{"id":"b63b0a08-1313-40dd-9158-89c8e6b80360","arxiv_id":"2607.15040","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An MLP reconstructs the plaquette phase of a three-level Δ system from eight simulated STIRAP transfer efficiencies.","lead":"This paper trains a neural network to read out a hidden gauge-invariant phase from the efficiency of STIRAP population transfer in a three-level Δ system. The demonstration uses only noiseless simulations, so the practical sensing claim still needs experimental confirmation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Exact φ→−φ degeneracy in the eight efficiency features makes the central sensing claim internally inconsistent: no MLP can reconstruct the sign of φ from populations.","rationale":"The reader correctly flagged missing invertibility and noise robustness. My stress test finds a sharper, prior problem: exact sign degeneracy from complex conjugation. This follows directly from Eq. (2) and the definition of the features as real populations; the detuning protocols in Table 2 contain no φ dependence, so the symmetry is not lifted. The paper's Sec. 4 claim of accurate reconstruction across [−π, π] is therefore internally inconsistent. One caveat: I have not run the code; if the simulation accidentally breaks the symmetry, the result could be an artifact. The proposed concrete test distinguishes these cases. Because the central 'quantum sensing' claim as stated cannot survive the symmetry, I move the verdict to REJECT. I credit the paper for a clear setup and for using a circular encoding, but the encoding cannot fix an unidentifiable sign.","tokens_in":9255,"tokens_out":7600,"duration_ms":90279,"concrete_test":"Insert a symmetry check in the data-generation pipeline: for a fixed φ0 ∈ (0, π), compute and compare the eight efficiencies x(+φ0) and x(−φ0) with identical pulses and detunings (Eqs. 9–11). If the L∞ difference is below numerical tolerance, the fingerprint is φ → −φ degenerate; then retrain/evaluate the MLP on a balanced test set containing both signs and record the sin-component error. Achieving the reported small error would require a code path that breaks H(−φ) = H(φ)*, e.g., non-Hermitian detuning, a sign error in the lower off-diagonal, or phase sampling restricted to [0, π].","verdict_should_be":"REJECT","load_bearing_attack":"For the Hamiltonian in Eq. (2) with real pulses and detunings, H(−φ) = H(φ)*. Since the STIRAP initial state is real and every feature is a final population η = |⟨1|ψ(tf)⟩|², complex conjugation implies η_k(φ) = η_k(−φ) for each of the eight configurations. Hence the 8-dimensional fingerprint satisfies x(φ) = x(−φ), while the circular target y(φ) = (cos φ, sin φ) is not invariant under φ → −φ (except φ = 0, ±π). A deterministic model trained on x cannot separate φ and −φ; the reported low test error over φ ∈ [−π, π] (Sec. 4, Fig. 4) is therefore not reproducible unless data generation or evaluation breaks this symmetry. This is not a noise-robustness issue: even in the ideal noiseless limit the mapping is two-to-one. The paper's Sec. 1 acknowledges the inverse problem is 'generally non-invertible' but never identifies this exact sign degeneracy, and no feature in Table 2 introduces a chiral observable such as a complex coherence or current.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a machine-learning approach to estimate the plaquette phase φ in a closed-loop three-level Δ system from STIRAP population transfer efficiencies. The Hamiltonian (2) contains a residual complex coupling e^{±iφ}; Sec. 2.3 shows that a non-zero φ breaks coherent population trapping, producing a phase-dependent transfer efficiency. Sec. 3 describes an eight-efficiency feature vector built from three driving conditions and several detuning protocols, and an MLP with circular output (cos φ, sin φ). Sec. 4 reports low test error on a synthetic dataset of 2000 phases. The authors conclude that the phase can be accurately reconstructed from experimentally accessible observables.","tokens_in":9564,"tokens_out":5698,"duration_ms":64150,"significance":"The analytical treatment of the dark-state breakdown is standard and the data-generation pipeline is transparent, including a held-out test set and early stopping; these are strengths. However, the central sensing claim is invalidated by an exact symmetry: all eight input efficiencies are invariant under φ→-φ while the target is not. Thus the reported MLP performance cannot be reproduced as stated, and the proposed method does not sense the signed plaquette phase. The practical significance of the work therefore hinges on a symmetry obstruction that is not addressed; the paper would need a fundamentally different observable or a reduced target (e.g., |φ|) to be salvageable.","major_comments":[{"comment":"For the Hamiltonian in Eq. (2), H(−φ)=H(φ)^*. Since the pulses in Eq. (10) are real, the detunings are real, and the STIRAP initial state is |0⟩, the propagator satisfies U(−φ)=U(φ)^*. Every feature is a final population η=|<1|ψ(t_f)>|² (Eq. 12), so η_k(−φ)=η_k(φ) for each of the eight configurations. The input vector therefore satisfies x(−φ)=x(φ) while the target (cos φ, sin φ) in Eq. (13) is not invariant under φ→−φ. A deterministic MLP cannot separate φ and −φ from identical inputs. The low test error in Fig. 4 is inconsistent with the stated data-generation protocol unless the symmetry is broken by an unstated step. This is a load-bearing flaw, not a noise issue: it persists in the ideal noiseless limit.","section":"Secs. 3.3 and 4, Eq. (2)"},{"comment":"The abstract claims estimation from 'experimentally accessible observables' and the conclusions frame the result as a quantum-sensing protocol. However, the entire study uses noiseless synthetic data with no detector noise, no parameter fluctuations, no error bars, and no comparison with a direct estimator or baseline. The only generalization demonstrated is interpolation within the same model distribution. This overstates the experimental scope and should be corrected even if the symmetry issue were resolved.","section":"Abstract; Sec. 5"}],"minor_comments":[{"comment":"'three-level ∆ system with.' contains a dangling 'with'.","section":"Sec. 5, first paragraph"},{"comment":"'STImulated' should be 'Stimulated'.","section":"Abstract"},{"comment":"The notation eδ_p is undefined; it presumably denotes the shifted detuning defined in Eq. (7).","section":"Eq. (8)"},{"comment":"The definitions of δ_pump and δ_Stokes are described only in prose; the table would be clearer if the matched detuning formulas were written explicitly.","section":"Sec. 3.3, Table 2"},{"comment":"The conversion from the two output components to a predicted angle is not specified; the convention used to report 'predicted phase' should be stated.","section":"Sec. 4 and Fig. 4"}],"recommendation":"reject","confidential_remarks":"The φ→−φ symmetry argument is decisive and I verified it directly from Eq. (2), the real initial state, and the definition of the features. Retraining or adding noise will not fix it; the protocol must be changed, for example by measuring a coherence phase or by estimating only |φ|. The paper's self-citations [23,24] are not a concern. The scope is appropriate for the journal, but the central claim is mathematically untenable as written."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read, before you spend referee time: this is a competent numerical proof-of-principle, not a field-changer. The genuinely new piece is using an MLP to estimate the gauge-invariant plaquette phase from eight simulated STIRAP transfer efficiencies. The phase-dependent STIRAP physics comes from Pope et al. [15], and the ML approach is a direct continuation of the group's own noise-classification work [23,24], but I don't know of an earlier demonstration of this specific estimation task. The circular (cos phi, sin phi) target encoding is the right call, the test set is genuinely held out, and the reported training/validation behavior is consistent with a stable fit.\n\nSoft spots, in order of importance. No noise, no error bars, no experimental data, no baseline comparison, no code or data release. The abstract's phrase 'experimentally accessible observables' is defensible in principle—final populations are measurable—but calling this 'quantum sensing' overstates what is shown: the mapping is learned and evaluated on the same noiseless model that defines the phase. Section 5 explicitly defers noise robustness. That is a scope issue, not a fatal flaw. The conditioning of the 8-dimensional fingerprint is also never characterized; they note the inverse problem is 'generally non-invertible' but do not say where the ambiguities are.\n\nOn the stress-test: the exact phi -> -phi degeneracy worry does not hold up. H(-phi)=H(phi)* is a time-reversal relation, and complex conjugation alone does not map the forward STIRAP evolution at phi to the forward evolution at -phi—the counterintuitive pulse sequence breaks time-reversal symmetry. The paper's own dressed-state leakage is proportional to sin phi, so the efficiencies are not obviously even in phi. If the eight features were exactly two-to-one, the reported test-set error over [-pi,pi] would be impossible; I see no hidden symmetry forcing that. A one-line symmetry comment in the paper would have preempted this, but the concern does not invalidate the result.\n\nWho should read it: anyone working on ML-assisted quantum control, STIRAP in closed-loop systems, or synthetic gauge-field characterization. It deserves a serious referee. I would send it out with the request that the authors add a noise model, a baseline comparison, and ideally a small experimental or analog demonstration before the word 'sensing' is used. As-is, it is a solid simulation-level extension with a clearly bounded claim.","headline":"A clean simulation-level proof of principle for estimating the plaquette phase from STIRAP efficiencies; the 'sensing' language runs ahead of the evidence, and the sign-degeneracy objection doesn't survive contact with the paper.","tokens_in":10012,"tokens_out":15846,"would_cite":true,"duration_ms":173573,"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":"A multilayer perceptron, trained on STIRAP transfer efficiencies measured under eight driving and detuning configurations, can reconstruct the gauge-invariant plaquette phase of a three-level Δ system across the full 2π interval.","keywords":["plaquette phase","three-level Δ system","STIRAP","coherent population trapping","machine learning","multilayer perceptron","quantum sensing","synthetic gauge fields"],"falsifier":"Sweep the plaquette phase on a controllable platform (or simulate the full Lindblad master equation with realistic noise) and record the eight efficiencies; if two distinct phases in [−π,π] yield fingerprint vectors that are statistically indistinguishable at the experimental noise level, or if the network's prediction error grows sharply once noise is added to the ideal efficiencies, then the central sensing claim fails. A numerical check: compute the Jacobian of the eight-efficiency map with respect to φ; if its singular values collapse toward zero anywhere, the map is locally non-invertible","tokens_in":9170,"feed_emoji":"🧠","tokens_out":4908,"duration_ms":53007,"temperature":0.7,"pith_summary":"The paper proposes a method to measure the gauge-invariant plaquette phase of a closed-loop three-level system—a quantity that is not directly observable—by turning the known phase-dependent degradation of STIRAP population transfer into a sensing signal. The central claim is that a multilayer perceptron, trained on simulated STIRAP efficiencies collected under eight different driving and detuning conditions, can accurately reconstruct the phase across the full 2π range. If true, this gives a cheap, measurement-tolerant way to sense an effective magnetic flux in engineered quantum systems, with applications to synthetic gauge fields and quantum metrology. The insight is that an imperfection—the breakdown of coherent population trapping—carries precisely the information one wants.","feed_headline":"Neural net reads hidden loop phase from STIRAP data","feed_subtitle":"The phase that breaks perfect state transfer leaves an eight-point fingerprint a network can decode.","key_machinery":"The gauge-invariant plaquette phase φ = φ0 + φs − φp, which remains after all local gauge transformations are exhausted; the generalized dark-state (trapped-state) condition that defines the detuning baseline; the dark state |D⟩ = cosθ|0⟩ − sinθ|1⟩; and the fact that for φ ≠ 0 the dressed states acquire a component along |D⟩ proportional to sin φ, which produces phase-dependent leakage. The sensing protocol uses eight STIRAP efficiencies (three driving configurations × detuning settings) as features, and a two-output MLP trained with mean-squared error on the circular encoding.","core_discovery":"The paper demonstrates that the plaquette phase φ, a gauge-invariant quantity that cannot be directly measured, is encoded in the transfer efficiency of STIRAP. When the detuning is set to the value that would ensure perfect trapping at φ=0, any nonzero φ causes leakage from the dark manifold proportional to sin φ, so the final population of the target state becomes a phase-dependent observable. The authors construct an eight-dimensional fingerprint of efficiencies from three driving conditions and three detuning schemes, and show that a multilayer perceptron trained on noiseless simulated data predicts φ across the full interval with small error. The MLP outputs (cos φ, sin φ), avoiding the","pith_inferences":["The paper only tests noiseless synthetic data; the immediate next experiment is to add measurement noise and parameter fluctuations. If the eight-efficiency mapping is not robust, the authors will need to augment the training set or the feature vector.","The choice of eight features is heuristic (three driving conditions, three detunings, two matched configurations). A cheaper protocol with fewer features may be possible; the paper lists reducing the number of input features as future work.","In a lattice of many triangular plaquettes, the same method could serve as a local magnetometer for synthetic gauge fields: measuring each plaquette's phase map would reconstruct the effective magnetic-field texture.","The network's generalization is only demonstrated within the same parameter range used for training (fixed pulse width, delay, and amplitude ratio). Whether the estimator is robust to changes in pulse shape or timing is an untested extension."],"forward_implications":["A single final-population measurement per configuration replaces full state tomography, making the phase estimate cheap and experimentally accessible.","The phase-dependent degradation of STIRAP, normally a drawback, becomes the physical resource that carries the sensing signal.","Because the method relies only on the closed-loop geometry, it transfers to any engineered platform—superconducting circuits, NV centers, quantum dots—with the same fingerprint structure.","The circular (cos φ, sin φ) output means the estimator works uniformly across the 2π boundary, so no phase discontinuity artifacts appear."],"fun_headline_variants":["Machine learning decodes hidden loop phase in STIRAP","Neural net reads plaquette phase from STIRAP data","Decoding the plaquette phase with a neural net","ML extracts gauge-invariant phase from STIRAP efficiencies"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The eight measured transfer efficiencies form a well-conditioned, effectively invertible fingerprint of the plaquette phase, so a network trained on ideal noiseless simulations will infer the phase from real measurements made with noise and experimental imperfections.","fun_headline_variants_meta":{"raw":{"variants":["Machine learning decodes hidden loop phase in STIRAP","Neural net reads plaquette phase from STIRAP data","Decoding the plaquette phase with a neural net","ML extracts gauge-invariant phase from STIRAP efficiencies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000575,"raw_usage":{"total_tokens":2506,"prompt_tokens":652,"completion_tokens":1854,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":396,"completion_tokens_details":{"reasoning_tokens":1797}},"tokens_in":396,"tokens_out":1854,"duration_ms":14305,"temperature":1.0,"reasoning_tokens":1797,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T00:19:08.976909+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sweep the plaquette phase on a controllable platform (or simulate the full Lindblad master equation with realistic noise) and record the eight efficiencies; if two distinct phases in [−π,π] yield fingerprint vectors that are statistically indistinguishable at the experimental noise level, or if the network's prediction error grows sharply once noise is added to the ideal efficiencies, then the central sensing claim fails. A numerical check: compute the Jacobian of the eight-efficiency map with respect to φ; if its singular values collapse toward zero anywhere, the map is locally non-invertible","supporting_citations":[],"review_version":1}