{"id":"0245fc3c-3dcd-4bf3-a00d-d3d7a2edf0c4","arxiv_id":"2508.09839","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Starlink in flight gives median downlink throughput around 64 Mbps and uplink between 24 and 33 Mbps depending on altitude, with delay dependent on ground station and satellite links.","lead":"This paper measures Starlink internet speeds and delays on flights over the Baltic Sea and the Pacific Ocean. It is a first look at how the satellite service behaves for airline passengers, with median downloads around 64 Mbps.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Altitude uplink slowdown may be confounded with descent path and ground-station geography; need per-phase control.","rationale":"The reader's verdict is UNVERDICTED largely because abstract-only review prevents verification. My stress-test identifies the same weakest assumption: the altitude attribution is confounded. The suggested test is directly implementable from the paper's own data (if shared) and would settle whether the headline altitude effect is real or an artifact of descent geography. Since the paper is already UNVERDICTED, my recommendation is UNCHANGED; the concern reinforces the need for a conditional acceptance once data is available, but does not alter the current verdict.","tokens_in":770,"tokens_out":1538,"duration_ms":16791,"concrete_test":"Request the per-segment measurement logs with timestamps, GPS altitude, route coordinates, and ground-station IDs. Redo the analysis with altitude bins matched to geographic regions: e.g., compare uplink throughput during descent at airport X with cruise segments at the same region and same ground stations, or use a mixed-effects model with altitude as a continuous predictor and flight-phase/region as covariates. If the altitude coefficient loses significance after controlling for distance-to-airport and served ground station, the 17,000-ft split is not causal. Also check whether the split was chosen post hoc by testing multiple thresholds and reporting the false-discovery rate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper attributes the drop in median uplink throughput from ~33 Mbps above 17,000 ft to ~20 Mbps below to altitude. But the 17,000-ft split is likely a post hoc threshold, and altitude strongly co-varies with flight phase: descent begins near destination airports, over different ground-station footprints, with different satellite elevations and ISL routing. The abstract provides no equipment details, sample sizes, number of flights, measurement times, or path maps, so we cannot separate the altitude effect from geography. If the lower-altitude segments are all near airports or over different regions, then the observed 13 Mbps gap may reflect ground-station congestion or satellite handovers, not altitude. The claim's \"approximately 33 Mbps\" and \"around 20 Mbps\" also lack confidence intervals, so even the magnitude is unverified. This is a load-bearing concern because it undercuts the paper's most interpretable finding—altitude-dependent uplink behavior.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports an empirical measurement study of Starlink in-flight performance, based on measurements over the Baltic Sea and the Pacific Ocean. The abstract claims that a single user device achieves median downlink throughput of 64 Mbps and median uplink throughput of 24 Mbps, with uplink throughput approximately 33 Mbps above 17,000 feet and dropping to around 20 Mbps during descent. It also reports that round-trip time (RTT) depends strongly on the location of the pinged ground station and on the use of inter-satellite links, and it compares in-flight performance to a ground terminal from a residential user's perspective.","tokens_in":1006,"tokens_out":2366,"duration_ms":26580,"significance":"If the empirical results hold, the paper provides a useful first benchmark for aviation-grade Starlink performance, a domain that has seen rapid deployment but little published measurement. The intercontinental scope—Baltic Sea and Pacific Ocean—and the comparison with fixed ground-terminal performance are valuable. The RTT analysis using 5.5 hours of ping data also addresses an under-explored area of LEO satellite network behavior. However, the claims as stated in the abstract lack the methodological detail and statistical support needed to establish the central altitude-related finding.","major_comments":[{"comment":"The central claim that uplink throughput drops from ~33 Mbps above 17,000 ft to ~20 Mbps at lower altitudes is confounded with flight phase and geography. Descent typically coincides with approach to an airport, changes in ground-station access, satellite elevation, and ISL routing. The abstract gives no route information, ground-station locations, or data on whether the lower-altitude segments occur over different regions. Without per-phase controls or matched pairs, the 13 Mbps gap cannot be attributed to altitude rather than co-varying location or coverage conditions.","section":"Abstract"},{"comment":"No sample sizes, number of flights, measurement dates, or equipment details are provided. The term 'single user device' does not identify the Starlink terminal hardware, antenna generation, or measurement software. This missing information is essential for reproducibility and for judging whether the results generalize across aircraft, routes, or network configurations.","section":"Abstract"},{"comment":"The throughput and RTT results are reported as point values ('approximately 33 Mbps', 'around 20 Mbps') without confidence intervals, percentiles, or statistical tests. The phrase 'significant reduction' is used for the uplink drop, but no significance test or effect-size measure is mentioned. Without this, the reader cannot assess whether the observed differences are stable or within measurement noise.","section":"Abstract"}],"minor_comments":[{"comment":"The 17,000-ft threshold appears to be chosen after observing the data. If so, a brief justification or a data-driven selection procedure (e.g., changepoint detection) would strengthen the analysis; otherwise the threshold should be described as a post hoc split.","section":"Abstract"},{"comment":"The RTT discussion mentions 5.5 hours of ping measurements over the Pacific but does not state how many flights, ping targets, or measurement sessions this includes. Clarifying this would improve interpretability.","section":"Abstract"},{"comment":"The abstract uses imprecise language ('approximately 33 Mbps', 'around 20 Mbps') for median values. Reporting exact medians and interquartile ranges would increase precision.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"This review is based solely on the abstract because the full text was not provided. The abstract's empirical claims are plausible and timely, but the load-bearing altitude effect is confounded as presented. In the full paper, the authors should provide methodology details, flight/route information, and statistical bounds. If the full manuscript already addresses these points, this revision may be straightforward; if not, the altitude attribution needs additional analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this is an abstract-only read, but what's there is a genuine first: in-flight Starlink measurements from actual commercial flights over the Baltic and the Pacific, with concrete median numbers (64 Mbps down, 24 Mbps up). That fills a real gap—prior work covered fixed sites and ground vehicles, not aviation. The paper also includes a comparative look at ground terminal versus in-flight performance, which is the right framing for anyone who wants to know what the passenger experience actually buys you.\n\nWhat the abstract does well: it gives a clean empirical baseline, the RTT analysis based on 5.5 hours of pings is a reasonable first pass, and the authors are honest that this is a “first look.” That is a legitimate contribution, even if it is not a new method or theory.\n\nThe soft spot is the altitude split. The abstract says uplink drops from ~33 Mbps above 17,000 feet to ~20 Mbps below, but it gives no methodology, sample size, flight count, or confidence intervals. The threshold looks post hoc, and altitude co-varies with flight phase: descent happens near airports, over different ground-station footprints, with different satellite elevations and routing. So the 13 Mbps gap could easily be geography or handover effects rather than altitude. This is a load-bearing concern because the abstract presents the altitude effect as the key interpretable result. If the full paper controls for flight phase and route geometry, the finding might survive; if not, it's just a confounded observation. I'd also want to know equipment details—what device, what antenna, what Starlink terminal—because that changes what the numbers mean.\n\nOn the citation pattern, I can't see the full reference list, but the abstract credits prior Starlink studies, which is the right thing to do. No invented entities, no circular derivation; this is straight empirical measurement.\n\nWho is this for? People working on LEO satellite networks, aviation connectivity, and maybe regulators looking for real-world performance numbers. It deserves a serious referee, not a desk reject, because the dataset is useful and the aviation use case is under-documented. But the referee should press hard on the altitude analysis and whether the author's have separated altitude from geography. If they haven't, the paper is still a useful baseline, but the headline causal claim should be softened.\n\nRecommendation: send it to peer review, but with a referee who will demand the controls and the full flight metadata.","headline":"A genuinely first in-flight Starlink measurement set with usable numbers, but the altitude analysis in the abstract is under-specified and likely confounded with descent phase and geography.","tokens_in":1445,"tokens_out":1417,"would_cite":true,"duration_ms":19129,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An in-flight Starlink terminal on an aircraft achieves median downlink throughput of 64 Mbps and uplink throughput of 24 Mbps, with uplink performance varying by altitude.","keywords":["Starlink","in-flight connectivity","satellite internet","throughput measurement","round-trip time","inter-satellite links","LEO satellite networks","aviation networking"],"falsifier":"Overfly the same ground station at two altitudes in level flight, pinging a fixed target: if uplink throughput is the same at high and low altitude, the reported descent slowdown is caused by route or handover rather than altitude; if uplink drops at low altitude with the ground station held fixed, the altitude explanation survives.","tokens_in":724,"feed_emoji":"✈️","tokens_out":6603,"duration_ms":61158,"temperature":0.7,"pith_summary":"This paper reports the first in-depth in-flight measurement study of Starlink, based on flights over the Baltic Sea and the Pacific Ocean. It aims to establish what a single user device can expect from Starlink aboard an aircraft: median downlink throughput of 64 Mbps and median uplink throughput of 24 Mbps, with uplink at roughly 33 Mbps above 17,000 feet and about 20 Mbps during descent. It also shows that round-trip time is shaped more by which ground station is being pinged and by inter-satellite-link routing than by the satellite overhead alone. This matters because airlines are beginning to offer Starlink connectivity to passengers, and aviation use differs from fixed or ground-vehicle use in motion and antenna elevation. If these baselines hold, they give airlines and network operators a concrete performance target and a way to reason about where in-flight connectivity falls short.","feed_headline":"In-flight Starlink measured: 64 Mbps down, 24 Mbps up","feed_subtitle":"Uplink is near 33 Mbps at cruise but about 20 Mbps during descent, from flights over the Baltic Sea and Pacific.","key_machinery":"The measurement campaign itself is the machinery: a Starlink terminal onboard an aircraft, with a single user device running throughput and ping tests during flights over the Baltic Sea and the Pacific Ocean. Altitude-stratified analysis splits results at 17,000 feet. Ping measurements aimed at different ground stations, together with inter-satellite links (ISLs, the links that pass traffic directly between satellites), are used to separate the latency contributors.","core_discovery":"The central discovery is an empirical baseline for in-flight Starlink performance measured from a single user device. Across intercontinental flights, the service delivered median downlink throughput of 64 Mbps and median uplink throughput of 24 Mbps. The uplink degraded with the aircraft's descent: median throughput was about 33 Mbps above 17,000 feet and fell to around 20 Mbps at lower altitudes. Round-trip time depended strongly on the location of the ground station being pinged and on the use of inter-satellite links, and the paper uses 5.5 hours of ping measurements over the Pacific Ocean to explore queuing at satellites, ISL routing, and feeder-link congestion as sources of deviation f","pith_inferences":["If the altitude effect is causal, upload-heavy tasks on flights could be scheduled during cruise to roughly double effective uplink throughput compared with descent; the paper does not state this operational consequence.","Because RTT tracks ground-station location, a route-level latency forecast for a planned flight could be built from the constellation's ground-station map without new onboard hardware.","The descent-phase uplink drop may also occur for terrestrial mobile users whenever the elevation angle to the satellite changes quickly; a roadside or maritime test would show whether the effect is altitude-specific or more general.","A controlled test overflying the same ground station at high and low altitudes would tell whether the 17,000-foot split is altitude itself or the geography and handovers that accompany descent."],"forward_implications":["Aviation connectivity baselines should be altitude-aware: uplink capacity at cruise altitude appears to be roughly 1.6 times the uplink capacity during descent.","User-perceived latency on in-flight Starlink cannot be read off satellite altitude alone; ground-station location and ISL routing are primary factors.","The 5.5 hours of Pacific ping data give a starting point for separating satellite queuing, feeder-link congestion, and ISL routing as distinct latency sources.","The paper's comparison of a ground terminal and an aircraft terminal frames in-flight performance against residential expectations, giving airlines a concrete service-quality reference."],"supporting_citations":[],"fun_headline_variants":["Starlink in flight: 64 Mbps down, 24 up; descent cuts uplink to 20","First intercontinental Starlink flight test: 64 down, 24 up, altitude matters","From 35,000 feet: Starlink delivers 64 down, 24 up, but uplink dips","In-flight Starlink: uplink averages 33 at cruise, 20 during descent","Real Starlink flight data: latency depends on ground station and ISL routing"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The paper attributes the uplink slowdown during descent to altitude, but descent also changes the flight's location, ground-station distances, and satellite handovers, so altitude and those co-varying factors are entangled.","fun_headline_variants_meta":{"raw":{"variants":["Starlink in flight: 64 Mbps down, 24 up; descent cuts uplink to 20","First intercontinental Starlink flight test: 64 down, 24 up, altitude matters","From 35,000 feet: Starlink delivers 64 down, 24 up, but uplink dips","In-flight Starlink: uplink averages 33 at cruise, 20 during descent","Real Starlink flight data: latency depends on ground station and ISL routing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000459,"raw_usage":{"total_tokens":2168,"prompt_tokens":804,"completion_tokens":1364,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":1240}},"tokens_in":548,"tokens_out":1364,"duration_ms":15058,"temperature":1.0,"reasoning_tokens":1240,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:45:28.022785+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Overfly the same ground station at two altitudes in level flight, pinging a fixed target: if uplink throughput is the same at high and low altitude, the reported descent slowdown is caused by route or handover rather than altitude; if uplink drops at low altitude with the ground station held fixed, the altitude explanation survives.","supporting_citations":[],"review_version":1}