{"id":"da17f566-41f5-4b36-a9a3-0c35fbb50d89","arxiv_id":"2506.03286","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Demonstrates a two-mode SRF cavity platform with 20.6 ms and 15.6 ms lifetimes, error-resilient sideband control, Fock state preparation up to n=20 with >95% post-selected fidelity, and a virtual Raman beamsplitter with inferred coherence-limited fidelity up to 99.9%.","lead":"A two-mode superconducting cavity controlled by a transmon achieved record-long coherence times of 20.6 and 15.6 milliseconds, and used error-correcting feedback to prepare Fock states up to 20 photons with over 95% post-selected fidelity. The work shows that ultrahigh-Q cavities can be programmed without destroying their memory advantage, a step toward qudit-based quantum computing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Beamsplitter fidelity is a fitted model parameter, not a measured quantity; direct two-mode process tomography is needed before the 99.9% claim can be accepted.","rationale":"The reader's weakest_assumption targets exactly the coherence-limited fidelity extraction. The paper is otherwise strong: it reports credible coherence measurements (20.6 ms and 15.6 ms lifetimes, dephasing >40 ms), a genuine feedforward protocol, and reproducible open-system simulations. The Fock state fidelities are post-selected but are directly measured populations. The load-bearing weak point is the VRBS fidelity: it is a model-inferred quantity, not a directly measured gate fidelity. The paper itself admits (Appendix H) that control errors are uncalibrated and excluded from simulations. Since the 99.9% claim is the headline control achievement, it should be substantiated by direct two-mode tomography or at least by a randomized benchmarking-style protocol. This does not invalidate the coherence or the protocol, but it changes the strength of the central claim from 'demonstrated' to 'coherence-limited estimate.' Therefore the CONDITIONAL verdict is appropriate: accept conditionally on providing a direct gate fidelity measurement.","tokens_in":35554,"tokens_out":1497,"duration_ms":16052,"concrete_test":"Perform randomized two-mode process tomography (or quantum state tomography of the output Bell state) for the VRBS gate at the optimal detuning, without post-selection and with the mapping confusion matrix (M.1) corrected. Compare the measured process fidelity to the predicted F* ≈ 99.810 ± 0.003% from Eq. (2). If the measured fidelity falls more than ~1% below the predicted value, the coherence-limited formula is not capturing the dominant errors and the headline 99.9% claim must be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central control claim rests on Eq. (2): F* ≈ 1 - (π/4)(κBS/gBS), where κBS = κ1 + κφ/2 and gBS are extracted from a fit of Eq. (1) to damped Rabi oscillations in the single-photon subspace. This formula assumes the only error channels are Markovian exponential decay and dephasing of the oscillation amplitude; all preparation, mapping, and readout errors are taken to be perfect. The paper's own numerical simulation (Appendix J) omits control errors, and Appendix H explicitly states 'control errors are not calibrated, and thus are not included in the noise model for simulation.' Since the experimental scatter in Fig. 3(b) shows deviations from the simulated curve at small detuning and the heating rate is ~1% per swap, the fitted κBS already absorbs any residual non-Markovian or state-preparation error. The headline 'approaching 99.9%' is therefore an idealized upper bound from a two-parameter fit, not a directly verified gate fidelity. A direct two-mode process tomography in the single-photon subspace would settle whether the actual process fidelity matches the coherence-limited bound, or whether unseen control errors (drive crosstalk, Stark-shift miscalibration, sideband drive imperfections, mapping confusion matrix) limit it to a lower value.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports a two-mode superconducting radio-frequency (SRF) cavity platform in which a transmon ancilla is weakly coupled to two cavity modes while preserving long coherence. The authors measure single-photon lifetimes of 20.6±0.4 ms and 15.6±0.2 ms for the two modes and extract pure dephasing times above 40 ms, comparing against bare-cavity baselines. They introduce a sideband feedforward protocol (SFP) with a parity filter to prepare Fock states up to |20> with post-selected fidelities above 95%, and they demonstrate a virtual-Raman-assisted beamsplitter (VRBS) interaction in the single-photon subspace. From fits of damped Rabi oscillations to Eq. (1), they report coherence-limited beamsplitter fidelities up to 99.810±0.003% after erasure post-selection and infer a fidelity approaching 99.9% when mid-circuit heating checks are added. The paper includes extensive appendices on fabrication, noise modeling, simulation, and calibration.","tokens_in":35981,"tokens_out":2553,"duration_ms":30976,"significance":"If the coherence and control claims hold, this is a substantial experimental advance: the reported multimode single-photon lifetimes are an order of magnitude beyond previous 3D multimode memories, and the demonstration of high-fidelity Fock-state preparation up to n=20 in a weak-coupling architecture is a useful step toward qudit processors. The baseline bare-cavity characterization, the sideband feedforward error-correction scheme, and the repeated-readout checks that cavity lifetimes are unaffected are concrete strengths. The paper is also honest in its appendices about where control errors are not calibrated and where simulations omit them. The main reservation is that the headline beamsplitter fidelity is an inferred, coherence-limited quantity rather than a directly measured process fidelity, and the Fock-state fidelities are post-selected with the success probability not fully reported.","major_comments":[{"comment":"The claim of a beamsplitter fidelity approaching 99.9% is not a directly measured process fidelity. It is computed from F*_BS ≈ 1 - (π/4)(κ_BS/g_BS), where κ_BS = κ_1 + κ_φ/2 and g_BS are extracted from a fit of Eq. (1) to damped oscillations in the single-photon subspace. This formula assumes that the only error channels are exponential decay and dephasing of the oscillation amplitude and that state preparation, sideband pulses, and the final mapping/readout are error-free. The manuscript itself notes in Appendix H that control errors are not calibrated and are not included in the noise model, and the simulations in Appendix J omit control errors. Because the heating rate is ~1% per swap and the experimental data deviate from the simulated curve at small detuning, the fitted κ_BS can absorb residual non-Markovian or control errors. A direct two-mode process tomography in the single-photon subspace, or at least a randomized-benchmarking-style comparison, is needed before 'approaching 99.9%' can be stated as an experimental gate fidelity rather than an idealized upper bound.","section":"Sec. IV, Eqs. (1)-(2), Fig. 3(c)-(d)"},{"comment":"The reported Fock-state preparation fidelities (95.3±1.9% for Alice |20> and 94.6±1.1% for Bob |20>) are post-selected by the parity filter, but the paper does not report the corresponding post-selection success probability or the raw (un-post-selected) fidelity for the SFP+PF protocol. Since the parity filter discards shots with odd total photon number, the success probability directly determines the practical overhead of the protocol. Without this quantity, a reader cannot assess whether the 'error-resilient' preparation is efficient or whether the headline fidelity is obtained at a large acceptance cost. Please report the parity-filter success rate (Fig. A5 provides one but not in the main text) alongside each SFP+PF fidelity, or state explicitly that the reported fidelities are conditional on passing the filter.","section":"Sec. III, Fig. 2(c)-(d)"},{"comment":"The mapping from the two-mode states |10>, |01>, and |00> to transmon levels uses the confusion matrix in Eq. (M.1), with substantial off-diagonal elements: P(measured |e> | prepared |10>) ≈ 0.049 and P(measured |f> | prepared |01>) ≈ 0.017. The main text does not state clearly whether the population oscillations shown in Fig. 3(c) are corrected using this confusion matrix before fitting to Eq. (1). If the correction is applied, its propagation into the extracted κ_BS and F*_BS uncertainties should be described; if not, the 99.810±0.003% value inherits uncontrolled readout-mapping errors. Please specify the exact processing of the measured populations and include the mapping confusion in the error budget.","section":"Appendix M and Sec. IV, Fig. 3(c)"}],"minor_comments":[{"comment":"The abstract's 'approaching 99.9%' should be qualified as a coherence-limited inferred fidelity, not a measured process fidelity, to match the body text and Appendix disclosures.","section":"Abstract and Sec. IV"},{"comment":"There is a typo in 'dispserive interaction'; it should read 'dispersive interaction'.","section":"Sec. II"},{"comment":"In the text after Eq. (B.2), 'decay pfP_PS' should read 'decay of P_PS'.","section":"Appendix B"},{"comment":"The phrase 'beasmplitter rate' should be corrected to 'beamsplitter rate'.","section":"Appendix I"},{"comment":"The text twice says 'Hanh-echo experiment'; the standard spelling is 'Hahn-echo'.","section":"Appendix A"},{"comment":"The caption contains '0/uni27E9' and '1/uni27E9', which appear to be encoding artifacts from the |0> and |1> ket symbols; these should be typeset properly.","section":"Fig. A7 caption"}],"recommendation":"major_revision","confidential_remarks":"The experimental work itself is strong and the appendices are unusually transparent about model limitations. My main concern is that the most eye-catching quantitative claim, the ~99.9% beamsplitter fidelity, is an inferred upper bound rather than a directly characterized process fidelity. I believe this is fixable within a revision by adding a direct two-mode process-tomography measurement or by reframing the claim as a coherence-limited bound with the necessary caveats in the abstract and conclusion. I do not see evidence of circular reasoning or fabrication; the self-citations are frequent but appear related to the SQMS program and prior work on the same hardware."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Name],\n\nI've read the manuscript carefully. The punchline: this is a strong experimental paper that deserves a serious referee, but the headline '99.9%' entanglement fidelity is an inferred coherence-limited upper bound, not a measured gate fidelity.\n\nWhat's genuinely new: record multimode coherence in a two-cell SRF cavity (20.6/15.6 ms T1, dephasing >40 ms) with a weakly coupled transmon, achieved by careful participation engineering. The sideband feedforward protocol (SFP) is a real contribution: it uses three-state transmon readout to correct ancilla errors during Fock ladder climbing, and the data show clear improvement (|20> population from ~0.51 to ~0.83 pre-parity-filter). The virtual Raman beamsplitter is a nice theoretical idea with an experimental proof-of-principle, and the derivation in Appendix I is thorough. The paper is also honest: it states upfront that control errors are not calibrated and that fidelities are coherence-limited estimates.\n\nThe soft spots, in order. First, the Fock-state fidelities exceeding 95% are post-selected on the parity filter; the unconditional SFP-only population for |20> is ~0.83. That's still a strong result, but the abstract could mislead. Second, the 99.9% beamsplitter fidelity is not a direct process fidelity. It comes from fitting Eq. (1) to damped Rabi oscillations and plugging the fitted rates into Eq. (2), which assumes Markovian decay/dephasing and perfect control. Appendix J's simulation explicitly omits control errors. So 'approaching 99.9%' is an upper bound, not a verified gate fidelity. The heating-check post-selection also discards ~4% of shots, so the reported number is conditional. These are not fatal flaws--the authors label them clearly--but a referee should ask for direct two-mode process tomography (or at least randomized benchmarking) in the single-photon subspace, and for unconditional Fock-state fidelities.\n\nWho this is for: anyone working on bosonic error correction, qudit processors, or high-coherence cavity QED. The engineering and the SFP protocol will be useful even if the fidelity numbers are refined.\n\nRecommendation: accept the paper, but only after the authors either (a) provide direct tomography of the beamsplitter/swap gates to back up the coherence-limited estimate, or (b) clearly rephrase the abstract and conclusions to avoid implying a measured 99.9% gate fidelity. The central engineering and protocol contributions stand on their own.\n\nRegards,\n[You]","headline":"Strong experimental platform with a genuinely new feedforward protocol and record multimode coherence; the 99.9% entanglement fidelity is an inferred bound, not a measured gate fidelity.","tokens_in":36468,"tokens_out":3205,"would_cite":true,"duration_ms":34383,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A weakly coupled transmon can control ultracoherent two-mode superconducting radio-frequency cavities well enough to prepare Fock states up to $|20\\rangle$ with over 95 percent post-selected fidelity and to entangle the modes with…","keywords":["superconducting radio-frequency cavity","bosonic qudit","transmon ancilla","sideband feedforward","Fock state preparation","virtual Raman beamsplitter","erasure post-selection","two-mode cavity"],"falsifier":"Run full quantum process tomography or interleaved randomized benchmarking on the virtual Raman beamsplitter in the dual-rail subspace $\\{|10\\rangle, |01\\rangle\\}$, without discarding erasure or heating shots, and compare the process fidelity to $F^*_{BS}$ from Eq. (2). If the tomographic fidelity falls short by more than the quoted 0.2 percent infidelity, the missing control errors are the dominant limitation and the 99.9 percent figure is an upper bound rather than a realized gate fidelity.","tokens_in":35386,"feed_emoji":"🧊","tokens_out":7840,"duration_ms":83579,"temperature":0.7,"pith_summary":"This paper reports a two-mode niobium superconducting radio-frequency cavity in which a weakly coupled transmon ancilla controls two long-lived modes without destroying them. Single-photon lifetimes reach $20.6 \\pm 0.4$ ms and $15.6 \\pm 0.2$ ms with pure dephasing beyond 40 ms, records for a multimode quantum memory. Using sideband transitions through the transmon's $|f\\rangle$ level, plus a feedforward protocol that corrects ancilla errors after each ladder step and a parity filter, the authors prepare Fock states up to $|20\\rangle$ with post-selected fidelities of $95.3 \\pm 1.9\\%$ and $94.6 \\pm 1.1\\%$. A virtual Raman beamsplitter then entangles the two modes with a coherence-limited fidelity of $99.810 \\pm 0.003\\%$ after erasure post-selection, approaching 99.9 percent when transmon heating is detected mid-circuit. The paper's point is that ultrahigh-coherence storage and high-fidelity programmable control can live in the same module, a step toward qudit-based quantum computing and modular quantum networks.","feed_headline":"20 ms photon lifetimes pair with 99.8% cavity gates","feed_subtitle":"A two-mode niobium cavity with a weak transmon link prepares Fock states up to |20> and entangles near the coherence limit.","key_machinery":"The load-bearing mechanism is the ancilla-mediated sideband interaction together with measurement-based error correction. Sideband pulses at $|f,n\\rangle \\leftrightarrow |g,n+1\\rangle$ climb the Fock ladder; after each pulse the transmon is measured in $|g\\rangle$, $|e\\rangle$, or $|f\\rangle$, and a feedforward pulse either repeats the step or repumps before repeating, correcting the dominant ancilla errors. The parity filter implements a parity-dependent phase gate: a $\\pi_{ge}/2$ pulse, a wait of $\\pi/|\\chi_e|$, and a reverse $\\pi_{ge}/2$ pulse map wrong-parity cavity states onto $|e\\rangle$ for post-selection. The virtual Raman beamsplitter uses two detuned sideband drives through the virtual $|f00\\rangle$ level; its coherence-limited fidelity is governed by $F^*_{BS} \\approx 1 - (\\pi/4)(\\kappa_{BS}/g_{BS})$ with $\\kappa_{BS} = \\kappa_1 + \\kappa_\\phi/2$, where $\\kappa_1$ and $\\kappa_\\phi$ are extracted by fitting the damped population oscillation to Eq. (1). The transmon anharmonicity $\\alpha$ supplies the $|\\alpha/\\Delta|$ enhancement that makes the Raman route faster than direct four-wave mixing.","core_discovery":"The central discovery is that the tension between long cavity coherence and fast control can be resolved by engineering the transmon participation down while keeping sideband coupling up. In a two-cell elliptical cavity, the two normal modes, called Alice and Bob, are each weakly dispersively coupled to one transmon, with dispersive shifts of $-71$ kHz and $-96$ kHz, yet sideband drives at the $|f,n\\rangle \\leftrightarrow |g,n+1\\rangle$ transition implement the control that SNAP gates would be too slow to provide. The authors verify the harmonic-oscillator scaling $\\gamma_n = n\\gamma_1$ up to $n = 20$, and use a three-state transmon readout with about 98 percent fidelity to feed forward corrections for transmon decay and dephasing after each sideband pulse, then a parity filter to discard wrong-parity shots. For two-mode entanglement, driving both sidebands with a common detuning $\\Delta$ from the virtual $|f00\\rangle$ level produces a beamsplitter interaction whose rate is enhanced by $|\\alpha/\\Delta|$ over four-wave mixing; fitting the damped Rabi oscillation to the model of Ref. [7] yields the coherence-limited fidelities. The paper also shows the beamsplitter rate is photon-number-dependent, which makes the gate a natural nonlinear primitive for qudit control, and reports numerical synthesis of a two-qutrit CSUM gate at 99.1 percent fidelity from five VRBS operations plus single-qutrit rotations.","pith_inferences":["Because the coherence-limited fidelities come from a fit that omits control errors, the realized gate fidelity is expected to be lower; direct randomized benchmarking of the VRBS gate would quantify the gap and likely guide pulse calibration.","The same feedforward logic should extend from Fock-basis ladder climbing to arbitrary cavity state synthesis: measuring the ancilla after each sideband operation and correcting in real time could protect superposition states too, not just number states.","If the two-cell module scales to longer multi-cell cavities, each additional cell could add a mode with similar coherence; a single cryogenic module could then host many addressable modes, making compact qudit simulations of lattice gauge theories and cascaded random-access memories more concrete.","The parity filter's post-selection discard rate grows with photon-number errors; converting it into genuine erasure conversion, repumping and reusing the state, would be a testable next step toward fault-tolerant dual-rail computation."],"forward_implications":["The two-mode module can act as an order-of-magnitude longer-lived multimode quantum memory than prior three-dimensional multimode devices, with both modes usable for storage and processing.","Fock states up to $|20\\rangle$ with over 95 percent post-selected fidelity, together with the measured $\\gamma_n = n\\gamma_1$ scaling, make 21-level qudit encodings practical in a single mode.","The virtual Raman beamsplitter realizes a dual-rail qubit gate near the coherence limit; post-selecting on erasure and on the transmon ground state improves swap fidelity by a factor of two over 100 swaps.","The photon-number-dependent beamsplitter is a nonlinear entangling primitive: numerical synthesis reaches 99.1 percent for the two-qutrit CSUM gate with five VRBS operations plus single-qutrit rotations.","Sideband-aided driven-dissipative reset empties a cavity in about 25 ms instead of the natural roughly 160 ms decay, enabling fast reinitialization between runs."],"supporting_citations":[{"why":"Sets the prior single-mode superconducting cavity coherence benchmark that this two-mode module extends.","marker":"[15]"},{"why":"Provides the surface-treatment and quality-factor methodology behind the millisecond-scale photon lifetimes.","marker":"[17]"},{"why":"Supplies the low-loss niobium coaxial cavity fabrication baseline for the high internal quality factors.","marker":"[16]"},{"why":"The fast sideband control scheme for weakly coupled multimode bosonic memories that this paper builds on.","marker":"[31]"},{"why":"Explains sideband-drive-induced Purcell decay scaling, used to choose the VRBS detuning.","marker":"[32]"},{"why":"Provides the damped-oscillation model of Eq. (1) and the parametric beamsplitter benchmarking approach.","marker":"[7]"},{"why":"Frames the single-photon two-mode subspace as a dual-rail qubit with erasure errors.","marker":"[50]"},{"why":"Motivates mid-circuit ancilla measurement for heating detection and dual-rail surface code implementations.","marker":"[49]"}],"fun_headline_variants":["20 ms cavity lifetimes with error-resilient control to 99.9%","Cavity qudits beyond memory: 20 ms coherence, 99.9% gates","Ultracoherent two-mode cavity: Fock |20> and 99.9% entanglement","Error-corrected sideband control for long-lived cavity qudits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 99.9 percent headline is not a measured process fidelity: it follows from fitting damped oscillations to a formula that assumes only exponential photon loss and dephasing are present, treating state preparation, the sideband pulses, and the final mapping and readout as perfect, and the paper's own open-system simulations omit control errors.","fun_headline_variants_meta":{"raw":{"variants":["20 ms cavity lifetimes with error-resilient control to 99.9%","Cavity qudits beyond memory: 20 ms coherence, 99.9% gates","Ultracoherent two-mode cavity: Fock |20> and 99.9% entanglement","Error-corrected sideband control for long-lived cavity qudits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000812,"raw_usage":{"total_tokens":3645,"prompt_tokens":1113,"completion_tokens":2532,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":729,"completion_tokens_details":{"reasoning_tokens":2448}},"tokens_in":729,"tokens_out":2532,"duration_ms":21389,"temperature":1.0,"reasoning_tokens":2448,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:06:43.768581+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run full quantum process tomography or interleaved randomized benchmarking on the virtual Raman beamsplitter in the dual-rail subspace $\\{|10\\rangle, |01\\rangle\\}$, without discarding erasure or heating shots, and compare the process fidelity to $F^*_{BS}$ from Eq. (2). If the tomographic fidelity falls short by more than the quoted 0.2 percent infidelity, the missing control errors are the dominant limitation and the 99.9 percent figure is an upper bound rather than a realized gate fidelity.","supporting_citations":[{"cited_title":"Reinhold, S","cited_arxiv_id":null,"evidence_quote":"Frames the single-photon two-mode subspace as a dual-rail qubit with erasure errors."}],"review_version":1}