{"id":"19db25c0-0936-440c-a078-456aa5589b88","arxiv_id":"2506.02831","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 29-day survey with the EDA2 prototype found Starlink satellites emit unintended broadband radio emission across SKA-Low frequencies, including ITU protected bands, at levels that could harm EoR science.","lead":"Using an SKA-Low prototype station, researchers detected Starlink satellite radio emissions across 73 to 235 MHz, including in bands protected for radio astronomy. The survey of 76 million sky images shows satellite interference could threaten future Epoch of Reionization observations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Automated TLE-based identifications drive all headline statistics, but the paper reports no false-positive rate or trajectory-matching tolerance; the 30%-of-images and 93 Jy/beam numbers are therefore less secure than the qualitative detection claim.","rationale":"The reader's weakest-assumption analysis correctly identifies the TLE-based identification pipeline as the load-bearing element of every quantitative headline in the paper. I agree with that diagnosis. The concern is not that Starlink emission is absent—previous EDA2 studies and LOFAR observations (Bassa et al. 2024; Di Vruno et al. 2023) independently establish that Starlink satellites emit at SKA-Low frequencies, including in protected bands. The concern is that the paper's most striking numbers—approximately 30% of images containing a detected satellite, and a 93 Jy/beam mean sky flux density, five orders of magnitude above the EoR threshold—are outputs of a single unpublished matching algorithm with no quantified false-positive rate, no completeness estimate, and no error bars. Since the public data products are provided, this is directly testable, which is why the conditional verdict is appropriate rather than a rejection. The abstract's protected-band count for 150.05-153.00 MHz is also imprecise because it draws on the 153.13 MHz dataset whose band only partially overlaps the protected allocation; this is a real but secondary issue, and it does not change the qualitative finding that protected bands are affected, since the 150.78 MHz dataset alone shows hundreds of detections inside the band. The recommended action is to keep the conditional verdict and require the authors to provide the false-positive audit and uncertainty quantification in a revision.","tokens_in":11949,"tokens_out":14574,"duration_ms":160033,"concrete_test":"Using the public data release (Zenodo doi:10.5281/zenodo.15089852), re-run the identification pipeline on the 170.31 MHz dataset (the highest-contamination dataset) with an explicit trajectory-matching tolerance, for example a time residual below 10 s and a position residual below 0.5 degrees, and with a shuffled-timestep control to measure the false-positive rate. If the unique-satellite count and the mean flux density change by less than 20% relative to the paper's values, and the false-positive rate is below 1%, the concern is resolved; otherwise the 30%-of-images and 93 Jy/beam statistics need to be revised with a corrected matching criterion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claims—112,534 detections, 1,806 unique satellites, up to ~30% of images contaminated, and the 93 Jy/beam mean flux density—all come from the automated identification pipeline described in Section 2.1. A detection is a 2D Gaussian fit at a TLE-predicted location, and an identification requires five or more detections to match a single pass, but the matching tolerance is never specified ('closely matched' is not defined). No false-positive rate is measured, no independent ephemeris comparison (e.g., against high-precision SGP4 or the LOFAR catalogs) is presented, and no injection/recovery test establishes completeness. TLEs for Starlink can be stale, especially during orbit raising, so predicted positions may be offset; the 3-degree / 2-pixel position tolerance could then either attach real emission to the wrong satellite or miss it entirely. Because the 93 Jy/beam figure is formed by summing detected fluxes and dividing by the number of timesteps and polarisations, any false positives directly inflate it, while any missed detections are silently omitted, and the paper gives no uncertainty on this number. This matters most for the EoR-impact claim: the 'five orders of magnitude above the Wilensky et al. threshold' statement rests entirely on that flux estimate. The qualitative existence of Starlink UEMR is well supported by prior LOFAR and EDA2 work, but the quantitative impact statistics are not yet audited.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents a survey of Starlink satellite emission using the EDA2 prototype station at the SKA-Low site, analyzing ~76 million full-sky images from 29 separate ~24-hour observations at 24 frequencies between 73 and 235 MHz. The authors report 112,534 individual detections of 1,806 unique Starlink satellites, with up to ~30% of images contaminated in some datasets, and a mean sky flux density of 93 Jy/beam from unintended broadband emission, which they compare to the Wilensky et al. (2020) threshold for EoR power-spectrum corruption. They also report detections in ITU-protected radio astronomy bands, reflections of terrestrial FM radio off satellite surfaces, and anti-correlated XX/YY polarization behavior. The catalog and data products are made publicly available via Zenodo.","tokens_in":12236,"tokens_out":7318,"duration_ms":71851,"significance":"If the quantitative claims hold, this is an important and timely result for radio astronomy and spectrum policy. The survey is the largest of its kind, provides open data products, and systematically extends prior EDA2 and LOFAR work across the SKA-Low band. The qualitative detection of broadband unintended electromagnetic radiation (UEMR) across 73–235 MHz is strongly supported by the waterfall plots and is consistent with previous studies. The paper's main quantitative claims—the high fraction of contaminated images and the 93 Jy/beam mean flux—need the pipeline validation described below, but the overall phenomenon is not in doubt.","major_comments":[{"comment":"The criterion for promoting detections to identifications—'five or more detections closely matched the trajectory of the single pass'—is not defined: no matching tolerance, time window, or goodness-of-fit threshold is given. Because the headline numbers (112,534 detections, 1,806 unique satellites, per-dataset image fractions, and the 93 Jy/beam mean) are all products of this pipeline, the paper should state the matching algorithm explicitly and provide a false-positive/recovery analysis (e.g., by running the pipeline with scrambled or time-shifted TLE positions, or with injected synthetic sources). Without this, the quantitative results are not fully auditable.","section":"Section 2.1"},{"comment":"The abstract states that 703 satellites were identified between 150.05 and 153.00 MHz, but the 153.13 MHz dataset in Table 1 is centered outside that protected band. If that dataset is included in the 703 count, the wording is misleading; if excluded, the paper should clarify. The same issue affects the statement in Section 3 (Discussion), and the 153.13 MHz dataset's partial spectral overlap with the 150.05–153.00 MHz band should be described explicitly (or the abstract corrected).","section":"Abstract and Section 3 (protected-band counts)"},{"comment":"The 93 Jy/beam figure is presented as a 'mean' in the Results but as a 'lower limit' in the Discussion, and no uncertainty is given. The text should specify how images with no detections are handled in the average, why the 136.7 MHz and >320 MHz datasets are excluded, and what uncertainty (e.g., from dataset-to-dataset scatter or fitting errors) applies. The comparison to the Wilensky et al. (2020) 1 mJy threshold depends directly on this number, so a quantitative uncertainty or a conservative bound is needed.","section":"Section 3 (mean flux density)"}],"minor_comments":[{"comment":"The equation for the 1-D Gaussian contains a mismatched bracket in the text ('A·exp[− (ν−ν0)^2 / 2σ^2]'); please format it properly as A·exp(−(ν−ν0)^2/(2σ^2)).","section":"Section 2.1"},{"comment":"The 'Total' row appears misaligned: '2446692' is not labeled as the total timestep count, and the unique NORAD totals are omitted; please format the table so each column has a clear entry in the total row.","section":"Table 1"},{"comment":"The caption says 'the bottom line showing the median of all datasets'; this should be clarified to 'the median of the per-dataset median noise values' to avoid ambiguity.","section":"Figure 3 caption"},{"comment":"The 76% (v2-mini Ku) and 71% (v2-mini DTC) identification fractions are compared to the total in orbit on the final day of acquisition, but detections span 29 days during which the constellation grew by 477 satellites; please state whether the comparison is over the full survey period or the last day, and discuss the potential bias.","section":"Section 3 (Starlink models)"},{"comment":"The claim that the FM reflection originates from a 10 kW transmitter in Geraldton would be strengthened by providing the transmitter's frequency, the expected Doppler shift, and a calculation of the reflection geometry for at least one of the four passes.","section":"Section 3.4 (reflections)"},{"comment":"The paper notes that no v1.0 satellites were detected, but does not discuss possible reasons (e.g., deorbiting, frequency coverage, or sensitivity); a brief comment would help the reader reconcile this with Bassa et al. (2024) and Di Vruno et al. (2023).","section":"Section 3 (comparison with other studies)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for A&A and addresses a topic of broad interest. The qualitative result is solid, but the quantitative headline statistics require the validation and clarification described in the major comments. The authors have been transparent about some limitations, but the trajectory-matching specification is essential for a survey paper of this type."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is the most extensive survey of Starlink UEMR at SKA-Low frequencies so far, and the central claim—Starlink satellites emit broadband unintended radiation across 73–235 MHz, including inside ITU-protected bands—is credible and consistent with earlier LOFAR and EDA2 work. The paper deserves to be a reference point for how severe the problem is. But the quantitative headline figures (93 Jy/beam mean sky flux, up to ~30% of images contaminated, the protected-band counts) come from an automated TLE-matching pipeline whose tolerances and false-positive rate are not documented, so treat those numbers as order-of-magnitude estimates rather than audited statistics.\n\nWhat's new: 29 separate 24-hour datasets, roughly 76 million images, 112,534 detections of 1,806 unique Starlink satellites. That is a step change from the earlier EDA2 and LOFAR studies, both in sample size and in frequency coverage. They also report detections in the 73–74.6 MHz and 150.05–153 MHz protected bands, reflections of terrestrial FM at 99.7 MHz, and a qualitative anti-correlation between XX and YY polarization flux densities. Public data products are provided, which is genuinely useful for future monitoring.\n\nWhat it does well: the survey is systematic, the detection criteria are conservative (e.g., the 5% Gaussian-fit uncertainty cutoff), the comparison to prior work is honest, and the authors explicitly acknowledge that their numbers are lower limits because they miss faint or unresolved emission. The paper is readable and the method is reproducible from the released metadata.\n\nSoft spots, in rough order of importance. First, the identification pipeline: a detection is a 2D Gaussian fit at a TLE-predicted position, and an identification requires five detections to 'closely match' a pass, but 'closely' is never quantified. There is no false-positive rate, no injection/recovery test, and no independent ephemeris check. This matters for the exact counts and for the 93 Jy/beam number, because false positives would inflate it and missed detections would deflate it. The qualitative existence of Starlink UEMR does not hinge on this, but the EoR-impact claim ('five orders of magnitude') does. Second, the abstract's 703 satellites 'identified between 150.05–153.00 MHz' includes the 153.13 MHz dataset, whose band center sits outside that protected allocation. That is a minor overstatement and easy to fix. Third, flux densities are reported without uncertainties, and the 93 Jy/beam is called a 'mean' in the results and a 'lower limit' in the discussion—clarify which. Fourth, the polarization anti-correlation is shown but never quantified; a correlation coefficient or phase offset would strengthen it.\n\nThe citation pattern is fine; self-citations are to the same instrument and method, which is legitimate. The paper is squarely aimed at the radio-astronomy RFI community, spectrum regulators, and SKA-Low planning. Yes, it deserves a serious referee. I would send it out, but with a request that the authors document the matching tolerance, add a false-positive estimate (even a small manual inspection of random detections), and correct the protected-band tally. Those are revisions, not grounds for rejection.","headline":"Largest Starlink UEMR survey to date; the qualitative result is solid, but the headline numbers need an audit of the TLE-matching pipeline before they are used in regulatory claims.","tokens_in":12777,"tokens_out":3031,"would_cite":true,"duration_ms":32523,"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":"Starlink's unintended radio emission now pervades the SKA-Low band and may be bright enough to block the detection of the Epoch of Reionisation.","keywords":["radio astronomy","Starlink","unintended electromagnetic radiation","SKA-Low","radio frequency interference","megaconstellations","Epoch of Reionisation","Engineering Development Array 2"],"falsifier":"Rerun the same 76 million images through a blind moving-source finder that does not use predicted satellite positions; if the blind search does not recover roughly 112,000 detections and sky-averaged flux near 93 Jy/beam, the identification pipeline is biased. Alternatively, a single deep, well-calibrated observation of a field with and without predicted Starlink passes could directly measure the added sky flux and test whether it approaches 93 Jy/beam.","tokens_in":11730,"feed_emoji":"🛰️","tokens_out":7223,"duration_ms":61181,"temperature":0.7,"pith_summary":"This paper establishes that Starlink satellites now inject strong unintended radio emission across much of the SKA-Low band, and that the contamination is large enough to threaten the telescope's flagship science goal. Using the Engineering Development Array 2, an SKA-Low prototype station in the Australian radio-quiet zone, the authors analysed roughly 76 million full-sky images taken over 29 days and report 112,534 detections of 1,806 unique Starlink satellites between 73 and 235 MHz. In the worst dataset, nearly 30% of images contain a detectable Starlink satellite, and emission is seen inside two of the three ITU-protected radio-astronomy bands in the SKA-Low range. The paper's central quantitative claim is that the mean sky flux density from Starlink's unintended broadband emission is at least 93 Jy/beam, about five orders of magnitude above the level previously estimated to overwhelm Epoch of Reionisation power-spectrum measurements. If true, this single source of interference could prevent the SKA-Low from detecting the EoR unless the emission is mitigated.","feed_headline":"Starlink's stray radio noise is 100,000x the cosmic-dawn limit","feed_subtitle":"A 29-day survey finds Starlink satellites in up to 30% of images, inside protected bands too.","key_machinery":"The load-bearing instrument is the Engineering Development Array 2, a 256-dipole SKA-Low prototype that images the whole sky, horizon to horizon, without beamforming. The load-bearing method is an autonomous detection-and-identification pipeline: for every coarse-channel image, predicted positions of all satellites below 2,000 km perigee are calculated from orbital elements; a 2D Gaussian is fit at each predicted location subject to strict criteria; detections are promoted to identifications only when five or more fits match a single predicted satellite pass. This pipeline converts raw images into a catalogue of verified satellite events with measured flux densities, polarisation, and spectral structure, and its output underpins every prevalence and flux statistic in the paper.","core_discovery":"The paper's central discovery is that unintended electromagnetic radiation from Starlink satellites is now pervasive across 73-235 MHz, not confined to orbit-raising manoeuvres. The survey identifies 1,806 unique satellites, dominated by v2-mini Ku and v2-mini Direct-to-Cell models, with 76% and 71% of those populations detected respectively. It finds both broadband leakage and narrowband intended downlink pulses, including 13 satellites in the 73.00-74.60 MHz protected band and 703 in the 150.05-153.00 MHz protected band, plus four cases of terrestrial FM reflections off satellite structures at 99.70 MHz. The broadband emission shows anti-correlated flux in orthogonal polarisations and a time-varying comb-like spectral structure. Because the mean sky flux density from this UEMR is at least 93 Jy/beam, roughly five orders above the 1 mJy threshold that could overwhelm EoR integration, the authors conclude this emission alone could prevent detection of the Epoch of Reionisation.","pith_inferences":["If the UEMR per satellite stays constant while the constellation grows, the sky-averaged 93 Jy/beam floor and the ~30% image contamination rate will rise with the number of satellites; the released catalogue makes this trend directly measurable in follow-up surveys.","The time-varying comb-like spectral structure and anti-correlated polarisation may point to a specific on-board electronic source, such as a clock harmonic or switching regulator; if identified, shielding or filtering that component could reduce UEMR without redesigning the whole satellite.","The FM-reflection detections imply that even a fully electromagnetically quiet satellite could still produce episodic narrowband interference for radio telescopes by reflecting terrestrial transmitters, so mitigation strategies may need to consider satellite attitude and surface geometry, not just onboard emission."],"forward_implications":["At 161.7 and 170.5 MHz, roughly 30% of images contain at least one identifiable Starlink satellite, and the authors note the true fraction is likely higher because detection criteria were set conservatively to avoid misidentifications.","A mean sky flux density of at least 93 Jy/beam from Starlink UEMR is about five orders of magnitude above the level estimated to overwhelm Epoch of Reionisation power-spectrum integration, so this UEMR alone could prevent EoR detection with SKA-Low.","Starlink satellites are transmitting in ITU-protected radio astronomy bands, with 13 unique v2-mini Ku satellites in 73.00-74.60 MHz and 703 unique v2-mini satellites in 150.05-153.00 MHz.","No Starlink satellites were detected above 320 MHz, which may constrain the upper frequency limit of the broadband UEMR.","Open release of the identification catalogue gives regulators and operators a quantitative baseline for tracking future changes in this emission."],"supporting_citations":[{"why":"Previous EDA2 detection of Starlink emission at 137.5 and 159.4 MHz; defines the 100 s pulsing behaviour and the avionics/propulsion attribution that this survey extends.","marker":"Grigg, D. et al. 2023"},{"why":"LOFAR detections of Starlink narrowband and broadband emission with comb-like spectral structure; this survey compares its spectral and temporal findings against it.","marker":"Bassa et al. 2024"},{"why":"Earlier LOFAR study that detected 68 Starlink satellites and a reflection at 143.05 MHz; supplies the baseline for tracking model evolution and reflection phenomena.","marker":"Di Vruno, F. et al. 2023"},{"why":"Estimates that about 1 mJy of RFI can overwhelm Epoch of Reionisation power-spectrum integration; provides the threshold used to argue 93 Jy/beam is five orders of magnitude too high.","marker":"Wilensky et al. 2020"},{"why":"Earlier EDA2 study at 160.2 MHz whose detection and identification pipeline is reused here; 22 of its 26 satellites are re-detected across the new survey.","marker":"Grigg et al. 2024"},{"why":"Defines the EDA2 design and performance, including the sensitivity falloff above 300 MHz that explains the null detections at higher frequencies.","marker":"Wayth et al. 2021"},{"why":"Supplies the calibration method, derived from Benthem et al., that converts EDA2 visibilities to flux-calibrated images.","marker":"Sokolowski et al. 2021"}],"fun_headline_variants":["Starlink noise floods 30% of radio images in SKA survey","1,806 Starlinks leak radio into protected bands","SKA-Low sees Starlink in 1 in 3 images","Starlink broadband leak threatens epoch-of-reionization hunt","Largest Starlink radio survey finds pervasive leak"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire quantitative picture rests on the assumption that the TLE-based trajectory matching correctly identifies which detections belong to which Starlink satellite; if the predicted positions are wrong or the five-detection match rule is too strict or too lax, the reported counts, image fractions, and flux-density averages would be systematically off.","fun_headline_variants_meta":{"raw":{"variants":["Starlink noise floods 30% of radio images in SKA survey","1,806 Starlinks leak radio into protected bands","SKA-Low sees Starlink in 1 in 3 images","Starlink broadband leak threatens epoch-of-reionization hunt","Largest Starlink radio survey finds pervasive leak"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1531,"prompt_tokens":1051,"completion_tokens":480,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":391}},"tokens_in":667,"tokens_out":480,"duration_ms":5280,"temperature":1.0,"reasoning_tokens":391,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:14:07.101036+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the same 76 million images through a blind moving-source finder that does not use predicted satellite positions; if the blind search does not recover roughly 112,000 detections and sky-averaged flux near 93 Jy/beam, the identification pipeline is biased. Alternatively, a single deep, well-calibrated observation of a field with and without predicted Starlink passes could directly measure the added sky flux and test whether it approaches 93 Jy/beam.","supporting_citations":[{"cited_title":"J., Sokolowski, M., et al","cited_arxiv_id":null,"evidence_quote":"Previous EDA2 detection of Starlink emission at 137.5 and 159.4 MHz; defines the 100 s pulsing behaviour and the avionics/propulsion attribution that this survey extends."},{"cited_title":"G., Di Vruno, F., Winkel, B., et al","cited_arxiv_id":null,"evidence_quote":"LOFAR detections of Starlink narrowband and broadband emission with comb-like spectral structure; this survey compares its spectral and temporal findings against it."},{"cited_title":"J., Barry, N., Morales, M","cited_arxiv_id":null,"evidence_quote":"Estimates that about 1 mJy of RFI can overwhelm Epoch of Reionisation power-spectrum integration; provides the threshold used to argue 93 Jy/beam is five orders of magnitude too high."},{"cited_title":"2024, A&A","cited_arxiv_id":null,"evidence_quote":"Earlier EDA2 study at 160.2 MHz whose detection and identification pipeline is reused here; 22 of its 26 satellites are re-detected across the new survey."},{"cited_title":"2021, The Engineering Develop- ment Array 2: design, performance and lessons from an SKA-Low prototype station","cited_arxiv_id":null,"evidence_quote":"Defines the EDA2 design and performance, including the sensitivity falloff above 300 MHz that explains the null detections at higher frequencies."},{"cited_title":"B., Bhat, N","cited_arxiv_id":null,"evidence_quote":"Supplies the calibration method, derived from Benthem et al., that converts EDA2 visibilities to flux-calibrated images."}],"review_version":1}