law, chain noise | −0.0718 | +0.0741 | −0.0052 | +0.0088 | |
| law, noiseless | −0.1197 | +0.1179 | −0.0084 | +0.0137 | |
| Haar B, | measured | +0.0103(56) | −0.0111(57) | −0.0353(55) | +0.0985(56) |
| law, chain noise | −0.0027 | −0.0158 | −0.0228 | +0.0888 | |
| law, noiseless | −0.0075 | −0.0245 | −0.0367 | +0.1645 |
Table XIV. Registered memory demonstration. Measured with binomial standard errors in units of the last digit, the law with the chain noise, and the noiseless law. Bootstrap standard errors over shots agree with the binomial ones to 3%, and an independent evaluation from the raw bits reproduces every entry. Regenerated by memory_experiment.py analyze.
Outside the registered criteria. The witness of the dual-unitary gate at , on the new rail qubits, is (0.010, 0.004, −0.005, 0.006) with a standard error of 0.004, null to within , and at it is null to within . This was not among the criteria and is reported as exploratory.
Repetitions. The registration provided for repetitions on a second chain or a second processor, each analysed on its own against the same criteria. The same nine configurations, with the same gates, seeds, shot numbers and interleaving, were submitted to ibm_marrakesh at 20:33 UTC (daop93opqrnc739ab0og) and to ibm_kingston at 20:37 UTC (daopb1n8gn2s739o77g0). Only the chains differ. On ibm_marrakesh the chain is 33, 34, 35, 19, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, with the germ on qubit 2, and on ibm_kingston it is 93, 79, 73, 72, 71, 70, 69, 68, 67, 57, 47, 46, 45, 44, 43, 42, 41, 36, 21, 20, with the germ on qubit 36. Both jobs returned all 589824 shots, and the registered script analysed them without change (Table XV). The noise model fitted to the edge records gives , , , and on ibm_kingston, and , , , and on ibm_marrakesh. The calibration records in force when the two repetitions ran, last updated at 01:28 UTC on ibm_marrakesh and at 20:15 UTC on ibm_kingston, agree with the rail values stored at submission (Table XI). A rerun of the registered analysis on the stored records of the three jobs reproduces every number of their reports to the precision quoted here.
Markov-structure fits
The model is Eq. (4) with per-bit readout assignment errors and common to all steps. Its parameters are the first-step mean , one per germ sign, the kernel , either common to all steps or one per step, and the two assignment errors. They are fitted by maximizing the multinomial likelihood of the full distribution of outcomes under the constraint , from several starting points. Goodness of fit is measured by Pearson’s over cells with expected count above 5, with one constraint per germ sign and one degree of freedom per fitted parameter subtracted. Pooled values are . The rails of Table II are the datasets R1 = r12_clean, R2 = r12_v2, R3 = results_n12, R4 = r20_clean, R5 = r20_generic, R6 = inj_0_0, R7 = r20_haar2, R8 = results_n20, R9 = r50_clean, R10 = results_n50, R11 = r_big and R12 = r_100. Records not pooled there behave as follows. Records with a planted single-qubit rotation give 0.62, 0.65 and 2.83 against 1.93 for the control on the same chain, and the interleaved paired runs give 2.64 and 3.90 against 1.64 and 3.82 for the controls measured in the same jobs, so the plant leaves the quality of the fit unchanged. At shots the residual of planted and control records alike reflects the single readout error that the model assumes for all steps. One planted record on a chain without a same-chain control gives 0.27. Chains containing a qubit with a stuck readout fail even the per-step test, at 110 to 201. Rails with unital noise injected off the ray at give 0.84 to 2.43.
References
References
- [1]D. Hangleiter and J. Eisert, Rev. Mod. Phys. 95, 035001 (2023).
- [2]L. C. G. Govia, P. Jurcevic, S. T. Merkel and D. C. McKay, New J. Phys. 25, 123016 (2023).
- [3]D. Hothem, J. Hines, C. Baldwin, D. Gresh, R. Blume-Kohout and T. Proctor, Nat. Commun. 16, 5761 (2025).
- [4]J. Hines and T. Proctor, Phys. Rev. Lett. 134, 020602 (2025).
- [5]Z. Zhang, S. Chen, Y. Liu and L. Jiang, PRX Quantum 6, 010310 (2025).
- [6]T. Proctor, K. Young, A. D. Baczewski and R. Blume-Kohout, Nat. Rev. Phys. 7, 105 (2025).
- [7]B. Bertini, P. Kos and T. Prosen, Phys. Rev. Lett. 123, 210601 (2019).
- [8]L. Piroli, B. Bertini, J. I. Cirac and T. Prosen, Phys. Rev. B 101, 094304 (2020).
- [9]B. Bertini, P. W. Claeys and T. Prosen, Rev. Mod. Phys. 98, 025001 (2026).
- [10]L. E. Fischer, M. Leahy, A. Eddins, N. Keenan, D. Ferracin, M. A. C. Rossi, Y. Kim, A. He, F. Pietracaprina, B. Sokolov, S. Dooley, Z. Zimborás, F. Tacchino, S. Maniscalco, J. Goold, G. García-Pérez, I. Tavernelli, A. Kandala and S. N. Filippov, Nat. Phys. (2026), doi:10.1038/s41567-025-03144-9.DOI
- [11]X.-H. Yu, Z. Wang and P. Kos, Quantum 8, 1260 (2024).
- [12]M. A. Rampp, S. A. Rather and P. W. Claeys, arXiv:2606.23803 (2026).arxiv.org/abs/2606.23803
- [13]H.-R. Wang, X.-Y. Yang and Z. Wang, Phys. Rev. Lett. 133, 170402 (2024).
- [14]F. Hübner and S. W. P. Kim, arXiv:2606.12538 (2026).arxiv.org/abs/2606.12538
- [15]P. W. Claeys and A. Lamacraft, Phys. Rev. Research 2, 033032 (2020).
- [16]S. Singh and S. Dooley, arXiv:2605.23622 (2026).arxiv.org/abs/2605.23622
- [17]J. Li, R. Li, J. Jack, B. Bertini and J. P. Garrahan, Phys. Rev. B 111, 024309 (2025).
- [18]P. W. Claeys, M. Henry, J. Vicary and A. Lamacraft, Phys. Rev. Research 4, 043212 (2022).
- [19]F. Ciccarello, S. Lorenzo, V. Giovannetti and G. M. Palma, Phys. Rep. 954, 1 (2022).
- [20]C. W. Gardiner, Phys. Rev. Lett. 70, 2269 (1993).
- [21]H. J. Carmichael, Phys. Rev. Lett. 70, 2273 (1993).
- [22]A. Lerose, M. Sonner and D. A. Abanin, Phys. Rev. X 11, 021040 (2021).
- [23]H.-R. Wang, I. Vilkoviskiy and D. A. Abanin, arXiv:2510.21927 (2025).arxiv.org/abs/2510.21927
- [24]I. Vilkoviskiy and D. A. Abanin, arXiv:2511.03846 (2025).arxiv.org/abs/2511.03846
- [25]F. A. Pollock, C. Rodríguez-Rosario, T. Frauenheim, M. Paternostro and K. Modi, Phys. Rev. Lett. 120, 040405 (2018).
- [26]A. Foligno, T. Zhou and B. Bertini, Phys. Rev. X 13, 041008 (2023).
- [27]M. Frías-Pérez and M. C. Bañuls, arXiv:2201.08402 (2022).
- [28]P. Kos, B. Bertini and T. Prosen, Phys. Rev. X 11, 011022 (2021).
- [29]H. Jaeger, Neural Comput. 12, 1371 (2000).
- [30]F. Vatan and C. Williams, Phys. Rev. A 69, 032315 (2004).
- [31]S. H. Pickering, M. McGinley, B. Kumar and B. Bertini, arXiv:2607.25969 (2026).arxiv.org/abs/2607.25969
- [32]M. Ippoliti and V. Khemani, Phys. Rev. Lett. 126, 060501 (2021).
- [33]T.-C. Lu and T. Grover, PRX Quantum 2, 040319 (2021).





