{"id":"54d5a7c8-b398-4bdc-a135-6db557123ea9","arxiv_id":"2512.18037","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Subtractively-fabricated CMOS-compatible transmon qubits show T1 stability on par with lift-off qubits, with a universal σT1 ∝ ⟨T1⟩^{3/2} scaling and a few-percent junction aging over a year.","lead":"This paper measures how stable superconducting qubits made with a CMOS-compatible, subtractive fabrication process are over hours, days, and a year. It finds these qubits match the stability of conventional lift-off qubits, while also showing that longer-lived qubits fluctuate more—an important constraint for building large quantum processors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"T1-stability benchmark may be an artifact of the Rician fit: dropouts influence the fit despite integration claim, so the single-a collapse could reflect differential dropout censoring.","rationale":"The paper presents a genuinely valuable dataset and a transparent comparison, but the central quantitative benchmark — the collapse of all qubits onto σT1 = a⟨T1⟩^{3/2} with a single a — depends on how dropout events are handled. The authors state that integration of the fitted Rician distribution avoids outliers, but the fit is made to histograms containing those outliers, so the fitted parameters are not outlier-robust. This is an internal tension, not just a disagreement with community norms. If the test I propose shows that EMFT and lift-off points separate when analyzed with a more robust protocol, the headline 'on par' conclusion would be weakened. The long-term aging inference via Eq. 3 is also indirect, but it is secondary to the single-cooldown stability claim. I therefore agree with the reader's CONDITIONAL verdict: the paper's main practical conclusion is plausible but the universal-scaling evidence needs a robustness check before being accepted as definitive.","tokens_in":15456,"tokens_out":6286,"duration_ms":64750,"concrete_test":"Recompute all Fig. 5 points from the raw T1 time series using an explicit, pre-registered robust protocol: remove points below Median - k·MAD (k=3 and k=5) before estimating moments, and use nonparametric mean/σ (or e.g. median/1.4826·MAD). Refit Eq. 1 globally and separately for EMFT vs lift-off. If the two a values differ by more than the joint fit uncertainty, or if the collapse disappears, the universal scaling and on-par claim rest on the Rician/dropout treatment. Also repeat with a Gaussian fit to quantify fit-function sensitivity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the extraction of ⟨T1⟩ and σT1 in Fig. 5 (§III A, S.I. II.B). The authors fit a mirrored Rician (Eq. S9) to histograms that include dropout events, then integrate the fitted density to obtain mean and standard deviation, claiming this 'avoid[s] the influence of extreme and prolonged outliers.' But the fit itself is not robust: the long left tail produced by dropouts (e.g., B.1 in Fig. 4) pulls the fitted distribution toward lower T1 and increases its width. Integration of that fitted distribution therefore does not exclude dropouts; it propagates their influence into both moments. Because dropout frequency and severity vary across qubits and datasets (B.1 versus A.2), the resulting σT1 values are differentially biased. The claimed single proportionality factor a in Eq. 1 and the conclusion that subtractive devices are 'on par' with lift-off devices could then be an artifact of unequal dropout censoring rather than a universal TLS-limited scaling. The derivation in S.I. II.C.2 gives the 3/2 exponent but leaves a free; nothing in the model constrains a to be identical across fabrication methods unless the dropout handling is identical. A second, related weakness is that literature datasets differ in duration, binning, and availability of raw data, so the same Rician procedure may not be applied uniformly.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a temporal-stability study of eight transmon qubits fabricated with a subtractive, CMOS-compatible process, monitored continuously for ~95 h in a single cooldown (T1, T2*, Ramsey frequency, readout fidelity, Δm, Teff), plus a multi-cooldown study of two qubits over 10 thermal cycles spanning more than a year. The central claims are: (i) T1/T2* fluctuations are dominated by TLS-qubit interactions and follow the scaling σT1 = a⟨T1⟩^{3/2} (Eq. 1) with a single proportionality factor a for both the authors' subtractive devices and literature lift-off devices; (ii) subtractively fabricated devices are therefore stable 'on par' with lift-off devices; and (iii) long-term frequency shifts are dominated by an increase in the junction normal-state resistance RN, estimated indirectly at ΔRN/RN below 3.4% over 400 days. The analysis uses a modified (mirrored, offset) Rician fit to T1 histograms, with moments computed by integration of the fitted density; the fit is explicitly acknowledged as purely empirical.","tokens_in":15831,"tokens_out":9796,"duration_ms":89893,"significance":"If the reported claims hold, the practical significance is real: a wafer-scale-compatible subtractive process producing qubits whose short- and long-term stability matches lift-off devices is an important data point for scalable QPUs, and the year-long, ten-cooldown tracking of fq, fr, and T1 is a useful community dataset. The paper's strengths include the detailed and transparent experimental description, the continuous 95-hour multi-parameter monitoring of 8 qubits, the explicit acknowledgment that the Rician fit is empirical (Fig. 3 caption; S.I. II.B), the inclusion of literature-comparison datasets, and a supplementary section with complete pulse sequences and a derivation of the scaling law. The verification of the σT1 ∝ ⟨T1⟩^{3/2} law across fabrication methods is, however, a fit with a free prefactor rather than a parameter-free prediction; the statistical support for a universal prefactor is not yet demonstrated, and the long-term RN aging inference is indirect. The significance is therefore conditional on the robustness checks requested below.","major_comments":[{"comment":"The statement that computing moments by integration of the fitted Rician density 'avoid[s] the influence of extreme and prolonged outliers (dropouts)' is not supported. Eq. S9 is fitted to the full histogram, which contains the dropout tail; the fit parameters (ν, σ, Ti,max) are therefore pulled by the same outlier points the authors intend to exclude, and the integrated ⟨T1⟩ and σT1 inherit that pull. Because dropout frequency/severity varies strongly across qubits (B.1 vs A.2 in Fig. 4), the σT1 values entering Fig. 5 are differentially censored. Please quantify the sensitivity: e.g., recompute moments after explicit censoring of points below a defined threshold, compare with the fit-based moments, and report how the fitted a changes.","section":"S.I. II.B (Eq. S9); §III A 'T1 stability'"},{"comment":"The universal-scaling claim is under-supported statistically. The exponent 3/2 is fixed a priori and only the prefactor a is fitted; no goodness-of-fit statistic is reported, no free-exponent fit is shown, and no test is given of whether EMFT and literature qubits are consistent with a common a. This matters because the S.I. II.C.2 derivation (Eqs. S14–S15) leaves a dependent on the single-TLS decay-rate statistics, so a universal a is not a model prediction, and because Eq. S13 (delta-method) is questionable in the presence of the large fluctuations observed. Please report fit quality, residuals, a free-exponent fit, and an explicit common-a vs separate-a comparison; and soften 'confirm' (§IV) accordingly.","section":"§III A, Eq. (1), Fig. 5"},{"comment":"The comparability of the literature and EMFT datasets is not established. The inclusion criteria (>500 points, >10 h, single cooldown) do not control for measurement duration, sampling cadence, or histogram binning. Since dropout probability grows with observation time and the fit-based σT1 is sensitive to dropout contamination (previous comment), longer datasets may have systematically inflated σT1. The single-a collapse in Fig. 5 could then reflect disparate censoring rather than a universal TLS limit. Please demonstrate robustness to duration/cadence (e.g., restrict all datasets to a common 10-h window) or include these as covariates.","section":"§III A 'T1 stability'; dataset inclusion criteria"},{"comment":"The long-term aging result ΔRN/RN < 3.4% is an indirect estimate: RN is never measured, but reconstructed from fq shifts via Eq. (3), with Δ taken from Tc and EC from the design anharmonicity. The comparison with lift-off aging rates therefore inherits these model assumptions, and the 25%/75% split in §III B between RN-mediated and bare-resonator contributions to Δfr is stated without uncertainty. The conclusion that subtractive junctions age more slowly should be supported by direct RN measurements (or a sensitivity analysis) or presented as a preliminary inference. The authors partially acknowledge this in §III B, which mitigates the concern.","section":"§III B; Eq. (3)"}],"minor_comments":[{"comment":"The units of a are given as s^{2/3}, but Eq. (1) requires [a] = s^{-1/2}. Please correct the units.","section":"Fig. 5 caption"},{"comment":"The abstract reports an average total downward shift of approximately 61 MHz, while §III B states 50–70 MHz; please reconcile the numbers.","section":"Abstract vs §III B"},{"comment":"The phrase 'confirm that the TLS-qubit interactions remain the limiting factor' is stronger than the evidence supports, given the empirical fit and free prefactor; recommend 'are consistent with'.","section":"§IV"},{"comment":"Typos: 'form Zurich Instruments' should read 'from Zurich Instruments'; 'which occurs When certain numbers' has an erroneous capital W.","section":"Acknowledgments; S.I. II.C.1"},{"comment":"The mixed logarithmic/linear axis (log scale with a linear interval ±0.1 around zero) is hard to read; please clarify the scale convention in the caption.","section":"Fig. 4(d)"},{"comment":"Quantitative claims citing an annual report and a PhD thesis should indicate where in those documents the data appear, or prefer peer-reviewed sources.","section":"References [11], [44]"}],"recommendation":"major_revision","confidential_remarks":"The reader's conditional verdict is, in my view, the right one. The main risk is the statistical support for the scaling law: the paper's flagship claim (Eq. 1 with a single a) is currently a fixed-exponent fit without diagnostics, and the Rician-integration argument does not actually exclude dropouts. Both are fixable in revision. I would also encourage the editor to ask for the literature datasets' processing details (or shared analysis scripts), since cross-dataset comparability is load-bearing for the 'on par' conclusion. The manuscript is within scope for a quantum-hardware journal; novelty relative to previous stability studies is incremental, but the CMOS-compatibility angle and the one-year multi-cooldown data justify publication if the robustness checks pass."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the new thing here is real: a systematic stability benchmark of subtractively fabricated CMOS-compatible transmons against lift-off devices, with 95 hours of single-cooldown data on 8 qubits and a year-long 10-cooldown aging study. If you work on scalable superconducting fabrication, this is one of the first papers to give hard numbers.\n\nThe experiment is careful and the writing is honest — they explicitly flag the Rician fit as empirical, and they restrict the literature comparison to datasets with raw data or distributions and clear inclusion criteria. That is more than most benchmarking papers do.\n\nNow the soft spots. The load-bearing claim is the sigma_T1 = a <T1>^{3/2} collapse with one a for all fabrication methods. That collapse is only as good as the sigma_T1 estimates, and those estimates come from fitting a mirrored Rician to histograms that include dropout events and then integrating the fitted density. The problem is that the fit itself is vulnerable to the long left tail from dropouts — so integration of the fitted curve does not really avoid the influence of extreme and prolonged outliers as the paper claims. B.1's dropout is excluded in the T2* analysis, but it is not clear that a uniform, pre-defined censoring rule is applied to all qubits and all literature datasets. Differential dropout handling across datasets could easily generate a spurious single-a collapse. This is the main thing a referee should push on: show the analysis with and without dropout points, or report robust estimators like censored moments.\n\nThe other weakness is the long-term aging inference: R_N drift is back-computed from f_q shifts using Eq. 3, not measured. That is acceptable as an estimate, but the quantitative 'below 3.4%' claim is weaker than the raw frequency drift, which is direct. The paper does acknowledge some of this.\n\nNone of this sinks the central conclusion. The raw time traces and readout stability show subtractive devices behaving like lift-off ones; the 'on par' conclusion survives even if the universal scaling law is shakier. The scaling law is the cherry on top, not the cake.\n\nVerdict: worth a serious referee. I would send it to review with a request for sensitivity analysis on the outlier handling and a clearer statement of what was and was not excluded. I would also want the raw T1 sequences for their own qubits available. This is a useful, honest experimental contribution.","headline":"Useful benchmark data for CMOS-compatible qubits; 'on par' is credible, but the universal scaling law depends on fragile outlier handling.","tokens_in":16337,"tokens_out":5030,"would_cite":true,"duration_ms":51388,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Subtractively fabricated CMOS-compatible transmon qubits show temporal stability on par with lift-off devices, with all T1 fluctuations following a single scaling law.","keywords":["superconducting qubits","transmon","CMOS-compatible fabrication","subtractive fabrication","temporal stability","T1 fluctuations","two-level systems","qubit aging"],"falsifier":"Recompute σT1 and ⟨T1⟩ from the full time traces without excluding any dropout events, using a physically motivated model for the skewed distribution; if one constant a no longer fits both subtractive and lift-off devices, the universality claim fails.","tokens_in":15355,"feed_emoji":"⚛️","tokens_out":5131,"duration_ms":53962,"temperature":0.7,"pith_summary":"Superconducting qubits must keep their coherence times and control fidelities stable over hours and across many thermal cycles before large error-corrected processors become practical. This paper argues that transmon qubits made with a subtractive, CMOS-compatible fabrication process—where metal is etched away rather than lifted off—are just as temporally stable as qubits made with conventional lift-off methods. Over 95 hours in a single cooldown, eight qubits showed T1 and T2* fluctuations driven mainly by two-level-system defects, and the fluctuations follow the predicted scaling σT1 = a⟨T1⟩^{3/2} with one common factor a across all fabrication methods. Over ten cooldowns spanning more than a year, two tracked qubits kept a stable T1 baseline while their frequencies drifted downward by about 61 MHz, which the paper attributes to slow aging of the Josephson junction barrier. The upshot: the industrial-grade fabrication route does not appear to cost qubit stability, a key requirement for scaling up.","feed_headline":"Subtractive qubits match lift-off stability over a year","feed_subtitle":"Eight qubits tracked over 95 hours and two over a year stay coherent; fluctuations follow a single scaling law.","key_machinery":"The central object is the transmon qubit and its two-level-system (TLS) environment. The load-bearing identity is the scaling relation σT1 = a⟨T1⟩^{3/2}, derived from an ensemble of independent TLS defects each contributing to the decay rate; variance adds in quadrature, giving standard deviation proportional to ⟨T1⟩^{3/2}. To extract σT1 and ⟨T1⟩ from skewed histograms, the paper uses an empirical mirrored-Rician fit; for long-term aging it uses the relation fq ≈ (1/h)√(8EJEC) − EC/h and EJ ∝ 1/RN to infer junction resistance drift from frequency shifts.","core_discovery":"According to the paper, subtractively fabricated transmon qubits—made by etching rather than lift-off—show the same degree of short-term T1 stability as lift-off qubits: over 95 hours in one cooldown, T1 and T2* fluctuate because two-level-system defects cross the qubit frequency, but the fluctuations follow the same σT1 = a⟨T1⟩^{3/2} curve with a single constant a for all three fabrication approaches surveyed. Long-term, two qubits tracked over 10 cooldowns and more than a year keep a stable T1 baseline, while qubit frequencies drift downward by about 61 MHz on average, attributed to a slow increase in junction resistance RN; readout resonators shift much less.","pith_inferences":["A direct test of the aging story would be to measure room-temperature junction resistance alongside each cooldown; the paper infers RN from frequency shifts, and such a measurement would settle whether the barrier-thickening model is correct.","If the universal scaling law holds generally, it implies that fabrication-method comparisons of T1 stability should be normalized by ⟨T1⟩, not compared raw.","The same TLS-ensemble variance argument predicts similar scaling for pure dephasing or correlated T2* fluctuations; extending the analysis to T2* would test the model further.","The exclusion of dropout events is a modeling choice; a mechanistic model that includes strong TLS interactions might explain the drops and remove the need for empirical Rician fits."],"forward_implications":["Qubit arrays made by subtractive CMOS fabrication can be used in error-correction contexts without an extra stability penalty relative to lift-off devices.","Because σT1 grows as ⟨T1⟩^{3/2}, pushing qubit lifetimes higher makes T1 fluctuations larger; a QPU must budget for this.","Over many cooldowns, qubit frequencies drift downward by tens of MHz as junction resistance creeps up, so long-lived systems need recalibration or drift compensation.","Readout resonators move much less than qubit frequencies, so the dominant long-term correction is on the qubit drive frequency.","The observation that a single proportionality factor fits all fabrication methods points to TLS defects, not fabrication details, as the limiting source of short-term instability."],"fun_headline_variants":["Etched qubits match lift-off stability in year-long test","Subtractive qubits: stable T1, but frequencies drift 61 MHz","TLS-driven T1 wobble follows universal curve in etched qubits","95-hour and year-long checks: subtractive qubits stay coherent"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The 'on par with lift-off' and single-a conclusions depend on excluding dropout events and fitting T1 histograms with an empirical skewed distribution; if those choices are applied differently across datasets, the fitted proportionality factor could be an artifact of selection rather than a universal property.","fun_headline_variants_meta":{"raw":{"variants":["Etched qubits match lift-off stability in year-long test","Subtractive qubits: stable T1, but frequencies drift 61 MHz","TLS-driven T1 wobble follows universal curve in etched qubits","95-hour and year-long checks: subtractive qubits stay coherent"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000236,"raw_usage":{"total_tokens":1345,"prompt_tokens":752,"completion_tokens":593,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":516}},"tokens_in":496,"tokens_out":593,"duration_ms":6405,"temperature":1.0,"reasoning_tokens":516,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T15:05:07.071480+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute σT1 and ⟨T1⟩ from the full time traces without excluding any dropout events, using a physically motivated model for the skewed distribution; if one constant a no longer fits both subtractive and lift-off devices, the universality claim fails.","supporting_citations":[],"review_version":1}