{"id":"2244496e-a435-4684-8cc8-44d916e9e5b4","arxiv_id":"2607.18549","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Uplink SRS channel features plus a random forest locate a commercial 5G phone indoors with 0.12 m MAE on random splits, but 1.78 m on spatially separated test areas.","lead":"This paper builds a 5G testbed where a positioning app on an open radio network uses a phone's uplink sounding signals to locate it indoors, reporting 0.12 m average error on random test points. It is a feasibility demonstration: accuracy drops to 1.78 m on spatial regions excluded from training, so the method interpolates but does not extrapolate.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 0.12 m MAE rests on an 80/20 random split over dense grid points, so training and test frames share physical locations; the paper's own spatially separated split (1.78 m) shows this is interpolation, not independent localization.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing issue: the random train-test split over densely sampled grid points does not produce an independent test set because frames from the same physical location appear in both partitions. The paper's own spatially separated split is internal evidence of this leakage, showing a 14× error increase when locations are truly unseen. My review does not change the reader's CONDITIONAL verdict; it strengthens it by pinpointing the specific experimental control that would settle the question. I am not raising a consensus-based objection: the engineering is plausible, the feature extraction is physically motivated, and the authors do disclose the spatial-split limitation. However, the abstract's unqualified '0.12 m MAE' and the conclusion's 'sub-meter active positioning' overstate what the evidence supports. The proposed grouped-split test is concrete, inexpensive, and directly determines whether the central claim generalizes beyond interpolation. Until that test is run, the paper should be treated as a promising demonstration with a conditional verdict, not as a validated operational capability.","tokens_in":4377,"tokens_out":1967,"duration_ms":26713,"concrete_test":"Re-evaluate the model using a grouped split by physical grid coordinate: assign all frames from each surveyed grid point to either the training set or the test set, never both, and report MAE on the held-out grid points. If the resulting MAE remains below ~0.5 m, the sub-meter claim survives; if it rises toward the 1.78 m spatial-split value, the 0.12 m headline is an artifact of spatial leakage. As a secondary check, instrument E2SM-LLC delivery timestamps and xApp inference time to verify whether 'real-time' has any quantitative support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claim — sub-meter localization with 0.12 m MAE — depends entirely on an 80/20 random split over 23,000 SRS frames collected at dense reference grid points in a 6×6 m room. With dense sampling, frames from the same physical grid point almost certainly appear in both training and test partitions. The model can therefore memorize per-point radio fingerprints, and the reported MAE measures interpolation within the surveyed grid, not localization of unseen positions. The paper itself provides the decisive control: when entire spatial regions are held out, MAE degrades to 1.78 m, a 14× increase. This is disclosed, but the abstract and conclusion still headline 0.12 m and claim 'sub-meter active positioning.' The 42% improvement over the RSSI baseline (0.2153 m) is also computed under the same leaky random split, so it does not independently validate generalization. The 'real-time' claim is additionally unsupported because no E2SM-LLC or end-to-end latency is measured; the only timing figure is a ~1 s EMA lag. The load-bearing weakness is thus the evaluation protocol: the headline accuracy metric is not an independent test of localization capability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an uplink SRS-based indoor localization system implemented on an OpenAirInterface (OAI) 5G testbed. A Positioning xApp on the Near-RT RIC extracts Channel Frequency Response (CFR) from SRS telemetry received via E2SM-LLC, converts it into a 32-dimensional physics-aware feature vector (PDP, inter-antenna phase differences, received power/spatial symmetry), and feeds it to a Random Forest regressor for 2D position estimation. Experiments in a 6×6 m multipath-rich lab with over 23,000 SRS frames report 0.12 m MAE under an 80/20 random split, 1.78 m MAE under a spatially separated blind split, and 0.2153 m MAE for an RSSI-only baseline. The paper also describes a real-time dashboard with EMA smoothing and multi-UE probability heatmaps, and claims sub-meter active positioning with real-time operation.","tokens_in":4709,"tokens_out":3065,"duration_ms":34168,"significance":"If the claims hold, the work provides a valuable integration of standardized E2SM-LLC SRS collection with a lightweight RF positioning pipeline, using commercial UE and an O-RAN testbed. The paper's strengths include the end-to-end system implementation, the physics-aware feature design, and the explicit disclosure of a spatially separated evaluation that degrades accuracy to 1.78 m. The latter control is scientifically informative, and the testbed contribution is useful to the O-RAN/ISAC community. However, the headline 0.12 m accuracy and the 'real-time' claim currently outrun the evidence, so the paper's significance depends on revising the claims to match the actual evaluation protocol.","major_comments":[{"comment":"The 0.12 m MAE is obtained under an 80/20 random split of 23,000 SRS frames collected on a dense grid in a 6×6 m room. With dense sampling, frames from the same physical location almost certainly appear in both training and test partitions, making this an interpolation metric rather than a test of localization of unseen positions. The paper's own spatially separated evaluation gives 1.78 m MAE, a 14× degradation, and this is disclosed. Nevertheless, the abstract and conclusion still headline 0.12 m and claim 'sub-meter active positioning.' This is a load-bearing mismatch between the claim and the evidence. The spatial-split result should be presented as the primary generalization metric, or all claims must be explicitly scoped to interpolation within the surveyed grid.","section":"Abstract and Sec. IV"},{"comment":"The title and abstract advertise 'real-time' localization, and Sec. II states that the mapping f:H→y operates under 'strict microsecond-level latency constraints.' However, no end-to-end latency measurement is provided. The only timing figure in the paper is the EMA filter's approximately 1 s lag, which is a visual smoothing artifact rather than a system latency bound. Without measuring E2SM-LLC delivery time, inference time, and end-to-end delay, the real-time claim is unsupported. Either add latency measurements or remove 'real-time' from the title/abstract and qualify the claim as near-real-time.","section":"Secs. II and IV"},{"comment":"The 42% improvement over the RSSI baseline (0.1238 vs. 0.2153 m) is computed under the same random split that leaks spatial information, so it does not independently validate generalization to new positions. In addition, no error bars, confidence intervals, or repeated-trial statistics are reported for any of the MAE values. Given the order-of-magnitude difference between the random-split and spatial-split results, the paper should report variance across multiple random seeds and also provide the RSSI baseline for the spatial split. This is necessary to assess whether the improvement is robust.","section":"Sec. IV (RSSI baseline and uncertainty)"}],"minor_comments":[{"comment":"The sentence stating 'approximately 85% of predictions within a 0.5 m error threshold' does not specify whether this refers to the random split or the spatial split. The CDF curves for both settings should be clearly labeled in Fig. 3 and in the text.","section":"Sec. IV / Fig. 3"},{"comment":"The reference to 'strict microsecond-level latency constraints' in Sec. II is not tied to any system requirement or measurement. Either substantiate it with a latency budget or remove it.","section":"Sec. II / Sec. IV"},{"comment":"The definition of the peak reference tap n_ref and the truncation to 16 taps is not fully specified. Please state how n_ref is selected and clarify whether the truncation is independent of the number of SRS subcarriers N.","section":"Eq. (2)"},{"comment":"The claim that the lightweight pipeline is designed for 'rapid re-deployment in new environments' is plausible but not evaluated in the paper. Consider softening this claim or adding a brief discussion of the retraining cost.","section":"Sec. V"}],"recommendation":"major_revision","confidential_remarks":"The paper describes a usable testbed and the spatial-split control is an honest disclosure, but the headline accuracy and 'real-time' claim are not aligned with the evidence. The required changes—reframing the primary accuracy metric, adding latency measurements, and providing variance estimates—are within the scope of a major revision. I would not reject the manuscript, because the underlying system and the disclosed spatial-split result provide a solid basis for a revised paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine integration effort — an uplink SRS-based localization xApp on OAI/FlexRIC, using a commercial phone, with a 32-D physics-aware feature set and a random forest regressor. The 0.12 m MAE in the abstract is real but it is an interpolation metric. The paper's own spatially separated evaluation gives 1.78 m, a 14× jump, and that is the number that tells you what the system can do on unseen positions. The stress-test note is right, and the paper deserves credit for reporting the spatial split instead of burying it.\n\nWhat is new: as far as I can tell, this is the first end-to-end demonstration of uplink SRS positioning over OAI with a commercial UE and standard E2SM-LLC telemetry. The engineering is non-trivial: making SRS/CFR extraction work through the E2 interface, aligning PDPs, computing phase features, and building a live dashboard. The feature design is physically sensible, and the RSSI-only baseline (0.2153 m) supports the claim that delay/phase structure helps — though that comparison also runs on the leaky random split, so it only shows in-distribution benefit.\n\nSoft spots, in order of severity:\n1. Evaluation protocol. The 80/20 random split over 23k frames from a dense grid almost certainly puts frames from the same grid point in both train and test. The model can memorize per-point fingerprints. The 1.78 m spatial split is the honest generalization number. The abstract and conclusion still headline 0.12 m and say 'sub-meter active positioning' without qualifying that it is within-grid interpolation. That framing should change.\n2. 'Real-time' is asserted but not measured. No E2SM-LLC latency, no end-to-end latency; only ~1 s EMA lag is mentioned. The title makes a claim the evaluation does not support.\n3. No error bars, no repeated runs, no confidence intervals. Single 6×6 m room. That is acceptable for a feasibility study, but it limits what you can conclude.\n4. No code or data release. The demo video is not described in technical detail; for a reproducibility point, release would help a lot.\n\nNone of this makes the integration fake. The authors explicitly state that tree-based regressors struggle with extrapolation, which is the right acknowledgment. But the paper frames the interpolation result as the headline, and that is the load-bearing weakness.\n\nWho benefits: researchers building O-RAN/ISAC localization prototypes, or anyone needing a reference architecture for getting SRS telemetry from OAI to a RIC xApp. It is a solid feasibility data point, not a new localization algorithm.\n\nRecommendation: send to peer review. A good referee will force a corrected evaluation and proper latency measurement, and the paper will be stronger for it. If you only read one number, read 1.78 m, not 0.12 m.","headline":"A useful OAI integration that is honest about its own limitation: the 0.12 m MAE is interpolation; the real spatial-split number is 1.78 m.","tokens_in":5197,"tokens_out":3674,"would_cite":true,"duration_ms":43828,"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":"Uplink SRS telemetry, routed through the E2 interface, can locate a smartphone indoors with 0.12 m mean error — but only inside the surveyed area, since spatially unseen regions degrade to 1.78 m.","keywords":["Indoor localization","Sounding Reference Signal","Channel Frequency Response","Random Forest","O-RAN xApp","Integrated Sensing and Communication","5G positioning","Real-time 5G sensing"],"falsifier":"Run the positioned xApp at a grid coordinate never visited during training and measure the error: if it stays near 0.12 m rather than 1.78 m, the claimed interpolation dependency is wrong. Separately, record timestamps at the base station and at the positioning app to check whether end-to-end latency actually meets the real-time claim.","tokens_in":4298,"feed_emoji":"📡","tokens_out":6065,"duration_ms":60003,"temperature":0.7,"pith_summary":"The paper claims that the uplink Sounding Reference Signal telemetry a 5G base station already receives can be exploited to locate a commercial smartphone indoors with sub-meter error. The authors build a positioning application that converts the channel frequency response into a 32-dimensional feature vector and feeds a Random Forest regressor. In a dense grid inside a 6-by-6 meter multipath laboratory, they measure 0.12 m mean absolute error with a random 80/20 data split, and 1.78 m when entire spatial regions are held out. The result points to an economical path for integrated sensing and communication on standard 5G radio hardware, with the caveat that the current model only interpolates within surveyed areas.","feed_headline":"5G uplink signals locate indoor phones within 0.12 m","feed_subtitle":"A lightweight positioning app turns standard 5G signal features into coordinates without deep learning.","key_machinery":"The central object is the 32-dimensional physics-aware feature vector derived from the 4-antenna channel frequency response. It combines a 16-tap normalized power delay profile, 8 trigonometric inter-antenna phase difference components, and 8 received-power/symmetry heuristics. This representation lets a Random Forest regressor map radio fingerprints to coordinates without deep learning. The lower-layer control service model is the conduit that delivers the SRS telemetry; the Random Forest is the estimator.","core_discovery":"The central claim is that uplink SRS measurements, exposed through the E2 interface's lower-layer control service model, can be turned into a compact feature set — normalized power delay profile taps, inter-antenna phase differences, and received-power heuristics — whose Random Forest regression predicts 2D coordinates. The paper reports 0.1238 m MAE in the random-split evaluation and 1.78 m in spatially blind evaluation, concluding that tree-based regressors interpolate well within the surveyed region but extrapolate poorly. The pipeline runs in real time and visualizes single- and multi-UE tracking with confidence heatmaps.","pith_inferences":["The 0.12 m figure likely overstates accuracy in a real deployment, because the random split allows frames from the same physical grid point to appear in both training and testing; a truly unseen location is better predicted by the 1.78 m spatial-split number.","Since the authors never measure E2 interface latency or frame loss, the 'real-time' claim rests on an unverified delivery path; logging timestamps at both ends would quantify the actual end-to-end delay.","A natural extension is to swap the Random Forest for a spatially continuous regressor, such as a Gaussian process; if the spatial-split error then drops dramatically, the tree model's piecewise-constant output is the main bottleneck.","The overlapping probability heatmaps that detect multi-UE convergence could serve as a zero-cost proximity alert in future integrated sensing and communication systems, but the false-alarm rate needs quantification."],"forward_implications":["If correct, any O-RAN-compliant gNB with SRS and the lower-layer control service model can host sub-meter indoor localization without modifying the UE.","The 1.78 m spatially-split result bounds the approach: the model cannot extrapolate beyond surveyed regions, so deployment requires dense site surveys.","The RSSI-only baseline's 0.2153 m MAE shows delay and phase features contribute a 42% error reduction, making multipath structure essential.","At microsecond inference per sample, the pipeline is light enough for near-real-time tracking with an EMA smoothing lag of ~1 s.","The multi-UE collision-detection effect falls out of overlapping probability heatmaps, needing no extra computation."],"fun_headline_variants":["5G uplink SRS locates indoors to 12 cm, but only in mapped zones","RF-based indoor positioning hits 0.12 m error, but fails outside trained area","Real-time 5G localization: 12 cm accuracy, but only in known areas","Uplink SRS + Random Forest: indoor positioning to 12 cm in trained zones","5G-based indoor localization: 0.12 m MAE, but blind tests show 1.78 m"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The evaluation assumes that an 80/20 random split over densely sampled grid points produces an independent test set; because frames from the same physical location can appear in both training and testing, the 0.12 m MAE is an interpolation metric, and the spatially separated split (1.78 m) shows the real generalization gap.","fun_headline_variants_meta":{"raw":{"variants":["5G uplink SRS locates indoors to 12 cm, but only in mapped zones","RF-based indoor positioning hits 0.12 m error, but fails outside trained area","Real-time 5G localization: 12 cm accuracy, but only in known areas","Uplink SRS + Random Forest: indoor positioning to 12 cm in trained zones","5G-based indoor localization: 0.12 m MAE, but blind tests show 1.78 m"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000823,"raw_usage":{"total_tokens":3426,"prompt_tokens":726,"completion_tokens":2700,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":2580}},"tokens_in":470,"tokens_out":2700,"duration_ms":21880,"temperature":1.0,"reasoning_tokens":2580,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T15:03:39.014069+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the positioned xApp at a grid coordinate never visited during training and measure the error: if it stays near 0.12 m rather than 1.78 m, the claimed interpolation dependency is wrong. Separately, record timestamps at the base station and at the positioning app to check whether end-to-end latency actually meets the real-time claim.","supporting_citations":[],"review_version":1}