{"id":"51edbf4f-59f5-4500-8435-0f3cb1c0a4ff","arxiv_id":"2606.17245","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Without the internet, phones running CttF can still retrieve cached webpages from nearby phones using saved ratings, proxy signatures, and proofs-of-work, with simulated city-scale success.","lead":"A new phone-to-phone system caches webpages before an internet blackout and shares them over Bluetooth once the internet is gone. If it works as simulated, people under shutdowns could retrieve popular reference pages within a day instead of only exchanging short messages.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline numbers depend on assuming community ratings are a noisy estimate of exactly the 'usefulness' metric used for evaluation; §5.3 generates both from the same Zipf ground truth, so the caching signal and the success metric are coupled.","rationale":"The reader's weakest assumption is the mobility/contact model. That concern is legitimate: 500m cells are larger than Bluetooth range, and the 100% per-timestep contact probability is not directly observed. However, it is partially bounded by the paper's own sensitivity sweep: even at 5% contact, one-week satisfaction is 52% (Fig. 4a), and longer horizons are not shown but may converge higher. The rating/usefulness coupling is less well bounded and more fundamental. CttF's novelty is crowdsourced rating-based replication; if ratings do not track usefulness, the mechanism has no signal. The simulation generates ratings from the same Zipf distribution used to define the 'most useful pages,' so the evaluation cannot distinguish the system's ability to aggregate consensus from the assumption that consensus exists. The paper flags this when it admits the correlation may not hold in a blackout (§5.3), but no experiment varies it. An independent sensitivity run with user-specific rating models would settle whether the headline numbers are robust. Meanwhile, the prototype microbenchmarks support the throughput assumptions, so this is not an internal contradiction; it is an external validity risk. The verdict should remain CONDITIONAL.","tokens_in":17271,"tokens_out":12003,"duration_ms":134937,"concrete_test":"Run the §6.1 benign scenario with the same simulator and parameters but replace the rating-generation module. Control: current global Zipf ground truth + 0.5 noise. Treatment: each seeder rates 500 pages drawn from a per-user topical model (e.g., sample a Dirichlet(0.1) over page categories and rate pages within that user's top categories), while satisfaction is still measured against the global top-100k list. If top-10k satisfaction falls by more than ~15 percentage points relative to control, the headline numbers depend on the assumed rating/usefulness correlation. A stronger variant uses a real browsing-history dataset (e.g., Ruth et al. clickstreams) to define ground-truth utility and derives noisy ratings from it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Headline numbers are driven by an unvalidated coupling between the ratings that drive caching and the ground truth used to measure satisfaction. In §5.3 the simulator assumes a global Zipf ground-truth distribution for page accesses, maps it to ratings via f(x)=10(a·Zipf(x)^b+c) fit to Ruth et al.'s hit shares, and draws each seeder's ratings from that ground truth plus 0.5 noise. Satisfaction in Fig. 3c is then reported for 'page indices sorted from most to least useful' — the same ordering that generated the ratings. The caching mechanism is therefore evaluated in a world where community ratings are a noisy but unbiased estimate of exactly the 'usefulness' metric being measured. The paper itself notes in §5.3 that no work correlates webpage ratings with popularity and that 'such a correlation may not hold in a blackout.' If real ratings are user-specific, topically biased, or only weakly correlated with what a population needs during a blackout, the distributed-rating mechanism has much less signal; the 75% / top-10k / top-100k numbers in Fig. 3 would not follow. This is more central than the contact-probability assumption: at 5% contact the system still delivers something, whereas if the rating signal is absent the value proposition collapses.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Cache to the Future (CttF) proposes a system for providing access to cached webpages during internet blackouts. In a pre-blackout phase, users cache and rate pages; ratings are exchanged over Bluetooth and used to prioritize background replication. A trusted proxy signs page contents, and proofs-of-work plus local averaging are used to limit Sybil influence. During a blackout, users request pages opportunistically from nearby peers. The paper validates the design through an Android prototype and simulations based on the YJMob100K mobility traces (25,000 users, 75 days, 500m cells). The headline results are roughly 75% request satisfaction over a 60-day blackout for the 100,000 most useful pages, sub-day median latency for the top 10,000 pages, and only a 5% satisfaction drop with 25% Sybils. The authors also compare against epidemic routing and evaluate jamming, cache-manipulation, and proof-of-work costs.","tokens_in":17602,"tokens_out":7827,"duration_ms":82813,"significance":"The work addresses a real and underexplored design point: unlike messaging-focused blackout tools, CttF targets web knowledge access. If the quantitative claims survive scrutiny, the paper would be a useful step toward deployable blackout-resistant content delivery. Strengths include the use of large real mobility traces, direct prototype throughput and battery measurements, and an explicit sensitivity analysis for contact probability. The comparison with epidemic routing is valuable, as is the adversarial evaluation. However, the evaluation's central coupling between the rating signal and the success metric, together with the optimistic baseline contact assumption, mean that the current numbers should be read as upper bounds rather than demonstrated performance.","major_comments":[{"comment":"The evaluation is circular with respect to the rating signal. The simulator defines a 'ground-truth Zipf-like distribution for page accesses', generates every seeder's ratings as noisy observations of that same ground truth ('Each user's ratings follow the global ground-truth ratings with additive exponentially decaying normal noise'), and then reports satisfaction for 'Page Indices Sorted from Most to Least Useful' — i.e., the same ordering. The paper itself states (§4.3) that no work correlates webpage ratings with popularity and that 'such a correlation may not hold in a blackout.' Thus the simulation only exercises the regime where community ratings are an unbiased, monotone proxy for exactly what the population will request. If real-world ratings are user-specific, topically biased, or decoupled from blackout-time demand, the caching signal weakens and the 75% / top-10k / top-100k n","section":"§5.3 and Fig. 3c"},{"comment":"The baseline contact model is an upper bound. The YJMob100K traces supply positions every 30 minutes in 500m cells; two users in the same cell are assumed to connect with 100% probability per timestep and exchange up to four 2.5MB pages in a one-minute session. Bluetooth range is typically tens of meters, so co-location in a 500m cell does not imply encounter. The traces do not contain Bluetooth encounter observations, so the 100% value is not derived from data. The paper's own sensitivity analysis shows one-week satisfaction drops from 73% to 52% at 5% contact, with larger latency penalties for mid/low utility pages (Fig. 4b/c). The 60-day 75% headline therefore depends on an unvalidated optimistic parameter. Please justify the 100% baseline with empirical encounter-rate data (e.g., from Bluetooth/BLE encounter studies) or present the sensitivity sweep as the primary result.","section":"§5.2, §6.1, Fig. 4"},{"comment":"The comparison against epidemic routing does not equalize resources. CttF nodes store 1200 pages (~3GB) and use pre-blackout community ratings to populate leecher caches; the epidemic-routing baseline stores only 300 forwarding pages (~750MB) and no leecher-side cache. In addition, the adversary is described as making up to 512 spam requests per interaction, which is hard to reconcile with the 4-page forwarding limit used elsewhere. Please equalize the storage and request budgets (or explicitly justify the 4x difference) and specify the per-interaction message budget for spam. Otherwise the claimed advantage over epidemic routing may be an artifact of the resource allocation.","section":"§6.4 and Table 2"}],"minor_comments":[{"comment":"The statement 'page access trends follow a Zipf-like distribution [34]' cites a paper on Tor flow correlation attacks; this appears to be the wrong reference. Please verify and correct.","section":"§4.3"},{"comment":"The x-axis ranges in Fig. 8a/b are difficult to parse ('-400, -200, 0, +1e6'). Please provide a clear axis for page indices with consistent scaling.","section":"Fig. 8"},{"comment":"The legend defines '✓: Only somewhat supported' but uses the same symbol as the full checkmark in the column headings. Make the partial marker visually distinct (e.g., half-filled) to avoid ambiguity.","section":"Table 1"},{"comment":"The relationship between the two-month simulation (Fig. 3, 82% by end) and the one-week follow-up simulations (Fig. 4, 73%) should be explained earlier. As written, the reader may see an inconsistency between the headline 75% and the 73% figure.","section":"§6.1"},{"comment":"The abstract's 'no technology currently exists' is stronger than the related-work section supports (Ceno, Kiwix, Collateral Freedom). Suggest softening to 'no widely deployed system' or similar.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"I agree with the stress-test concern: the rating/ground-truth coupling in §5.3 is the central weakness and should be addressed with a decoupling experiment. The paper is otherwise well-presented, and the sensitivity analyses are useful. If the authors add a sensitivity sweep that breaks the coupling between ratings and demand, and if they calibrate or de-emphasize the 100% contact baseline, I would support acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CttF is a sensible design point that has no direct prior art: a blackout-resistant system for web content rather than messaging, combining proxy-signed pre-blackout caching, locally exchanged ratings with PoW, and non-epidemic request/response. The paper is well-structured, and the survey (Table 1) is genuinely useful for positioning. The prototype microbenchmarks are a plus: they validate the throughput assumptions, and the PoW difficulty measurements on two real phones are concrete.\n\nThe simulation work is thoughtful: real mobility traces (YJMob100K), 25k users, city-scale grid, and they vary contact probability and adversarial fractions. The main soft spot is the one the stress-test note flags. Section 5.3 fits a transform f(x)=10(a·Zipf(x)^b+c) to Ruth et al.'s hit shares to generate ground-truth ratings, and then evaluates satisfaction for 'page indices sorted from most to least useful' using that same ordering. The ratings that drive caching are a noisy version of the exact ground truth used for the success metric. So the evaluation assumes community ratings are an unbiased estimate of what users will actually need during a blackout. The paper itself admits no work correlates ratings with popularity and that the correlation 'may not hold in a blackout.' If real ratings are biased or idiosyncratic, the 75% satisfaction / top-100k numbers in Fig. 3 are not supported. That's a load-bearing gap, more central than the contact probability issue.\n\nThe contact probability is a secondary concern but real: the 100% within-cell default is optimistic, and at 5% contact the one-week satisfaction drops to 52%. That doesn't sink the system — it still delivers something — but the headline should come with the caveat.\n\nOther issues: no simulator code or data released, no variance/seed reporting, so the exact numbers are not independently reproducible from the paper alone. The epidemic routing comparison is a bit stylized (adversaries spam 512 requests), but it does illustrate the bandwidth problem.\n\nWho is this for? Researchers working on blackout-resistant systems, censorship circumvention, and mobile mesh networking. The design space is real, and the paper deserves a serious referee. The evaluation needs major revisions, especially decoupling the rating ground truth from the success metric, reporting confidence intervals, and releasing artifacts. I'd engage with it.","headline":"A genuinely new design point for blackout-resistant web access, but the headline numbers rest on an unvalidated coupling between community ratings and the usefulness metric.","tokens_in":18098,"tokens_out":1689,"would_cite":true,"duration_ms":17766,"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":"A Bluetooth-based distributed archive can keep the web's most useful pages reachable through a two-month internet blackout.","keywords":["internet blackouts","blackout-resistant technologies","mobile mesh networks","distributed caching","content availability","Bluetooth communication","proof-of-work","censorship resistance"],"falsifier":"Collect Bluetooth encounter traces from a few thousand city residents over several weeks (e.g., via a research app logging BLE beacons) and count how often two phones in the same 500m cell establish a session and how many megabytes they exchange per session. If the observed same-cell contact probability is near 5% rather than 100%, or if median sessions move less than one 2.5MB page, the simulation's headline satisfaction and latency numbers would not hold in the field.","tokens_in":17141,"feed_emoji":"📡","tokens_out":2803,"duration_ms":34332,"temperature":0.7,"pith_summary":"The paper argues that no existing blackout-resistant technology can give citizens continued access to web-based knowledge, and presents Cache to the Future (CttF) as the first system designed to fill that gap. CttF caches static webpages on user phones before a blackout, spreads community ratings over Bluetooth to decide which pages to replicate, and lets users fetch pages directly from nearby peers during the blackout. The claim is that crowdsourced ratings plus cryptographic safeguards (digital signatures for provenance, proof-of-work to curb Sybil rating attacks) make this work at city scale. If correct, residents in blackout-prone regions could keep reading Wikipedia, medical pages, and community resources offline, with the most useful content arriving within hours. The simulations suggest roughly 75% of requests for the 100,000 most useful pages are satisfied over a two-month outage, with only a 5% satisfaction drop even when a quarter of nodes are adversarial.","feed_headline":"Bluetooth mesh serves web pages through a two-month blackout","feed_subtitle":"Simulations on real city mobility show ~75% of the 100,000 most useful pages stay reachable, with top pages arriving in under a day.","key_machinery":"The load-bearing mechanism is the combination of locally averaged community ratings and rating-driven Zipf sampling: each user maintains per-page average ratings learned only from direct Bluetooth exchanges, and background caching samples pages with probability proportional to a transformed Zipf distribution, so popularity begets replication. Proof-of-work (SHA-256 prefix) gates rating exchanges to slow Sybil inflation, digital signatures from a trusted proxy provide verifiable page provenance, and a four-page-per-session forwarding limit models realistic Bluetooth throughput. The simulator's contact model—two users in the same 500m grid cell connect with configured probability for a one-min","core_discovery":"CttF is a distributed webpage archive that operates in two phases. Pre-blackout, users cache pages fetched through a trusted proxy that signs them, rate pages on a 1–10 scale, and exchange those ratings (and cached URLs) over Bluetooth; a locally maintained community rating per page is updated with each exchange, and a proof-of-work limits how often a MAC-spoofing adversary can re-submit ratings. Background caching samples pages proportionally to a Zipf-transformed rating distribution, so highly rated pages are replicated on more devices. During a blackout, users request pages and receive them from co-located peers over Bluetooth, verifying signatures when available; when signatures are abse","pith_inferences":["Editorial inference: The headline numbers depend on the assumption that two phones in the same 500m cell reliably connect for a minute every 30 minutes; the paper's own 5%-contact experiment cuts one-week satisfaction from 73% to 52%, so real-world Bluetooth encounter rates are the first thing to measure before trusting city-scale claims.","Editorial inference: Because satisfaction collapses from 100% to 13.5% for pages beyond the top 100,000, the system is strong for common knowledge but weak for exactly the obscure pages that censorship victims may need; a deployment focused on high-value local content would require seeding rather than relying on ratings alone.","Editorial inference: The paper's assumption that seeders rate pages by genuine usefulness may be optimistic; when as few as 5% of nodes are coordinated Sybils, over half of leechers cache useless pages, so the proof-of-work mitigation matters mainly for stalking adversaries, not for distributed rating pollution.","Editorial inference: A testable extension is to run a small field pilot where volunteers carry CttF-like devices through a city for a week and log actual session lengths and pages exchanged; those data would either confirm the contact model or force a re-parameterization of the simulations."],"forward_implications":["If the central claim holds, communities facing blackouts could pre-load a small app and accumulate a distributed archive within days of normal browsing, without dedicated infrastructure.","The top 10,000 most useful pages would typically arrive in under a day even under jamming of up to 250 km², meaning reference and emergency information stays fresh.","Sybil attacks are blunted: even at 25% adversarial nodes, request satisfaction falls only about 5% and latency rises modestly, suggesting rating manipulation cannot easily starve the system.","Epidemic routing is not required; direct user-to-user exchanges over sparse Bluetooth encounters suffice, which lowers bandwidth and battery overhead relative to messaging-first mesh apps.","The design extends blackout-resistant technology from texting to knowledge access, complementing existing tools that only carry short messages."],"fun_headline_variants":["Bluetooth mesh serves web pages during extended blackouts","Crowdsourced caching keeps top pages alive in outages","Distributed web archive uses peer ratings via Bluetooth","City-scale simulations show 75% of key pages survive blackout"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that two users in the same 500-meter grid cell will actually connect over Bluetooth for a full minute of page exchange every 30-minute timestep; the paper's own sparse-contact experiment shows that if real encounter rates are closer to 5% than 100%, one-week satisfaction drops from 73% to 52%.","fun_headline_variants_meta":{"raw":{"variants":["Bluetooth mesh serves web pages during extended blackouts","Crowdsourced caching keeps top pages alive in outages","Distributed web archive uses peer ratings via Bluetooth","City-scale simulations show 75% of key pages survive blackout"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000903,"raw_usage":{"total_tokens":3683,"prompt_tokens":663,"completion_tokens":3020,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":407,"completion_tokens_details":{"reasoning_tokens":2954}},"tokens_in":407,"tokens_out":3020,"duration_ms":22881,"temperature":1.0,"reasoning_tokens":2954,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T11:04:13.658319+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect Bluetooth encounter traces from a few thousand city residents over several weeks (e.g., via a research app logging BLE beacons) and count how often two phones in the same 500m cell establish a session and how many megabytes they exchange per session. If the observed same-cell contact probability is near 5% rather than 100%, or if median sessions move less than one 2.5MB page, the simulation's headline satisfaction and latency numbers would not hold in the field.","supporting_citations":[],"review_version":2}