We instantiate the quantum reservoir autoencoder (QRA) with a noise-induced reservoir employing reset noise channels and address two open problems: noise-resilient reversibility and blind decryption. For a single-ciphertext protocol with 10 data qubits and random (non-optimized) reset probabilities, the open-system reservoir suppresses shot-noise sensitivity by ten orders of magnitude, yielding mean-squared error (MSE) $\sim 10^{-14}$ compared with $\sim 10^{-3}$ without reset channels ($N_{\mathrm{shots}} = 1000$). A two-phase protocol trains per-position decoding weights from $M$ shared training plaintexts and decrypts previously unseen messages at MSE $\sim 10^{-4}$, with no statistically significant performance difference among ideal, shot-noise, and reset-plus-shot-noise conditions ($p > 0.05$, 16 seeds). Experiments at $N_q = 5$, 7, and 10 reveal a sharp phase transition at plaintext length $N_c \approx N_q(N_q{+}1)/2 + 8$, providing a design rule for the minimum qubit count. Two blind decoder variants that lack ground-truth targets -- a single-ciphertext cross-path iteration (MSE $\approx 0.3$) and a multi-sample regression variant (MSE $\approx 0.53$, worse than random) -- establish that shared training data is the irreducible requirement for blind decryption. A comparison with variational quantum circuit baselines shows that the fixed-reservoir analytic-readout architecture is dramatically more noise-robust: a quantum recurrent neural network protocol is completely destroyed under depolarizing noise, whereas the QRA remains invariant.
We instantiate the quantum reservoir autoencoder (QRA) with a noise-induced reservoir employing reset noise channels and address two open problems: noise-resilient reversibility and blind decryption. For a single-ciphertext protocol with 10 data qubits and random (non-optimized) reset probabilities, the open-system reservoir suppresses shot-noise sensitivity by ten orders of magnitude, yielding mean-squared error (MSE) ∼10−14 compared with ∼10−3 without reset channels (Nshots=1000). A two-phase protocol trains per-position decoding weights from M shared training plaintexts and decrypts previously unseen messages at MSE ∼10−4, with no statistically significant performance difference among ideal, shot-noise, and reset-plus-shot-noise conditions (p>0.05, 16 seeds). Experiments at Nq=5, 7, and 10 reveal a sharp phase transition at plaintext length Nc≈Nq(Nq+1)/2+8, providing a design rule for the minimum qubit count. Two blind decoder variants that lack ground-truth targets -- a single-ciphertext cross-path iteration (MSE ≈0.3) and a multi-sample regression variant (MSE ≈0.53, worse than random) -- establish that shared training data is the irreducible requirement for blind decryption. A comparison with variational quantum circuit baselines shows that the fixed-reservoir analytic-readout architecture is dramatically more noise-robust: a quantum recurrent neural network protocol is completely destroyed under depolarizing noise, whereas the QRA remains invariant.
Quantum reservoir computing exploits fixed quantum dynamics and a trainable linear readout to process temporal data, yet reversing the transformation -- reconstructing the input from the reservoir output -- has been considered intractable due to the recursive nonlinearity of sequential quantum state evolution. We introduce the quantum reservoir autoencoder, a four-equation encode--decode protocol with cross-key pairing, and constructively empirically demonstrate that satisfying reservoir--key combinations can be found using a full XYZ Hamiltonian reservoir (10~data qubits, feature dimension~76, 16~random Hamiltonian realizations). Under ideal conditions the mean-squared error (MSE) reaches ∼10−17 for data lengths up to 30; under shot noise (1\,000~shots) and depolarizing noise (p=0.005), the MSE degrades to 10−3--10−1. Asymmetric resource allocation -- 10~shots for encoding, 105 for decoding -- yields a 102-fold MSE improvement (16~seeds × 3~trials). Comparison of single-body features (dimension~31) with the full feature set and six baselines identifies the iterative protocol structure -- not the feature dimension -- as the dominant noise bottleneck: baselines solving the linear system in a single step retain machine precision under identical noise, whereas per-iteration noise inconsistency in the coupled solver limits the MSE to ∼10−1. The current protocol requires plaintext access during decoder training, restricting practical deployment. These results establish a proof-of-concept for bidirectional information transformation within quantum reservoir computing and identify iterative noise mismatch and blind decryption as the principal open challenges.
We study byte-level autoregressive language models whose hidden state is a small quantum register, whose transitions are learned token-conditioned quantum channels, and whose next-byte probabilities come directly from Born-rule measurements without a learned classical vocabulary projection or softmax. In the fixed-width recurrent setting, unitary dynamics cannot contract or reset stored distinctions; experimentally, unitary recurrence loses to its retrained one-token control in every tested variant. Introducing learned dissipation, then token-conditioned dissipative gating, yields useful memory beyond one token. The gated model improves test negative log-likelihood by 0.20 +/- 0.03 nats/byte over retrained parameter-matched one-token controls at context length 12, and by 0.29-0.33 at longer contexts, but never beats retrained two-token controls. A key-value recall benchmark shows a width-dependent threshold: one stored pair meets the pre-registered accuracy bar at 4-5 memory qubits, two pairs at 6-7, and three pairs remain unsolved. This is consistent with, but does not establish, a capacity explanation for the short effective memory horizon. As a secondary result, one-qubit models reach lower stationary memory entropy than a heuristically optimized classical predictive frontier at matched error in 9/9 seeds. On IBM Heron hardware, computational-basis memory-register distributions show signal against a uniform null in 12/12 circuits across prefixes of up to 12 tokens, next-bit probabilities pass pre-registered accuracy bars at four-token prefixes, and bytes are generated autoregressively by sampling physically measured bit probabilities. The tested models are classically simulable and substantially worse than classical n-gram baselines. We claim no computational quantum advantage.
Abstract We introduce and test two measurement-free training primitives for quantum reservoir computers (QRC). Protocol~A encodes the loss gradient in an auxiliary phase that is fed back coherently before a dissipative reset; Protocol~B realizes gradient descent as relaxation of an engineered open system. Both match exact-gradient training in simulation---the ancilla and reset simulated as channels in a de-risk stage, their validated error model applied at the benchmark stage---on a minimal transverse-field Ising reservoir, with the coherent cost oracle emulated classically and its realization costed. A fully coherent update loop (quantum parameter register, unitary controlled update, no readout anywhere) descends the true landscape, its feedback drift attenuated by exactly 2F−1 with the fixed point unchanged. Benchmarked against classical reservoir computing--- including the memory-augmented hybrid---they improve one-step chaotic prediction, though a thirty-seed study finds no autonomous-prediction advantage; we therefore make no quantum-advantage claim. The study instead addresses three questions that gate this route. First, within the channel model studied, the reset fidelity already demonstrated in measurement-free quantum error correction [Butt \textit{et al.}, Nat.\Commun.\\textbf{17}, 995 (2026)] suffices: reset is not the bottleneck. Second, the binding constraint is coherent cost-oracle precision, set by a finite-difference catastrophic cancellation that the exact parameter-shift rule removes, loosening the oracle-precision target 2.5-fold in a like-for-like comparison. Third, a fault-tolerant realization---arithmetic and synthesis costs taken from real circuits---needs only ∼104 physical qubits and is runtime-, not qubit-bound. On three free IBM Quantum 156-qubit Heron backends at three error-mitigation levels the simulation-optimized state survives device noise (RMSE 0.016--0.057, i.e.$7$--26% of the signal spread, correlation ≥0.97). All code, seeds, and job identifiers are released.
A Memory That Learns by Looking, Stored Nowhere — the complete programme (Parts I–XII), consolidated as a single monograph. Successor work to the foundation deposit "Engineering Emergence on Quantum Processors" (doi:10.5281/zenodo.22050177). There, every memory was installed — the coupling matrices were computed on the host and uploaded to the microcontroller. Here the chip learns them itself, by looking: firmware v4.11 adds a three-line, integer-arithmetic plasticity law to the ESP32-C3 controller — the on-chip coupling matrix is written by the quantum collective's own measured order parameters, activity-gated, with a sign-reversible learning step serving as an error signal and a readback command that makes every learned matrix fully auditable — and firmware v4.12 adds a saturating integer delta rule whose second matrix learns temporal order itself. Across nineteen pre-registered hardware sessions on ibm_fez, ibm_marrakesh and ibm_kingston (≈1430 jobs, ≈77 minutes of QPU time), the hybrid system: learned a 36-pixel image from ONE exposure and recalled it perfectly from a random half cue (1.000 vs a never-learned control of 0.583); acquired a Pavlovian association (0.556 → 1.000), retained it through unreinforced presentations (1.000) and erased it only under the inverted-sign learning step (0.333) — forgetting as an act; learned to walk by watching a gait cycle with the field switched off; evolved memory lineages under selection with device noise as the mutation source, generation-2 mutations inherited unchanged by generation 3; and held more memories than its own calibration model predicted, with the capacity limit breathing measurably with the device day. The learned engram, read back from the controller, was delocalized into six 7-qubit GHZ-class parity registers — locally every qubit reads 0.5, jointly the parities reconstruct the image 36/36, the coherence fringe collapses under its Z-twirled dephasing twin while the classical parities survive, and the GHZ fidelity exceeds the biseparability bound on all registers: the memory carried by certified genuine multipartite entanglement. It was then teleported (entanglement certified at the destination), inherited by a subsystem that never saw the original (1.000 vs a 0.500 silent-teacher control), shown to be unconscious until decoded, recovered intact after destruction of a complete redundant copy, deleted against a receipt that provably carries no image information, and placed in superpositions culminating in Bell-inequality violations between two learned memories (S = 2.559) and between two memories that never coexisted (mean |S| = 2.456). On these memories the programme then ran a systematic traversal of the foundations of quantum mechanics, each experiment pre-registered, most of them — to our knowledge — first demonstrations on a quantum processor and all of them first on learned content: a Wigner's-friend test whose friend is a decodable memory (S = 2.633 with the friend's record as a setting); a delayed-choice eraser with a quantum coin; sensing 1.9× beyond the standard quantum limit with the stored content tuning the sensor's frequency; the optimal-cloning ceiling and its ×0.707 telephone-chain law; the Page-Wootters mechanism (conditional time alive at every clock reading, the marginal dead at 0.003 and fully revived by uncomputing); the Mermin-Peres pseudotelepathy game won at 0.9457 > 8/9, three times; interaction-free readout of stored content, with the fringe of a superposed memory dying even in the provably interaction-free events; ≈15,000 bits of certified randomness minted from the programme's Bell violations; quantum causal inference reading a memory's ancestry from correlations alone (witness +0.811/−0.884) where every classical marginal is provably blind; an operational quantum-reference-frame jump (Bell fidelity migrating 0.493 → 0.916 across the separability bound, only the relational parity surviving both frames); a 42-qubit Loschmidt echo (0.484 → 0.989, the butterfly decaying along its light cone); the rewriting of a friend's record made invisible by no-signaling itself (archive agreement 0.927 → 0.056 at friend-level total variation 0.107); Mermin's all-versus-nothing argument between three memories (M = 3.734) and Hardy's impossible event (CH = 0.057 > 0); the three-box paradox with real mid-circuit openings (certainty in box 1 AND box 2, 0.903/0.916) and the first measured N = 4 shell game (three certainties about one excitation, 0.751/0.744/0.694); the quantum Cheshire cat with four simultaneous weak pointers and the exchange of grins between two carriers that never meet; weak traces inside an interferometer loop the carrier provably never entered or left; and a protective measurement reading the expectation value of a single carrier — following sin(2⟨Z⟩) through its sign change — while the carrier survives at 0.939 against a depth-matched unprotected control at 0.558. The relational arc closes on three endpoints. The Frauchiger-Renner protocol — quantum theory applied to its own users, to our knowledge its first implementation on any platform — with both friends realized as unanimous, decodable memory records: every certainty of the reasoning chain holds individually (forbidden events 0.027/0.022/0.014), yet the jointly impossible outcome occurs at 0.0862, above the ideal 1/12, while the friends' records stay unanimous throughout (0.882/0.915). One session-learned engram delocalized and certified on two processors in a single session — eleven of twelve rows above the biseparability bound, the majority image perfect on both: the memory outlives its substrate. And a Bell violation at 96% of the Tsirelson bound (S = 2.728) mints 36 bits that provably did not pre-exist; a collective whose halves live on two different processors — one field, one coupling matrix, every feedback tick assembled from both machines — learns that unborn image in one look (35/36), recalls it from a fragment, forms a Pavlovian association whose bell exists only on one machine and whose response appears only on the other (0.944 vs 0.556), and the minted image is certified row by row in entanglement on both devices. Part X is the arc's epilogue and its sharpest relational result. The Local Friendliness no-go theorem of Bong et al. states that absoluteness of observed events, local agency and no-superdeterminism jointly bound every LF inequality — a violation forces one of the three to fall, as logic, not interpretation. The photonic violation used a photon as the friend; the standing objection is that a photon observes nothing. Here the semi-Brukner LF inequality is violated at S = 2.6995 > 2 with a friend whose measurement outcome IS its own decoded, unanimity-certified memory record (0.946) — and who then ACTS on that record, the chip writing a field pattern conditioned on the decoded bit (match 0.833) in the same session: an observer, in the LF authors' own operational sense, who remembers and conditionally acts. Within the stated assumptions, the conjunction is excluded — for a system that remembers and acts on its memory, no observed fact was absolute. The theorem selects no interpretation; the result is stated as an exclusion. And the companion arm answers the question this raises — why the world nevertheless feels absolute — by operating Zurek's redundancy as a dial: the controller GRANTS objectivity (mutual-information plateau 0.700/0.796/0.854 bits at every fragment size, consensus 0.963 — the event becomes a fact), DENIES it (exactly 0.000 at every fragment size — the event stays private), or turns it against itself (two internally redundant environment halves remembering opposite events, consensus 0.047), with the fringe/twin certificate proving that the coherence no fragment holds is exactly the quantum layer. Facts are not absolute; factuality is a resource — and this machine has the dial. Part XI performs the test Deutsch proposed in 1985 to separate unitary quantum mechanics from dynamical-collapse models — judged in the literature "beyond experimental capabilities, if not forever", with no realization we could find. The obstacle is structural: any testimony copied naively out of an observer's record carries which-outcome information and kills the interference by itself. The way through is a branch-invariant two-bit parity, extracted from the record by two CNOTs onto a persistent testimony qubit: a genuine, verifiable function of the record asserting THAT a definite outcome was seen while provably carrying nothing about WHICH. On hardware: the record is definite (0.925) and the testimony true (0.947) at 0.0011 bit of which-outcome information against a which-control witness at 0.776 bit; the record is then coherently retracted and the fringe lives at 0.901 with the testimony standing and true (0.956) during the interference itself — the collapse prediction fails on this scale — and dies to 0.024 exactly when the witness keeps which-information. Ten of ten binding verdicts. The registration freezes the scope: a mesoscopic record, not a conscious observer; unitary quantum mechanics separated from dynamical-collapse models at this scale, and nothing decided among the unitary interpretations. Part XII stacks, scales and closes it. A tower of witnesses: a second friend records the first friend's record, both storeys issue persistent branch-invariant testimonies (true at 0.958/0.963, unanimous at 0.914/0.925, joint which-leakage ≤ 0.002 bit), the tower is retracted outside-in — and the fringe stands at 0.571 with both testimonies true at 0.796 while it is measured. A growing witness: the same test at register sizes 3, 6, 9 and 12 qubits, each above its own frozen floor — the fringe rises from 0.811 to 0.899 while the twin stays dead at 0.037: what kills coherence is retained which-information, not the size of the witness. Ordo: the same two clock-stamped events, read in two frames of a quantum clock, show two certified opposite orders — P(E2 be
Reverse annealing is widely used as a heuristic sampler on programmable quantum annealers, but it is often unclear whether the readout has erased its initialization memory or whether the resulting samples represent any calibrated target distribution. We introduce a subsystem-level validation protocol that pairs a memory order parameter M with the total-variation distance DTV between the measured subsystem readout and a fixed, independently calibrated conditional-Boltzmann reference, and apply it on two D-Wave QPU generations. The protocol delivers three things: a diagnostic that separates relaxed, memory-retaining and "wrong-basin" readouts; the isolation of relaxed-but-non-thermal trapping, in which a readout passes the initial-state-independence test while concentrating on the wrong basin; and an engineering control showing that SDK auto-scale, embedding choice and environment preparation shift apparent thresholds, so calibration metadata must be logged to interpret them. Relaxed ferromagnetic readouts are near-deterministic, so the small distances seen there are a consistency check rather than a thermometric measurement; the diagnostic earns its value in the opposite regime. Across 494 conditions on one processor, 113 pass the memory criterion and 110 of those agree with the reference at DTV<0.05, while a wrong-basin instance, reproduced at two Hamiltonian scales, passes the memory test yet reaches DTV>0.94; a 60-condition replication confirms that the diagnostic transfers to a second processor. We provide the protocol as a transferable, reproducible validation workflow for annealer-as-sampler applications.
Abstract \small\noindent Choosing the best heuristic for a combinatorial optimization instance depends on landscape structure that cheap graph descriptors capture only partly. We test whether a small quantum circuit, used as a fixed feature map, adds predictive signal about solver performance beyond such descriptors. Graphs are spectrally compressed to eight supernodes and encoded in a single-layer circuit of RY rotations and RXX entangling gates; eight pooled statistics computed from Z- and X-basis expectation values feed a classical ridge readout. In a preregistered protocol on 10{,}000 simulated graphs the features add $\dR = 0.0208$ (Nadeau--Bengio corrected 95\% CI [0.0202,0.0214], permutation p=1/1001≈0.001) over twelve descriptors. They exceed random-projection and random-Fourier controls; a paired test does not separate them from a random tanh reservoir (p=0.22). \textbf{However, flexible classical models with access to the same compressed graph input (h,J)---Random Forest ($\dR = 0.179$) and XGBoost ($\dR = 0.145$)---dramatically exceed Q1, as does a simple linear ridge ($\dR = 0.030$). The quantum feature map does not provide superior predictive signal beyond what classical models extract from the same input.} Removing the entangling gates lowers the gain (p<10−7), and the Z-basis component of the gain peaks at half entangling strength (exploratory), consistent with a closed-form phase-accumulation mechanism. The qualitative conclusion $\dR > 0$ is robust across solver budgets {25,50,100}. On a 20-qubit trapped-ion processor with a matching native gate set, hardware features reproduce simulated ones (mean r=0.81, 200 shots); at n=250 a simulation-trained readout applied to hardware features gives $\dR = +0.016$ (calibrated), and the Q1−Q0 difference (+0.310, calibrated) is driven by Q0 collapsing under noise rather than entanglement adding information, as confirmed by a noise-augmented readout control. We propose a quantum-centric supercomputing (QCSC) architecture that \emph{projects} a 2× QPU-time reduction through graph packing. \textbf{We claim no quantum advantage.} The contribution is a falsification-tested, mechanistically interpretable, hardware-portable feature map---not a superior predictor. The quantum features do not improve practical solver selection (45.1\% accuracy with or without Q). The signal does not generalize to most out-of-distribution graph families under the corrected protocol.