{"id":"18e85836-fc87-4c6d-bba2-ba912c383ab6","arxiv_id":"2505.16519","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A system that broadcasts simplified webpages and ChatGPT responses over FM radio with SMS uplink achieved 10 kbps downlink and low loss rates in a six-week Cameroon deployment.","lead":"SONIC uses FM radio broadcasts to send simplified webpages and AI chat responses to phones, with SMS as the return channel. In a six-week trial in Cameroon, 30 users requested pages and ChatGPT queries through the system, which showed low loss for most receivers with strong signal.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Figure 14's median losses in the dominant RSSI bins (about 20-22%) appear to contradict the abstract's claim of 'less than 20% loss for a majority of transmissions with signal strength above -90 dBm'.","rationale":"The reader's weakest assumption focused on the fragility of the FM station's audio path and off-peak scheduling. That concern is real and documented in Section 5, but there is a more direct, internal inconsistency in the paper's central quantitative claim. The abstract states a majority of transmissions have loss < 20% for RSSI > -90 dBm, yet Figure 14 shows median losses around 20-22% in the RSSI bins that contain most transmissions. This is a correctness issue within the paper's own data, not an external assumption. The reader did note that abstract numbers are not clearly measured in the body, which partially overlaps. However, the reader did not identify the specific contradiction with Figure 14. The concern does not fully overturn the central feasibility claim, because 20% loss is still within the tolerance of the pixel-interpolation user study, and the user engagement results are separate. But it does mean the abstract's headline loss figure is overstated or ambiguously defined, and it strengthens the need for artifact release and clear measurement definitions. Since the reader already recommended CONDITIONAL acceptance for related reasons, my verdict remains UNCHANGED; the condition should now explicitly include reconciling the loss statistics with Figure 14 and releasing the raw loss data.","tokens_in":19358,"tokens_out":3714,"duration_ms":31202,"concrete_test":"From the raw deployment logs (which the authors should release), compute the fraction of transmissions with measured RSSI > -90 dBm that have frame loss < 20%, including all transmissions that were attempted (even those where the receiver decoded no frames). If this fraction is not > 50%, the abstract's claim is false. Also recompute the per-bin medians from the raw data to confirm the Figure 14 values; if the medians are indeed ~20% for the -90 to -60 dBm bins, the abstract must be amended to state the distribution accurately.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim in the abstract is that 'less than 20% loss for a majority of transmissions with signal strength above -90 dBm'. The paper's own deployment data in Figure 14 directly bear on this. The violin plot shows median loss percentages per RSSI bin: 20.4% for (-90,-80], 22.1% for (-80,-70], 20.9% for (-70,-60], and 11.9% for (-60,-50]. These three bins together account for more than 60% of all transmissions (summing the widths listed: 20.4+22.1+20.9+11.9+... The exact proportions are given on the figure: 20.4%, 22.1%, 20.9%, 11.9%, 4.8%, etc.). Since a median of approximately 20% means that roughly half of the transmissions in each of those bins exceed 20% loss, the majority of transmissions with RSSI above -90 dBm are not below 20% loss as claimed. The abstract either uses a different definition of 'majority' (e.g., weighted across all bins, where the low-loss -60 to -50 and -50 to -40 bins pull the overall median down) or a different subset (e.g., only completed transmissions), but no such qualification is given. If the loss rates in the dominant RSSI ranges are actually centered around 20-22%, the headline number is misleading. This matters because the paper's own interpolation study (Figure 20) shows that content clarity remains acceptable at 20% loss, so the system could still be usable; but the abstract's specific numerical claim is not supported by the displayed data. The discrepancy also raises doubts about how 'loss' is computed (e.g., whether transmissions with zero received frames are included or excluded).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents SONIC, a system that delivers simplified web content and LLM-generated responses to low-end smartphones using FM radio as a broadcast downlink and SMS as an uplink. The authors describe the server and client architecture, a modified LineageOS build that exposes FM tuner audio to applications, image-based webpage compression with pixel interpolation, and a hyperlink-pushing heuristic. They report a six-week deployment at an FM radio station in Cameroon with 30 users, and claim a sustained downlink throughput of 10 kbps, less than 20% transmission loss for a majority of transmissions above -90 dBm, and strong user engagement across web browsing and ChatGPT interactions.","tokens_in":19722,"tokens_out":5709,"duration_ms":49258,"significance":"If the headline claims are supported, SONIC would be a genuinely useful contribution to low-cost connectivity in low-income regions: it repurposes near-ubiquitous FM radio and SMS infrastructure, and the six-week deployment with 30 real users is a substantial step beyond lab-only evaluations. The paper also makes a credible case that FM-capable phones are widely available and affordable in the target regions. The main concerns are not about the overall concept but about whether the abstract's quantitative claims are backed by the reported measurements.","major_comments":[{"comment":"The abstract claims 'less than 20% loss for a majority of transmissions with signal strength above -90 dBm,' but the paper does not report the statistic that directly supports this claim. Figure 14 shows violin plots with bin widths labeled as percentages of total transmissions, but it does not state the median loss per bin or the fraction of transmissions with loss below 20% for the RSSI > -90 dBm subset. The text in §6 says data points are 'more concentrated below 20%' for the -80 to -50 dBm range, which is not a quantitative majority statement. Relatedly, 'loss' is not defined: it could mean lost frames, lost bytes, or failed transmissions, and the denominator is unclear. Please report the exact fraction of transmissions with loss below 20% among those with RSSI above -90 dBm, and define the loss metric precisely.","section":"§6, Figure 14, and abstract"},{"comment":"The abstract and Section 1 describe 'a sustained downlink throughput of 10 kbps' as a demonstrated result, but Section 3.1.1 defines 10 kbps as the modulation rate of the Quiet OFDM profile (92 sub-carriers, 9.2 kHz center frequency). Section 6 contains no measured throughput or goodput data, such as bytes successfully delivered per hour or per transmission window. As written, the headline number is a configuration parameter, not a measured result. Please either remove 'sustained' and rephrase as the configured modulation bitrate, or add measured throughput/goodput from the deployment.","section":"§3.1.1 and §6"},{"comment":"Section 1 states that 'Mean decoding accuracy remained at 71% under real-world conditions,' but this metric is never defined and does not appear anywhere in Section 6. The evaluation instead reports transmission loss percentages and request completion rates. If 'decoding accuracy' refers to the fraction of successfully decoded frames, bytes, or transmissions, that definition and the supporting measurement must be provided; otherwise, the sentence should be removed from the introduction.","section":"§1 and §6"},{"comment":"The scalability analysis is based on 'FCFS queue simulations' whose methodology is not described. The paper does not state how the baseline traffic was scaled, how cache behavior was modeled, what service time was assumed per request, whether the simulated players experience transmission loss, or how the number of frequencies changes the service rate. Since the abstract and conclusion describe SONIC as 'scalable,' these simulation assumptions need to be specified, or the figure should be explicitly labeled as an illustrative back-of-the-envelope extrapolation rather than a simulation result.","section":"§6, Figure 16"}],"minor_comments":[{"comment":"The unit is written as '-90 dbM' and 'dBM'; it should be '-90 dBm' throughout.","section":"Abstract and §1"},{"comment":"The sentence 'These results sheds light on the affordability of web access' should be 'These results shed light...'.","section":"§2"},{"comment":"The percentages printed beneath the RSSI bins are the proportions of transmissions, but the violin plot itself is not accompanied by a legend describing what the median, quartiles, or whiskers represent; please add a legend or a short caption explaining the plot elements.","section":"§6, Figure 14"},{"comment":"The term 'completion rate' is used but not defined; it should state whether a request counts as complete only when all frames are decoded, or whether partial content counts.","section":"§6, Figure 17"},{"comment":"The line 'over 95% of transmissions experienced loss rates below 10%' refers to ideal lab conditions, but this is easy to misread as a field result because Figure 14 is discussed later; consider adding an explicit qualifier such as 'under controlled lab conditions' in the sentence.","section":"§4, Figure 10"}],"recommendation":"major_revision","confidential_remarks":"The paper leans on prior work [57] for the pixel-interpolation evaluation; since the current paper's readability claims depend on that study, the editor may want to confirm that [57] is publicly available and that the present manuscript clearly delineates new contributions. No code or data release is mentioned, which would help reproducibility of the Figure 14 and Figure 16 analyses."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Sarah,\n\nThis paper earns a real read. It is a genuine field deployment of a plausible idea: use FM broadcast as a low-cost downlink, SMS as uplink, and pre-rendered pages as images. The Cameroon study with 30 users over six weeks is the actual contribution. The LineageOS modification to expose FM audio without root is useful engineering, and the LLM-over-FM extension is a nice add. The authors are also transparent about the messiness: ACK SIMs blocked, station staff missing the 10 PM switch, power outages, earphone antennas. That gives the loss data credibility.\n\nThe strongest evidence is the loss-vs-RSSI relationship in Figure 14 and the per-user completion rates. The abstract's 'less than 20% loss for a majority' is defensible if you read the text, which says the -80 to -50 dBm range (the bulk of transmissions) is mostly below 20% loss. Note: a stress-test claim I saw misreads Figure 14, shifting the percentage labels by one bin; that concern does not hold up.\n\nWhere it gets soft: 'sustained downlink throughput of 10 kbps' is the modulation rate of their Quiet profile, not a measured end-to-end throughput. The 'mean decoding accuracy of 71%' appears in the introduction and never again—no definition, no measurement. No artifact or dataset is released, so the deployment numbers cannot be independently checked. The 'first system' claim is an overclaim; the core mechanism was in their own CoNEXT '24 paper. The link-pushing regression is a fine heuristic, but it is not load-bearing, so that is a minor issue.\n\nThe right verdict is conditional. The idea is not new, but the deployment is real and the system is plausible. This is for the systems-for-development crowd. A serious referee should ask for a corrected abstract, a defined and measured accuracy metric, a data release, and a toned-down novelty statement. Those are fixable. I'd send this to review.","headline":"A real deployment of FM-broadcast web delivery, honestly reported; the abstract overstates some numbers and the novelty claim needs toning down.","tokens_in":20307,"tokens_out":5178,"would_cite":true,"duration_ms":37833,"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":"SONIC claims that existing FM broadcast infrastructure plus SMS can deliver useful web content and LLM chat answers to ordinary Android phones, and backs this with a six-week, 30-user deployment in Cameroon sustaining 10 kbps downlink and…","keywords":["FM radio data broadcast","SMS uplink","digital divide","web access for developing regions","data-over-sound","OFDM","Android FM tuner","LLM over broadcast"],"falsifier":"Run the same nightly broadcast for one week at a second FM station, measuring per-frame loss against signal strength; if the majority of receptions above -90 dBm exceed 20% loss, or if any single day's transmission fails because station staff did not flip the switch, the paper's reliability claim does not generalize.","tokens_in":19157,"feed_emoji":"📻","tokens_out":7457,"duration_ms":62004,"temperature":0.7,"pith_summary":"SONIC is a complete data-delivery system that treats an FM radio station as a broadcast downlink and SMS as a narrow uplink, aiming to give people who cannot afford mobile data a usable slice of the Web. The paper's central claim is that this combination works in the field: a six-week deployment in Cameroon sustained 10 kbps, kept loss below 20% for most transmissions when signal strength was above -90 dBm, and drew regular use for both webpage requests and LLM chat queries. The authors argue this matters because 2.6 billion people remain offline, and FM receivers are built into many affordable Android phones while FM broadcast costs stay fixed regardless of audience size. If correct, the system shows that repurposing existing broadcast infrastructure is a viable interim path to connectivity, not just a laboratory demonstration.","feed_headline":"FM radio plus SMS delivers web pages and AI answers offline","feed_subtitle":"A six-week Cameroon trial kept a 10 kbps downlink with under 20% loss at good signal strength, pointing to a low-cost path to connectivity.","key_machinery":"The system's load-bearing piece is an audio-over-FM chain built on OFDM, a multi-carrier modulation that splits the channel into narrow orthogonal sub-carriers: 92 sub-carriers centered around 9.2 kHz carry 10 kbps of coded data, modulated by an open-source audio modem library. On the send side, webpages are rendered headlessly, captured as 320-pixel-wide WebP screenshots, wrapped in a structured file format whose headers mark metadata, link maps, and payload frames, and then pushed through the station's audio path. On the phone, a modified open-source Android build opens the built-in FM tuner to apps, the app decodes frames with CRC32 checksums plus inner and outer forward-error correction, and lost pixels are filled by nearest-neighbor interpolation, exploiting the left-to-right structure of text. Requests travel by SMS; interactive hyperlinks are stored as click coordinates, and a scored push mechanism prefetches the three most likely next links during idle time.","core_discovery":"The paper claims that pre-rendered webpage screenshots and LLM responses can be encoded as OFDM audio, broadcast over an FM station's off-peak hours, and decoded on unrooted Android phones using their built-in FM tuners, with SMS providing the request channel. In its Cameroon deployment, transmissions sustained 10 kbps; a majority of receptions with signal strength above -90 dBm lost less than 20% of frames, and mean decoding accuracy was 71%. Users made 1,737 URL requests and 2,936 query requests over six weeks, and the broadcast nature meant every tuned-in phone received every transmission, feeding a shared knowledge hub even for users who never sent a request. The paper further claims one FM frequency can serve roughly 30 active users in a seven-hour nightly window, and 300 users with ten frequencies.","pith_inferences":["Editorial inference: the same off-peak broadcast channel could prefetch popular pages and LLM answers into phone caches, reducing later cellular usage to interactive or personalized requests only.","Editorial inference: since LLM responses lose value with a single lost frame while webpage images tolerate partial loss, rebroadcasting short LLM responses with added redundancy would likely lift completion rates more than increasing the overall bitrate.","Editorial inference: moving from the analog audio path to an FM subcarrier or a dedicated digital channel could raise throughput well beyond 10 kbps without changing the client's request-by-SMS model.","Editorial inference: the push priority score was fit on click data from one online study population, so its coefficients may shift for other regions; logging clicks on pushed links during deployment would let the metric adapt to local browsing habits."],"forward_implications":["If the deployment numbers hold, a single seven-hour overnight window on one FM frequency can clear the backlog of about 30 active users, and scaling to 300 users requires roughly ten frequencies, suggesting capacity can grow by adding channels rather than redesigning the system.","Because FM is broadcast, content requested by one user reaches every tuned-in phone; the knowledge hub turns this into a free discovery channel and preserves downlink anonymity, since the server cannot know who is listening.","LLM responses arrive faster than full webpages because they are smaller, but a single lost 500-byte frame can break an entire response, whereas webpage images tolerate partial loss; completion rates therefore depend on content type and signal strength.","Users with median signal strength above -70 dBm consistently exceeded 80% request completion, while users below -90 dBm often fell below 30%, so receiver placement and antenna conditions determine whether the service is usable.","Once the station and SMS gateway exist, serving additional listeners adds almost no marginal cost, making advertising-supported or SMS-premium business models plausible."],"supporting_citations":[{"why":"Supplies the prior SONIC prototype, including the pixel-interpolation user study and the initial FM/SMS webpage-delivery approach that this deployment extends.","marker":"[57]"},{"why":"Provides the open-source audio modem library that performs the OFDM modulation and demodulation of SONIC files.","marker":"[13]"},{"why":"Is the Android port of that modem library, modified here to decode audio bytes coming from the FM tuner instead of the microphone.","marker":"[6]"},{"why":"Introduces click maps for making static page screenshots interactive, the mechanism SONIC uses for hyperlink navigation over SMS.","marker":"[24]"},{"why":"Demonstrates FM broadcast data services whose quality degrades with signal strength, motivating the image-based content that stays interpretable under loss.","marker":"[25, 26]"},{"why":"Provides the open-source Android distribution modified to give third-party apps access to FM chip audio without root privileges.","marker":"[50]"}],"fun_headline_variants":["FM radio + SMS: offline web for 2.6 billion","SONIC: 10 kbps web via FM, SMS request channel","Cameroon six weeks: FM radio delivers web and AI","Turn FM radios into data links for web access","Low-cost web: FM broadcast down, SMS up"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system assumes the radio station's audio chain, including the mixer, transmitter audio input, and the phone's FM receiver, carries the OFDM-modulated audio with little distortion, and that station staff reliably switch to the data broadcast at 10 PM; the paper reports days were lost when the switch was missed or the transmitter volume drifted from 100%.","fun_headline_variants_meta":{"raw":{"variants":["FM radio + SMS: offline web for 2.6 billion","SONIC: 10 kbps web via FM, SMS request channel","Cameroon six weeks: FM radio delivers web and AI","Turn FM radios into data links for web access","Low-cost web: FM broadcast down, SMS up"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000578,"raw_usage":{"total_tokens":2716,"prompt_tokens":927,"completion_tokens":1789,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":1706}},"tokens_in":543,"tokens_out":1789,"duration_ms":12803,"temperature":1.0,"reasoning_tokens":1706,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:58:59.278526+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same nightly broadcast for one week at a second FM station, measuring per-frame loss against signal strength; if the majority of receptions above -90 dBm exceed 20% loss, or if any single day's transmission fails because station staff did not flip the switch, the paper's reliability claim does not generalize.","supporting_citations":[{"cited_title":"https://github.com/ quiet/quiet","cited_arxiv_id":null,"evidence_quote":"Provides the open-source audio modem library that performs the OFDM modulation and demodulation of SONIC files."},{"cited_title":"https://github.com/quiet/ org.quietmodem.Quiet","cited_arxiv_id":null,"evidence_quote":"Is the Android port of that modem library, modified here to decode audio bytes coming from the FM tuner instead of the microphone."},{"cited_title":"DRIVESHAFT: Improving Perceived Mobile Web Performance","cited_arxiv_id":"1809.09292","evidence_quote":"Introduces click maps for making static page screenshots interactive, the mechanism SONIC uses for hyperlink navigation over SMS."},{"cited_title":"My traceroute (MTR)","cited_arxiv_id":null,"evidence_quote":"Provides the open-source Android distribution modified to give third-party apps access to FM chip audio without root privileges."}],"review_version":1}