{"id":"83dda083-5165-427b-9a40-44eba7d23ff6","arxiv_id":"2607.07629","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":5,"one_line_summary":"Polarization drift on a 15-km hybrid aerial-inground fiber link shows strong diurnal and seasonal cycles correlated with temperature, humidity, and wind, and is partially estimable from weather data using random forests.","lead":"This paper measured how light polarization drifts in a 15-km hybrid aerial-and-underground fiber link over 11 months, finding strong day-night and seasonal cycles tied to temperature, humidity, and wind. A smart generalist might read it to understand how weather drives signal instability in quantum-communication fiber networks and whether environmental data can predict it.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The <2% train-test RMSLE gap is the headline ML claim, but the lagged configuration achieving it was selected on the test set, and the small gap is more consistent with underfitting than genuine generalization.","rationale":"The reader correctly identified test-set hyperparameter selection as a weakening factor in the rationale and noted the large absolute errors and underfitting. However, the reader's designated 'weakest_assumption' focused on the single-station weather proxy issue, which is a legitimate but less load-bearing concern — it questions whether the input features are meaningful proxies, but even with perfect weather data the test-set selection and underfitting issues would remain. The most load-bearing concern is that the headline <2% gap is an artifact of (a) selecting hyperparameters on the test set and (b) a model too simple to overfit, not evidence of genuine generalization. The paper is honest about these limitations (acknowledging underfitting, large absolute errors, and the baseline outperforming at night), which is why the CONDITIONAL verdict remains appropriate. The empirical correlation findings (diurnal/seasonal structure, Spearman correlations) are well-supported and do not depend on the ML estimation claims. The theoretical model is presented transparently as qualitative motivation rather than validated prediction. No code or data availability limits independent verification, but the paper's own acknowledgments of limitations are sufficiently detailed that a reader can assess the claims with appropriate caution. The verdict should remain CONDITIONAL: the observational findings are sound, but the ML generalization claim is overstated relative to what the methodology supports.","tokens_in":22906,"tokens_out":2839,"duration_ms":156678,"concrete_test":"Re-select the lag parameters (τ, dτ) using only training-set cross-validation (not test RMSLE), then evaluate the resulting model on the held-out test set. If the test RMSLE exceeds 0.905 (the threshold for a >2% gap relative to the training CV RMSLE of 0.887), the <2% generalization gap claim does not hold. Additionally, compare the full weather model against the hour-of-day-only baseline on the full test set (not split by day/night): if the baseline achieves comparable or lower test RMSLE, the weather features are not contributing to generalization beyond diurnal structure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central ML claim is a <2% relative gap between training CV RMSLE (0.887) and test RMSLE (0.888) for the spectral area estimator with lagged inputs F_X = G_{6,4}[X]. Two issues undermine this. First, the lag parameters (τ=24, dτ=4) were explicitly selected by minimizing test RMSLE: Section 3.2 states 'we determined that (τ, dτ) = (24,4) hours yielded a 5.4% relative reduction in the test RMSLE.' This is test-set hyperparameter selection, making the reported test RMSLE optimistically biased. The unlagged configuration (where no test-set selection occurred) has test RMSLE 0.939 vs. train CV 0.959 — a gap of -2.1% (test better than train), which already exceeds 2% in absolute terms. Second, both train and test RMSLE values are ~0.89, corresponding to multiplicative errors of e^0.89 ≈ 2.4 (i.e., ~140% relative error). The paper itself acknowledges 'a strong possibility of underfitting contributing to increased bias' (Section 4) and that the estimator captures diurnal cycles but 'not the amplitudes of the peaks and troughs' (Fig. 9). A small train-test gap in an underfit model is not evidence of generalization — a constant predictor would also show a small gap. The paper's own baseline comparison (Fig. 8) shows that an hour-of-day-only estimator outperforms the full weather model during nighttime for spring, summer, and fall, suggesting the weather features add noise rather than signal for a substantial portion of the data. The reader's weakest_assumption (single-station weather proxy) is a valid concern but is secondary to this methodological issue, which directly determines whether the headline metric is meaningful.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This manuscript presents an 11-month study of polarization drift on a 15-km hybrid aerial-inground fiber link at ORNL's QNET testbed. The authors compute sub-Hz FFT spectral features (spectral area, centroid, variance, entropy, and beta-exponent) from Stokes-parameter measurements and correlate them with environmental variables (temperature, humidity, wind speed, hour of day). A theoretical model based on SO(3) rotation products and stochastic birefringence is developed to motivate the analysis qualitatively. A random forest regressor is then trained on weather data to estimate FFT spectral features, with the headline result being a <2% relative gap between training CV RMSLE (0.887) and test RMSLE (0.888) for the spectral area estimator using lagged inputs. The work targets applications in quantum networking, specifically weather-aware link-state estimation for polarization stabilization.","tokens_in":23928,"tokens_out":1450,"duration_ms":227894,"significance":"The dataset—11 months of continuous polarization measurements on a hybrid aerial-inground link spanning multiple seasons—is a valuable contribution to a literature where most studies cover shorter periods or single fiber types. The correlation analysis (Spearman, mutual information) is well-executed and the seasonal/diurnal structure is clearly documented. The theoretical model (Section 2.4, Appendix B), while not directly solved, provides physically grounded motivation for the lagged-input approach via Duhamel's principle. The baseline comparisons (Fig. 8) are a commendable design choice. However, the central ML generalization claim requires qualification: the <2% train-test gap is reported on a configuration whose lag parameters were selected using test-set performance, and the absolute RMSLE values (~0.89, corresponding to ~140% multiplicative error) indicate substantial underfitting, which the authors themselves acknowledge. These issues affect how the ML results should be interpreted but do not invalidate the correlation findings or the dataset contribution.","major_comments":[{"comment":"§3.2: The lag parameters (τ=24, dτ=4) were selected by minimizing test RMSLE, as the text states: 'we determined that (τ, dτ) = (24,4) hours yielded a 5.4% relative reduction in the test RMSLE.' This constitutes test-set hyperparameter selection, making the reported test RMSLE of 0.888 optimistically biased. The <2% train-test gap—the headline ML claim—is thus not an unbiased measure of generalization. The authors should either (a) re-select lag parameters using only training/CV data and re-report test performance, or (b) explicitly acknowledge this as a limitation and reframe the <2% gap claim accordingly.","section":null},{"comment":"§3.2, Table 1, §4: Both train and test RMSLE for the spectral area are ~0.89, corresponding to multiplicative errors of e^0.89 ≈ 2.4 (~140% relative error). The paper acknowledges 'a strong possibility of underfitting contributing to increased bias' (§4) and that the estimator captures diurnal cycles but 'not the amplitudes of the peaks and troughs' (Fig. 9). A small train-test gap in an underfit model is weak evidence of generalization—a trivial constant predictor would also show a small gap. The authors should add a baseline comparison reporting the train-test gap for a naive estimator (e.g., hour-of-day mean) to contextualize whether the <2% gap is meaningful or merely a consequence of underfitting. Fig. 8 partially addresses this but does not report the train-test gap for the trivial baseline's gap.","section":null},{"comment":"§2.5, §3.2: The test dataset spans only September–November 2025 (one season, fall). The <2% gap therefore does not demonstrate cross-season generalization; it shows performance within a single held-out season. The paper notes that fall test RMSLE is higher than other seasons even when fall is included in training (§3.2), suggesting distributional shift. The claim in the abstract and §4 that estimators showed 'small overall gap (<2% relative error) between test and train datasets' should be qualified to note that the test set covers only fall, and cross-season generalization to unseen seasons is not demonstrated.","section":null}],"minor_comments":[{"comment":"§2.3: The spectral area is defined as the sum of the first 40 FFT bins, and the beta-fit uses bins 5–100. The rationale for these specific bin ranges is not provided. A brief justification (e.g., excluding white-noise floor, capturing dominant drift frequencies) would strengthen this choice.","section":null},{"comment":"§2.5: The random forest hyperparameters are reported as ranges (e.g., 'minimum leaf size of 10-15; maximum tree depth of 10-15'). The final selected values should be reported for reproducibility.","section":null},{"comment":"Fig. 8: The figure caption does not specify what 'CV' and 'test (fall)' labels refer to in the context of the seasonal folds. Clarifying whether 'winter,' 'spring,' 'summer' folds are from cross-validation on the training set or from the test set would help.","section":null},{"comment":"§3.1: The mutual information scores (0.57 in summer, 0.28 in winter) are described as 'less reliable' than Spearman's correlation, yet they are still used to draw conclusions about the predictive value of hour-of-day. A brief note on why they are included despite this caveat, or removal of the quantitative claims, would help.","section":null},{"comment":"Eq. (4): The symbol K_O (outermost layer) is introduced but the subscript notation is inconsistent with the later K_j notation in Eq. (5). Consistent notation would improve readability.","section":null},{"comment":"§2.4.3: The Duhamel's principle argument motivating lagged inputs is concise but could benefit from one sentence clarifying that the response kernel G_j(τ) is not estimated from data but motivates the use of lagged features in the ML model.","section":null},{"comment":"Fig. 9: The y-axis label and units for the spectral area are not clearly specified. Adding units or noting that the quantity is dimensionless (as stated in §2.3) would aid interpretation.","section":null},{"comment":"The abstract mentions 'spectral moments9' with a stray numeral, likely a formatting artifact.","section":null}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about test-set hyperparameter selection for the lag parameters is well-founded and is the primary reason for the major revision recommendation. The reader's weakest-assumption point about single-station weather data being a coarse proxy is valid but is already acknowledged by the authors in §4. The circularity concern does not land: the theoretical model is explicitly stated to be qualitative motivation only, and the ML estimator is trained and evaluated on empirical data with no definitional circularity. The paper's dataset and correlation analysis are solid contributions; the ML claims need reframing rather than removal."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive review. The referee correctly identifies that the lag parameter selection used test-set performance, that the absolute RMSLE values indicate underfitting, and that the test set covers only one season. We agree with all three major comments and will revise the manuscript accordingly. Specifically, we will (a) re-select lag parameters using only training/CV data and re-report, (b) add a naive baseline train-test gap comparison, and (c) qualify the generalization claim to note single-season test coverage. These are honest limitations that do not affect the correlation findings or the dataset contribution.","responses":[{"response":"The referee is correct. The text in §3.2 states that (τ, dτ) = (24, 4) was chosen based on test RMSLE, which is test-set hyperparameter selection and makes the reported <2% gap optimistically biased. We will adopt option (a): re-select lag parameters using only training/CV data (ten-fold shuffled cross-validation on the December 2024–August 2025 dataset) and then report the resulting test RMSLE on the held-out fall period. We will also add option (b) as an explicit limitation in §4, noting that even after CV-based selection, the small number of lag configurations searched means some residual optimistic bias may remain. The headline claim in the abstract and §4 will be reframed accordingly—we will describe the train-test gap as a consistency check rather than an unbiased generalization measure.","revision_made":"yes","referee_comment":"§3.2: The lag parameters (τ=24, dτ=4) were selected by minimizing test RMSLE, constituting test-set hyperparameter selection. The <2% train-test gap is optimistically biased. Authors should either (a) re-select using only training/CV data and re-report, or (b) acknowledge as limitation and reframe."},{"response":"We agree that a small train-test gap in an underfit model is not strong evidence of generalization on its own. Fig. 8 already compares RMSLE values for the random forest against two baselines (hour-of-day-only random forest and a product estimator), but it does not report the train-test gap for those baselines. We will add this: for each baseline, we will report both CV RMSLE and test RMSLE so the reader can see whether the trivial baseline also exhibits a small gap. Our expectation is that the hour-of-day baseline will show a comparable or smaller gap (since it has even less capacity to overfit), which would confirm the referee's point that the gap alone is not meaningful without considering absolute error and relative performance. We will revise §3.2 and §4 to explicitly state that the <2% gap should be interpreted alongside the baseline comparisons and absolute error levels, not as standalone evidence of generalization.","revision_made":"yes","referee_comment":"§3.2, Table 1, §4: Both train and test RMSLE ~0.89 (~140% multiplicative error), indicating underfitting. A small train-test gap in an underfit model is weak evidence—a trivial constant predictor would also show a small gap. Authors should add a baseline comparison reporting the train-test gap for a naive estimator (e.g., hour-of-day mean)."},{"response":"This is correct. The test set covers only fall 2025, so the <2% gap demonstrates within-season consistency, not cross-season generalization. The paper already notes that fall RMSLE is higher than other seasons even when fall is included in training (§3.2), which suggests distributional shift. We will revise the abstract, §3.2, and §4 to explicitly state that (i) the test set covers only the fall season, (ii) cross-season generalization to unseen seasons is not demonstrated, and (iii) the higher fall RMSLE relative to other seasonal folds suggests distributional shift that would need to be addressed in future work with multi-season test data. We will also note in §4.1 (Ongoing work) that additional data collection is needed to enable cross-season validation.","revision_made":"yes","referee_comment":"§2.5, §3.2: Test dataset spans only September–November 2025 (one season, fall). The <2% gap does not demonstrate cross-season generalization. The claim in the abstract and §4 should be qualified."}],"tokens_in":22891,"tokens_out":936,"duration_ms":81714,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"This paper's real contribution is empirical: 11 months of continuous polarization drift measurements on a 15-km hybrid aerial-inground fiber link at ORNL, with diurnal and seasonal patterns characterized against weather variables. That dataset and the correlation analysis built on it are legitimate new work. The Spearman correlation analysis is straightforward and well-executed, and the observation that wind speed dominates winter drift behavior (possibly via ice-coated fiber galloping) is physically sensible and worth reporting. The theoretical model in Section 2.4 and Appendix B is detailed and internally consistent, though it is never directly solved or fitted — it serves as qualitative motivation only, which the authors acknowledge honestly.","headline":"Solid empirical characterization of polarization drift on hybrid fiber; ML claims need scrutiny","tokens_in":23767,"tokens_out":708,"would_cite":false,"duration_ms":51441,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Weather predicts polarization drift on hybrid fiber links","keywords":["polarization drift","optical fiber","aerial fiber","quantum networking","random forest","SOP drift","birefringence","Stokes vector"],"falsifier":"If the spectral area of the polarization drift FFT showed no diurnal structure, no seasonal variation, and no improvement over a time-of-day-only baseline when weather variables are added, the central claim would be refuted.","tokens_in":22954,"feed_emoji":"🌡️","tokens_out":1723,"duration_ms":136166,"temperature":0.7,"pith_summary":"This paper argues that the sub-Hertz polarization drift of a 1550-nm optical signal traveling over a 15-km hybrid aerial-inground fiber connection is structured — not purely random — and can be partially estimated from routine weather measurements. Over 11 months of continuous Stokes-vector measurements, the authors compute FFT spectral features (spectral area, spectral centroid, spectral variance, spectral entropy, and a noise exponent β) and show that these features obey strong diurnal and seasonal cycles: drift peaks during daytime temperature maxima and wind-speed peaks, drops at night, and shows higher mean and variance in summer than winter. A physical model casts polarization drift as a product of SO(3) rotations along fiber segments, where temperature and humidity modulate the deterministic birefringence while mechanical stresses from wind and thermal gradients introduce multiplicative noise. This model predicts that the slow-axis birefringence suppresses stochastic drift, and that temperature changes — not absolute temperature — drive the drift amplitude, which the data confirm. The authors then train random forest regressors on weather variables (temperature, humidity, wind speed, hour of day, and their 45-minute changes) plus lagged histories to estimate the FFT features, achieving a test RMSLE of 0.888 for the spectral area with 24-hour lagged inputs, with less than 2% relative gap between training and test errors. The estimator captures diurnal cycles and seasonal trends but not individual peak amplitudes, consistent with the high multiplicative noise the physical model predicts.","feed_headline":"Weather data estimates fiber polarization drift within 2% error","feed_subtitle":"Eleven months of measurements on a hybrid aerial-buried fiber link show that temperature, humidity, and wind drive predictable diurnal and季节","key_machinery":"Theoretical model: polarization drift as a product of SO(3) rotation operators along discrete fiber segments, derived from a stochastic differential equation for the Stokes vector. The deterministic birefringence term depends on temperature and humidity through thermo-optic and photoelastic coefficients; the stochastic term captures mechanical perturbations (wind, bending, twisting) and produces multiplicative noise. The Magnus expansion and Baker-Campbell-Hausdorff formula are used to compose segment rotations into a total drift operator. The model motivates the use of lagged weather inputs via Duhamel's principle, which gives the fiber's internal temperature as a linear functional of past","core_discovery":"The central object is the spectral area of the polarization-drift FFT — the low-frequency power in the first 40 FFT bins of the normalized Stokes ratio — and the paper's core claim is that this quantity carries extractable environmental structure. The spectral area tracks diurnal and seasonal temperature cycles, with daytime maxima aligning with temperature peaks and nighttime minima during cooling. The noise spectrum follows an f^{-β} profile with β ≈ 1.0 ± 0.2, indicating flicker noise, and the signal energy concentrates within 5.1% of the 0.76 Hz FFT window. A random forest estimator trained on weather measurements and their lagged histories can estimate the spectral area with a test RMSL","pith_inferences":["The single-station weather proxy limitation could be directly tested by deploying distributed temperature sensors along the fiber route and comparing estimator performance; if local measurements substantially reduce RMSLE, the current model is learning coarse regional correlations rather than fiber-specific physics.","The model's Duhamel-principle justification for lagged inputs predicts a characteristic fiber thermal time constant; systematically varying the lag spacing (dτ) and history length (τ) in the grid search should reveal this time constant, which would be a fiber-specific physical parameter independent of the ML model.","The heteroskedasticity observed in spectral area — variance increasing with temperature — is a direct prediction of the model's birefringence-suppression mechanism, and could be tested by measuring whether fibers with deliberately engineered high birefringence show flatter variance-temperature curves in controlled incubator experiments.","The winter-specific wind-drift correlation (galloping of ice-coated fibers) suggests that seasonal fiber coating conditions (ice, dirt accumulation) change the effective thermal time constant, which would manifest as seasonally varying optimal lag parameters in the ML estimator."],"forward_implications":["If weather-driven polarization drift is partially predictable, quantum network operators could proactively schedule entanglement distribution during low-drift periods (nighttime, winter) and pre-tune polarization compensation systems based on weather forecasts rather than reactive feedback alone.","The finding that temperature changes (ΔT) correlate more strongly with drift than absolute temperature suggests that polarization stabilization systems should incorporate rate-of-change sensors or thermal-gradient measurements rather than point temperature readings.","The physical model's prediction that slow-axis birefringence suppresses stochastic drift implies that fibers engineered with higher intrinsic birefringence (e.g., polarization-maintaining fibers) would show reduced weather sensitivity in deployed links, quantifying the tradeoff between fiber cost and drift resilience.","Extension to higher-frequency FFT components (kHz, MHz) could reveal wind-gust-specific signatures distinct from the thermal-driven sub-Hz drift studied here, enabling frequency-segmented environmental attribution of polarization noise.","The <2% train-test gap suggests the random forest has learned generalizable environmental structure rather than overfitting, but the absolute RMSLE of 0.888 indicates substantial unexplained variance — likely from fiber-segment-specific microclimate effects and intrinsic stochastic perturbations that no single-station weather model can capture."],"fun_headline_variants":["Polarization drift spectra track diurnal weather cycles on aerial fiber","Flicker noise in fiber polarization drift follows temperature and wind patterns","Random forest estimates fiber polarization drift spectra from weather data","Spectral area of polarization drift reveals seasonal structure in aerial fiber","Wind speed and temperature peaks align with daytime polarization drift maxima"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The estimator assumes that weather readings from a single station at 15-meter altitude adequately represent the local environmental conditions experienced by each segment of a 15-km fiber that alternates between aerial and buried sections, traversing areas with potentially different microclimates. The paper itself acknowledges that local weather differences across the fiber route could be nontrivial and that climate patterns are not accounted for.","fun_headline_variants_meta":{"raw":{"variants":["Polarization drift spectra track diurnal weather cycles on aerial fiber","Flicker noise in fiber polarization drift follows temperature and wind patterns","Random forest estimates fiber polarization drift spectra from weather data","Spectral area of polarization drift reveals seasonal structure in aerial fiber","Wind speed and temperature peaks align with daytime polarization drift maxima"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":564,"prompt_tokens":495,"completion_tokens":69,"prompt_tokens_details":null},"tokens_in":495,"tokens_out":69,"duration_ms":26953,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T04:24:36.419719+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If the spectral area of the polarization drift FFT showed no diurnal structure, no seasonal variation, and no improvement over a time-of-day-only baseline when weather variables are added, the central claim would be refuted.","supporting_citations":[],"review_version":1}