{"id":"7c38d5be-a03b-4a7c-a768-1b68b073be9a","arxiv_id":"2601.02177","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"On commodity ESP32 WiFi sensors, six blind-source-separation methods all fail to achieve reliable multi-person gait identification (39–56% accuracy), suggesting a sensing-quality limit rather than an algorithmic fix.","lead":"This paper tested six signal-separation algorithms on low-cost ESP32 WiFi sensors for multi-person gait identification and found all of them peaked near 56% accuracy with very high feature overlap. The authors conclude that commodity WiFi hardware, not algorithm choice, is the bottleneck for dense multi-person identification, which matters for anyone building contactless authentication or privacy-sensitive monitoring on cheap devices.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported accuracy may be an artifact of unaligned BSS component order; no source-to-person matching is described.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the absence of any source-to-person alignment procedure for BSS outputs. This is indeed the most critical threat to the central claim. The paper's contributions and conclusions depend entirely on the accuracy numbers in Table I and the diagnostic metrics derived from labeled features. If component order is arbitrary and unaligned across trials, the SVM labels are meaningless, and the reported metrics cannot support the hardware-ceiling conclusion. The proposed concrete test—adding a permutation-correction/oracle alignment and comparing results—would directly settle whether the reported numbers are artifacts. I also note the paper's internal contradiction: Section III-D reports negative PDR (performance improves from 2-person to 10-person), while the abstract and conclusion claim 'severe performance degradation as person count increases.' This is a serious inconsistency, but it is secondary because even if resolved, the alignment issue would still undermine the empirical foundation. The reader's CONDITIONAL verdict is appropriate; no change in verdict is needed, but the conditions should explicitly include demonstrating the alignment procedure or providing the raw data to test it.","tokens_in":7852,"tokens_out":4447,"duration_ms":46581,"concrete_test":"Obtain the raw CSI recordings or, failing that, re-run the pipeline with an oracle permutation-alignment step: for each multi-person trial, use the known ground-truth person identities to compute the best one-to-one assignment between the p separated components and persons (e.g., Hungarian algorithm maximizing feature correlation with single-person enrollment references), then train/test the SVM on the aligned features. Compare accuracy and ISV/ISD with the paper's Table I. If aligned accuracy is materially higher (or method ranking changes), the reported results are dominated by source-ordering artifacts and the hardware-ceiling claim is unsupported; if aligned accuracy is essentially unchanged, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that ESP32 sensing quality, not separation algorithm, caps multi-person identification—rests on the SVM accuracies in Table I. Those accuracies are only meaningful if each separated component is labeled with the correct person identity for every trial. All six BSS methods used (FastICA, SOBI, PCA, NMF, Tensor; Wavelet thresholding is also order-sensitive) have permutation indeterminacy: the output components have no canonical ordering. The paper describes no source-to-person alignment, no permutation-correction, no Hungarian assignment, and no mention of how y_i in Eq. (9) is assigned for multi-person trials. If component order varies across trials, the training labels are effectively random with respect to identity; the SVM cannot learn a stable mapping, and the reported 39–56% accuracy, ISV, and ISD would measure a labeling artifact, not a hardware ceiling. This is the weakest load-bearing assumption because it jeopardizes the empirical foundation of the conclusion. The PDR contradiction (negative values meaning accuracy improves with more people) is secondary; it could be explained away, but arbitrary ordering cannot.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a systematic empirical study of multi-person gait identification using commodity ESP32 WiFi CSI. Six blind source separation methods (FastICA, SOBI, PCA, NMF, Wavelet, Tucker) are compared across seven scenarios with 1–10 people. The authors introduce three diagnostic metrics (ISV, ISD, PDR) and report accuracies between roughly 39% and 56%, with high feature-space overlap and inconsistent environmental effects. The main claim is that, under commodity ESP32 constraints, the bottleneck is hardware sensing quality and spatial diversity rather than the choice of separation algorithm, and that commodity ESP32 sensors cannot provide sufficient signal quality for reliable multi-person separation.","tokens_in":8129,"tokens_out":5323,"duration_ms":58344,"significance":"If the empirical result holds, this is a valuable negative result: it would redirect effort away from incremental algorithm refinement on ESP32-class devices toward hardware upgrades (massive MIMO, mmWave) or alternative sensing modalities. The systematic comparison of six separation methods and the attempt to diagnose failure modes beyond accuracy are useful in principle. However, the central conclusion rests on a multi-person classification pipeline that is not adequately specified, and the paper contains direct internal contradictions in the reported accuracy range, statistical claims, and performance-degradation narrative. The diagnostic metrics are a potentially useful contribution, but their definitions and units need to be made coherent before they can support the stated hardware-ceiling conclusion.","major_comments":[{"comment":"All six separation methods have permutation indeterminacy: FastICA, SOBI, PCA, NMF, wavelet thresholding, and Tucker decomposition output components in no canonical order. Section II-F trains an SVM with labels y_i in Eq. (9) as person identities, which requires each separated component to correspond to a fixed person consistently across trials. No source-to-person alignment, permutation-correction procedure, Hungarian assignment, or template matching is described. If component order varies across trials, the labels become effectively random with respect to identity, and the reported 39–56% accuracies—and the ISV/ISD metrics—measure a labeling artifact rather than a hardware sensing ceiling. This is load-bearing: the paper's conclusion that commodity ESP32 sensors cannot support reliable multi-person separation rests on these numbers. Please add an explicit alignment step and report its","section":"II-D, II-F, III (Table I)"},{"comment":"The abstract and contributions state that all methods achieve 45–56% accuracy with σ=3.74% and statistically insignificant differences (p>0.05). Table I lists PCA at 39.4% and Wavelet at 42.3%, both below 45%, and the mean of the six reported accuracies is 47.05% with sample σ≈5.3%, not 3.74%. Furthermore, no test statistic or p-value is reported in Section III. Please correct the accuracy range and standard deviation, and provide the actual statistical test (name, statistic, p-value) used to support the 'no significant differences' claim.","section":"Abstract, Section I, Table I"},{"comment":"The abstract and conclusion claim 'severe performance degradation as person count increases,' but Section III-D reports PDR values of -51.5% to -6.7% for most methods, meaning accuracy improves from the 2-person to 10-person scenarios, and the text explicitly calls this 'counterintuitive.' These two statements cannot both be true for the same dataset. Please report accuracy broken down by scenario (A–G) and reconcile the PDR results with the degradation narrative. As written, the inconsistency undermines the core 'hardware ceiling' interpretation.","section":"Abstract vs. Section III-D"},{"comment":"The ISV/ISD ratio is used as the central evidence that within-class variance exceeds between-class separation by factors of 73–1,266,000×, and the conclusion proposes viability thresholds (ISV/ISD < 10, |PDR| < 20%). However, ISV is defined as a squared distance and reported in units of ×10³ in Table I, while ISD is reported as a percentage. The claimed ratio of 1,266,000 for Tensor does not follow from the stated definitions unless the units are handled consistently. Without compatible units or clear normalization, the ISV/ISD ratio and the proposed thresholds are not interpretable. Please define ISV and ISD with consistent units and recompute the ratios.","section":"III-D, IV"}],"minor_comments":[{"comment":"Equation (1) uses a fixed eigenvalue threshold of 0.95 for source-count estimation. No validation is reported showing that this threshold yields the correct number of people across the seven scenarios. Please report estimated vs. true person counts.","section":"II-C"},{"comment":"The wavelet 'separation' is described as thresholding detail coefficients in gait frequency bands; this is a denoising/feature-selection step rather than a blind source separation method in the same sense as ICA, SOBI, PCA, or NMF. This weakens the claim that six diverse separation methods were compared.","section":"II-D.5"},{"comment":"No train/test split or cross-validation procedure is described. It is unclear whether the SVM classifiers are evaluated on the same participants seen in training and how many trials are used per class. Please specify the evaluation protocol.","section":"II-F / III-A"},{"comment":"The notation for ISV is inconsistent: Table I labels the column '(×10³)' and lists values such as 30.6, 2.3, and 49370.0, while the text says ISV ranges from 2.3×10³ to 49,370×10³. The corresponding ISD values are labeled '%' but are later used as raw distances in ratio calculations. Please clean up these units and notation.","section":"Table I and III-D"}],"recommendation":"major_revision","confidential_remarks":"The permutation-indeterminacy issue is the most serious concern; it goes to the validity of every multi-person accuracy and the derived diagnostic metrics. If the authors can add a credible source-to-person alignment and re-report the results, the paper may become publishable. The internal inconsistencies (39–56% vs. 45–56%, negative PDR vs. 'severe degradation') also need careful correction. I do not recommend reject because the fix is within the scope of a revision, but the revision must be substantial."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this paper gives you a genuinely useful comparative benchmark—six BSS methods on real ESP32 hardware, 440 trials, across 1–10 people—but the headline conclusion about a fundamental hardware ceiling rests on an unstated assumption about component ordering that the paper never justifies, and the abstract contradicts the results table on basic numbers. Worth a careful look, not as published.\n\nThe good: this is the kind of systematic negative result the subfield needs. The flat accuracy range (roughly 39–56%, depending on which part of the paper you trust) across algorithms is consistent with the weaker claim that algorithmic refinement alone won't crack multi-person identification on this hardware. The ISV/ISD/PDR diagnostics are simple summary statistics, but they do add a useful way to talk about why accuracy fails. The authors designed realistic scenarios across lab and classroom and report per-environment splits, which is more than most papers do.\n\nThe soft spots are serious. First, the permutation problem. All six BSS methods are order-indeterminate, yet nowhere is there a description of how each separated source got assigned to a person identity for the SVM labels. Without an alignment step like Hungarian matching or a fixed ordering heuristic, the reported accuracies, ISV, and ISD could be measuring a labeling artifact. This is load-bearing: the central conclusion depends on those numbers. Second, the internal inconsistency: the abstract says 45–56% accuracy, while Table I lists 39.4% for PCA and 42.3% for Wavelet. And the abstract claims \"severe performance degradation as person count increases\" while Section III-D reports negative PDR for most methods—meaning accuracy improved from 2 to 10 person scenarios. You can't have both. Third, no baselines, no error bars, no confidence intervals, no code or data. The diagnostic thresholds (ISV/ISD < 10, |PDR| < 20%) are asserted without support.\n\nThe core question—whether the bottleneck is sensing or algorithms—is important, and this paper is a reasonable first attempt. But as it stands, the evidence doesn't support the strong hardware-ceiling claim. A referee should ask for the alignment procedure, corrected numbers, and released data. If those come through, the benchmark could become a useful reference. I wouldn't cite it in its current form; I'd bring it to a reading group to discuss the BSS labeling failure mode.\n\nRecommendation: send to peer review, with the expectation of major revision.","headline":"Useful comparative benchmark on real ESP32 hardware, but the 'hardware ceiling' conclusion is undercut by an unaddressed BSS source-ordering problem and internal contradictions between the abstract and the results table.","tokens_in":8585,"tokens_out":2536,"would_cite":false,"duration_ms":28172,"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":"The paper argues that multi-person gait identification on commodity ESP32 WiFi is capped by sensing quality and spatial diversity rather than by the separation algorithm, with all six tested methods landing between 39% and 56% accuracy.","keywords":["WiFi sensing","Channel State Information","CSI","ESP32","gait identification","blind source separation","multi-person identification","biometric limits"],"falsifier":"A concrete test: rerun the identical pipeline but align separated components to person identities using a deterministic assignment rule (e.g., match each component to the closest single-person reference template). If accuracy on 5- and 10-person scenarios rises substantially above 56%, or if ISV drops by orders of magnitude, then the reported ceiling is at least partly a labeling artifact rather than purely a hardware limit. Another falsifier: collect CSI from an ESP32 with a 4-antenna external array (or 160 MHz bandwidth) and show ISV/ISD drops below the proposed viability threshold while acc","tokens_in":7723,"feed_emoji":"📶","tokens_out":5775,"duration_ms":55608,"temperature":0.7,"pith_summary":"The paper sets out to answer whether poor multi-person gait identification over WiFi is an algorithmic problem or a hardware one. Using commodity ESP32 sensors and six blind source separation methods across 1-10 people and two environments, it finds all methods land within 39-56% accuracy with statistically indistinguishable differences. It introduces three diagnostic metrics—intra-subject variability, inter-subject distinguishability, and performance degradation rate—that attribute the failure to extreme feature-space overlap, unstable within-person representations, and unpredictable environmental sensitivity. The authors conclude that commodity ESP32 sensors hit a fundamental sensing ceiling, so further algorithmic refinement on such hardware is unlikely to yield reliable multi-person identification. If correct, the result redirects research toward more capable hardware and places bounds on the passive identification threat from low-cost WiFi devices.","feed_headline":"Cheap WiFi sensors cap multi-person gait ID at 56% accuracy","feed_subtitle":"Even the best of six separation algorithms leaves 44% of walkers misidentified, pointing to a hardware ceiling.","key_machinery":"The load-bearing machinery is the ESP32 CSI acquisition front end together with three diagnostic metrics. The ESP32 provides k=3 antennas and m=52 OFDM subcarriers, which the paper shows yield insufficient spatial resolution (angular resolution roughly 67 degrees) for separating co-located walkers. The metrics—intra-subject variability (ISV), inter-subject distinguishability (ISD), and performance degradation rate (PDR)—quantify how far within-person feature scatter exceeds between-person separation, and how accuracy changes as crowd size grows. The six blind source separation methods (FastICA, SOBI, PCA, NMF, Wavelet, Tucker tensor decomposition) serve as probes: because their accuracies cl","core_discovery":"The paper claims and presents evidence for a negative result with a causal attribution: under the constraints of commodity ESP32 Channel State Information (three antennas, 52 OFDM subcarriers), multi-person gait identification is limited more by sensing quality and spatial diversity than by the chosen separation algorithm. Across 440 trials and seven scenarios spanning 1-10 people, FastICA, SOBI, PCA, NMF, Wavelet, and Tensor decomposition all produce similar accuracy (39-56%), with feature-space overlap above 97% and within-subject variability exceeding between-subject separation by factors of 73 to 1,266,000. The paper argues the ESP32's coarse angular resolution and correlated subcarriers","pith_inferences":["The paper does not describe a permutation-correction step for the six BSS methods; if separated component order is not aligned across trials, the reported accuracies and ISV/ISD values could partly reflect a labeling artifact rather than a true sensing ceiling. A re-run with source-to-person alignment (e.g., using per-person reference templates) would test whether NMF's 56% is a genuine ceiling or","One direct testable extension: apply the same pipeline to an 8-antenna array or a 160 MHz recording while holding the algorithms fixed; the ISV/ISD gap should shrink if the paper's hardware-attribution is correct, and should remain large if the bottleneck is algorithmic.","The negative PDR values (accuracy rising from 2 to 10 people) hint that separation benefits from richer mixtures; this is consistent with the paper's spatial-diversity argument but also suggests that a 'more people, more sources' regime may be where commodity sensors are least hopeless—a point worth probing with intermediate person counts.","If the hardware-ceiling conclusion holds, it also bounds passive privacy risk from commodity WiFi sniffing: an attacker with the same class of hardware cannot reliably identify individuals in a crowd, so security mitigations can focus on higher-end sensing or multi-modal fusion."],"forward_implications":["Algorithm choice will not rescue multi-person gait identification on commodity ESP32-class hardware; all six methods tested fall in the same narrow accuracy band with p > 0.05.","Commodity WiFi CSI is not a dependable multi-user biometric for authentication or continuous verification; deployments should treat it as a weak single-user channel at best.","Research effort on multi-person WiFi identification should shift from separation algorithms to sensing hardware with more antennas and bandwidth, such as massive MIMO or mmWave.","The proposed diagnostic thresholds (ISV/ISD < 10, |PDR| < 20%) give later systems a concrete pre-classification check for whether separation is even feasible before building a classifier.","Environmental unpredictability (2.2% to 18.9% lab-classroom gaps, including negative PDR) means models trained in one setting cannot be expected to transfer to another."],"fun_headline_variants":["WiFi gait ID hits wall: 56% max accuracy on ESP32","Commodity WiFi sensors cap multi-person gait ID at 56%","Six algorithms, one ceiling: WiFi gait ID fails beyond 56%","Hardware, not algorithms, limits multi-person WiFi gait ID"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper assumes that after blind source separation, each separated output component corresponds to a fixed, identifiable person in a consistent order across trials, so that labels can be attached to feature vectors for SVM training; none of the six BSS methods produces a canonical ordering, and no permutation-correction step is described.","fun_headline_variants_meta":{"raw":{"variants":["WiFi gait ID hits wall: 56% max accuracy on ESP32","Commodity WiFi sensors cap multi-person gait ID at 56%","Six algorithms, one ceiling: WiFi gait ID fails beyond 56%","Hardware, not algorithms, limits multi-person WiFi gait ID"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000541,"raw_usage":{"total_tokens":2478,"prompt_tokens":838,"completion_tokens":1640,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":1562}},"tokens_in":582,"tokens_out":1640,"duration_ms":11982,"temperature":1.0,"reasoning_tokens":1562,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T12:35:50.832621+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: rerun the identical pipeline but align separated components to person identities using a deterministic assignment rule (e.g., match each component to the closest single-person reference template). If accuracy on 5- and 10-person scenarios rises substantially above 56%, or if ISV drops by orders of magnitude, then the reported ceiling is at least partly a labeling artifact rather than purely a hardware limit. Another falsifier: collect CSI from an ESP32 with a 4-antenna external array (or 160 MHz bandwidth) and show ISV/ISD drops below the proposed viability threshold while acc","supporting_citations":[],"review_version":1}