arXiv CSopen access

The Towers Were Standing: A Cause Decomposition of Cellular Outages During Hurricane Helene

The Towers Were Standing: A Cause Decomposition of Cellular Outages During Hurricane Helene

arXiv:2609.10944v1 [cs.NI] 10 Sep 2026

Oluseyi Olukolaa,∗, Oare Danielle Addehb , Esther Abiodun Konanb , Nick Rahimia a

School of Computing Sciences and Computer Engineering, University of Southern Mississippi, Hattiesburg, MS, 39406, United States b Department of Communication, College of Humanities and Social Sciences, George Mason University, Fairfax, VA, 22030, United States

Abstract Hurricane Helene produced the largest absolute cell-site outage in the public FCC record, peaking at 4562 sites. The conventional model is physical: towers destroyed. Helene did destroy over 1700 miles of fibre, but almost none of it was cell sites. We present the first cause-decomposed study of the FCC’s Disaster Information Reporting System, reconstructing 80 state-days and 580 county-days from 24 daily filings by two reconciled independent extractions. Damage to cell sites is negligible: 1.1 % of attributed cell-site-days across six states, at most 3.8 % anywhere. The sites were standing. What took them out divides by terrain: pooled, power dominates at 63.2 %, but in mountainous North Carolina severed transport (backhaul) reaches 52.2 % against 47.3 %, and in Tennessee 69.9 %. North Carolina’s transport share rises from 7.0 % to 85.0 % across the event (ρ = 0.92). Seventeen days after landfall, on 15 October, 47 sites lost transport across six contiguous North Carolina counties with no rainfall, no power loss, no damage, and recovery by the next report. Independent active-probe measurement corroborates it: responsive /24s fall 1.02 % for twelve hours while Tennessee stays flat. We release the dataset. Backup power is the standard resilience investment; here it addresses the smaller half of the problem. ∗

Corresponding author. Email addresses: [email protected] (Oluseyi Olukola), [email protected] (Oare Danielle Addeh), [email protected] (Esther Abiodun Konan), [email protected] (Nick Rahimi)

Keywords: network outages, disaster resilience, cellular networks, backhaul, network measurement, FCC DIRS, telecommunications regulation 1. Introduction On 26 September 2024 Hurricane Helene made landfall in Florida and tracked inland across Georgia, the Carolinas, Tennessee and Virginia. In the mountains of western North Carolina it produced catastrophic flooding, and with it the largest absolute cell-site outage in the public DIRS record: 4562 cell sites out of service on 28 September, across the six states reporting that day (FL, GA, NC, SC, TN, VA) (Federal Communications Commission, 2024).1 Over 1700 miles of fibre-optic cable were destroyed, 19 North Carolina counties were left technologically isolated, and, on 28 September, 17 Public Safety Answering Points lost the ability to receive 911 calls (Hauser, 2025). Clinicians at a rural hospital in Henderson County, one of the six counties in the October event we isolate in §5.4, describe treating patients without electronic records, laboratory systems or their usual communication tools for the duration (Gamboa et al., 2026). The public and engineering narrative of such an event is physical: towers fall, equipment floods, and restoration is understood as reconstruction. That framing shapes what gets funded, what gets hardened, and what researchers build: if the problem is destruction, the answer is sturdier infrastructure and faster rebuilding. What this paper does and does not claim. We do not claim Helene destroyed little. It destroyed a great deal, and the 1700 miles of severed fibre are 1

We say “largest in the DIRS record” rather than “largest ever” because the comparison does not extend further. Hurricane Irma reached 4370 sites out on 11 September 2017, only 4 % below Helene (Federal Communications Commission, 2017a), and for Superstorm Sandy the Commission published only that more than a quarter of cell sites in 158 counties across ten states and the District of Columbia were disabled at peak, with no absolute count (Federal Communications Commission, 2013). DIRS counts also scale with the county set the FCC designates, which differs by event: Helene’s reporting area held 37,717 sites (12.1 % out) against Irma’s 16,352 (26.7 %). Proportionally the largest DIRS outage remains Hurricane Maria, at 95.6 % of Puerto Rico’s 1789 sites on 23 September 2017 (Federal Communications Commission, 2017b). For scale within the mainland record, Hurricane Ida peaked at 1824 sites out across Alabama, Louisiana and Mississippi on 30 August 2021 (Federal Communications Commission, 2021).

2

themselves physical damage of the most literal kind. Our claim is about where on the network graph that damage landed, and it is a claim the reporting instrument makes precise. DIRS asks providers to attribute each out-ofservice cell site to damage at the site, to loss of transport (the backhaul linking that site to the network core), or to loss of commercial power. Read in those terms, the answer is unambiguous: across six states, damage at the site accounts for 1.1 % of attributed cell-site-days, and never exceeds 3.8 % in any state. The radio access equipment was overwhelmingly intact; what had failed, in the mountains, was the transport link between that equipment and the network core. That distinction is not bookkeeping. A site down for lack of power and a site down for lack of backhaul present identically to a subscriber, but they are different engineering problems, with different restoration timelines, different mitigations, and (as we show) different temporal behaviour. Backup power, the standard hardening investment, addresses only one of them. Scope. The balance between the two shifts with terrain, and the paper scopes its claims accordingly. Pooled across all six reporting states, commercial power remains the larger cause (63.2 % of attributed site-days against 35.7 % transport). It is in the mountainous inland states (North Carolina and Tennessee, where fibre follows a small number of valley routes) that transport overtakes it. The finding that generalises across our whole sample is the negligibility of damage at the site; the transport-dominance finding is scoped to constrained-route terrain, which is also where the outage was largest and longest. 1.1. Background: the measurement instrument DIRS is a voluntary reporting system the FCC activates for major disasters (Federal Communications Commission, 2026a). Communications providers file daily status for a defined disaster area; the Public Safety and Homeland Security Bureau publishes a daily aggregate as a PDF, stating the instant its data describe (09:00 EDT on the report date, for the Helene activation). Crucially for this work, the published tables break each day’s cell-site outages into three causes (sites out due to damage, due to transport, and due to power ) plus a count of sites operating on backup power. This decomposition is the measurement instrument we exploit. Prior work has used FCC hurricane reports as a source of outage counts (§2); to our knowledge the cause columns have not previously been analysed. 3

DIRS’s mandatory sibling, the Network Outage Reporting System (NORS), collects richer per-incident detail but is presumed confidential under 47 C.F.R. §4.2 (of Federal Regulations, 2026; Federal Communications Commission, 2026b), and is therefore unavailable for research without a confidentiality process (claffy and Clark, 2022). The Commission has recently modernised the regime. Its Third Report and Order on Resilient Networks, adopted 20 May 2026, expands the ability of DIRS filers to submit geospatial information voluntarily, including cellsite locations, and harmonises transport-facility reporting to optical-carrier circuits (Federal Communications Commission, 2026c,d). Our results speak directly to that proceeding. 1.2. Contributions • A dataset and a reconciliation method (§3). A per-cause, percounty daily record of Helene’s cellular outage (80 state-days and 580 county-days), reconciled by the cross-table method of §3.2. • Cause decomposition (§5.1). Damage at the site is negligible in all six states; transport exceeds power only in the two mountainous ones. We report both denominators and the unattributed residual. • The temporal inversion (§5.2). North Carolina’s transport share rises near-monotonically across the event (ρ = 0.92), reaching 80.9 % of attributed outages in the recovery tail. We test the trend against both serial correlation and the mid-event rescoping of the reporting area. • An isolating event, externally corroborated (§5.4). A 15 October transport failure across six contiguous counties, with weather, power, damage and rescoping eliminated against primary sources, and the timing confirmed by independent active-probe measurement. 2. Related Work Outage measurement from the outside. A substantial literature infers outages from active and passive Internet measurement. Dainotti et al. characterised politically-motivated outages from BGP, network-telescope traffic and traceroute (Dainotti et al., 2011, 2014); Trinocular adaptively probes a sample of 4

addresses within each responsive /24 and infers block-level reachability outages (Quan et al., 2013); and Bischof et al. characterised Internet outages and shutdowns at scale (Bischof et al., 2023). These approaches can attribute cause, but only at a coarse and largely political granularity: Bischof et al. separate government-ordered shutdowns from spontaneous outages, and Dainotti et al. identify censorship. None reports the operating cause an operator would file (damage, transport or power) and, for our purposes, the platforms are least able to see exactly the networks we care about: Bischof et al. note that active probing of publicly routable IPv4 space has limited visibility into NAT-heavy mobile networks. Our work is complementary in both directions: DIRS reports operating cause directly for cellular, at the price of covering only declared disasters and only participating providers, and we use an external platform to corroborate DIRS in §5.5. Disasters and physical infrastructure. Cho et al. measured ISP impact of the 2011 Tōhoku earthquake (Cho et al., 2011); Heidemann et al. analysed Hurricane Sandy (Heidemann et al., 2012); Padmanabhan et al. linked weather events to residential link failures at scale using ThunderPing (Padmanabhan et al., 2019); Durairajan et al. assessed long-run sea-level-rise risk to US Internet infrastructure (Durairajan et al., 2018). Maitland and Peha studied Puerto Rico’s post-Maria restoration from a policy standpoint (Maitland and Peha, 2018). None decomposes a cellular outage record by reported cause. Transport and power are occasionally separated in passing (Cho et al. describe severed backbone circuits and, separately, a data centre losing mains power) but qualitatively, for a single operator, and not as a decomposition of the outage. Cellular resilience engineering. Booker et al. (2010) and Griffith et al. (2015) model base-station survival under hurricanes; Malandrino and Chiasserini consider disaster impact on wireless communication (Malandrino and Chiasserini, 2017). Closest to our finding, Yang et al. analyse backhaul failure and resilient backhaul design using operator-internal data (Yang et al., 2017). This literature largely models physical survival, which our results suggest is the smaller problem. Backhaul is not absent from it (Booker et al. simulate the resilience benefit of fibre-ring and mesh topologies) but it is reached either by simulation or, in Yang et al.’s case, from proprietary operational data unavailable to the community, and in neither case as an observed share of a real outage. 5

FCC outage data outside networking. FCC outage reporting has been used before, but as a source of outage counts rather than causes, and largely outside the networking literature. Reed and Wang built cell-tower fragility and recovery curves for Hurricanes Harvey and Irma from FCC cell-tower outage reports alongside utility and Department of Energy data (Reed and Wang, 2018). Du models county-level telecommunications outages across ten hurricanes and eight states, using outage magnitude as the prediction target and power outage and demographics as predictors (Du, 2024).2 Kuhn analysed the pre-NORS FCC outage-reporting regime for the public switched telephone network, and did decompose those outages by cause (into human error, natural events, hardware, software, overload and vandalism) though that taxonomy has no transport or commercial-power category and does not concern cellular (Kuhn, 1997). claffy and Clark survey NORS and DIRS as public-interest measurement instruments and document their access restrictions (claffy and Clark, 2022). Recent Congressional analysis describes both systems, the restoration of cellular service after Helene, and their data limitations (Service, 2025; Gallagher, 2024). In all of this work, an outage is a scalar. The cause columns are an unused resource. Feeny et al. illustrate this. They reconstruct cellular coverage under natural hazards from public data and, needing failure causes, model structural damage and power loss from first principles, scoping backhaul out of the analysis (Feeny et al., 2026). That is a reasonable design given the sources they draw on. It also indicates that a 2026 paper reaching for cause-resolved cellular outage did not have the DIRS decomposition in view, which is better evidence that the columns are unexploited than an absence of citations could give. We therefore make the narrow claim, and only it: this is the first study to exploit the DIRS cause decomposition, and hence the first to separate transport from power in a public cellular outage record. We do not claim to be the first to use DIRS. 2

Du does not name the outage source in the abstract and the article is paywalled; we therefore describe what it models rather than assert which FCC series it draws on.

6

3. Methodology Figure 1 summarises the pipeline. Its central feature is the fork: two extractions are performed independently from the same source documents and compared only afterwards, so agreement between them is evidence rather than tautology. E1 decomposition

FCC DIRS 24 daily reports 26 Sep– 19 Oct

independent

Extraction A state totals + cause columns

Extraction B

Reconcile P

county = state 45/45 days exact

county totals per state cross-validation, not self-consistency

Dataset 80 state-days 580 countydays

cause shares

E2 temporal trend, phases

E3 15 Oct event alternative elimination

Figure 1: Measurement pipeline. Two independent extractions (one targeting state-level tables with their cause columns, one targeting county-level tables) are reconciled against each other before any analysis. The two read disjoint tables, so the reconciliation is a cross-table identity rather than a repeat of the same read. Agreement is exact on all 45 reconcilable state-days: 23 in North Carolina and 22 in Tennessee.

3.1. Data acquisition We retrieved every published DIRS Communications Status Report for Helene directly from the Commission’s document server (Federal Communications Commission, 2024). Twenty-four reports exist, one for each day from 26 September to 19 October 2024 inclusive, including the weekend of 12–13 October. The 26 September report predates state-level cause reporting and is used only for county baselines, giving 23 days of analysable cause data. Each report states the instant it describes (09:00 EDT on the report date), so the series is a sequence of point samples at a fixed daily offset, not a set of daily aggregates. §5.5 depends on that. Table 1 lists the fields extracted from each report. 3.2. Extraction and reconciliation The FCC publishes these tables as PDFs whose text layer collapses adjacent numeric columns, so naive extraction is unreliable. We therefore performed two independent extractions (Extraction A targeting state-level tables and their cause columns, Extraction B targeting county-level tables) and cross-validated them. 7

Table 1: Fields extracted from each daily DIRS report. The three cause columns are the measurement instrument this study depends on.

Level

Field

Used for

State State State State State County Area Area

sites served, sites out out due to damage out due to transport out due to power sites on backup power sites served, sites out subscribers out (cable/wireline) PSAP status

denominators E1, E2, E3 E1, E2, E3 E1, E2, E3 E3 reconciliation, E3 E3 context

The independence that matters here is not two runs of the same parser. A and B read different tables, on different pages, with different column layouts, and the check between them is an accounting identity that the source documents never print: for every day, the sum of the county rows must equal the independently extracted state row. A shared tooling bug would have to produce compensating errors in two differently-shaped tables to survive that test. The identity holds on every day it can be checked. Across all 45 statedays for which both a county table and an independently extracted state row exist (23 in North Carolina and 22 in Tennessee), the sum of the county rows reproduces the state totals exactly, in both sites-served and sites-out. The one county table that cannot be checked this way is North Carolina’s on 26 September, whose report predates state-level cause reporting; it is excluded rather than counted. Where a row’s text layer was corrupt, the value was recovered by solving against the report’s own printed percentage and column totals, and accepted only when the solution was unique; one such row (Buncombe, 29 September) occurred and was resolved and verified two ways. No value is interpolated, and days without a published report are absent rather than filled. 3.3. Dataset composition Table 2 gives the resulting dataset. 3.4. Definitions and derived quantities Cell-site-day. One site out of service for one reporting day. We integrate over the event rather than reporting a single snapshot, because snapshots 8

Table 2: Composition of the reconstructed dataset. Panels are unbalanced because states and counties enter and leave the FCC’s disaster area as the event evolves.

Level

Scope

State County County

FL, GA, NC, SC, TN, VA North Carolina Tennessee

Days

Units

Records

23 24 22

6 states 21 counties 8 counties

80 420 160

Total county-days

580

are dominated by whichever day is chosen. Two denominators, both reported. The three cause counts do not sum to the printed outage total. Across North Carolina they account for 6373 of 6845 out-of-service site-days, leaving 472 (6.9 %) unattributed; pooled across six states the residual is 7.6 %. We therefore define the transport share against attributed causes, stransport =

ntransport , ndamage + ntransport + npower

and report the alternative (transport as a fraction of all out-of-service sitedays) alongside it wherever a headline number appears. For North Carolina the transport share is 52.2 % of attributed causes and 48.6 % of all out-ofservice site-days; these are the same quantity under two denominators, and should not be confused with power’s 47.3 %, which is a different cause under the first of them. The choice of denominator changes no ordering, because both causes are scaled by the same residual: transport exceeds power either way (52.2 % vs 47.3 % attributed; 48.6 % vs 44.0 % of all out-of-service sitedays). Within-state shares, because the panel moves. The FCC rescopes the disaster area as the event evolves, and it does so twice in ways that matter here: Georgia and South Carolina leave after 7 October, and 12 of North Carolina’s 21 counties leave after 12 October, dropping its served count from 1452 sites to 901. Absolute counts are therefore not comparable across those boundaries: a naive reading of the raw subscriber series would show an overnight fall from 325,348 to 84,085 on 8 October, which is Georgia and South Carolina leaving, not recovery. 9

We consequently compute every trend on within-state shares, never on absolute counts, and we never compare a count across a rescoping boundary. That convention removes the arithmetic problem but not the compositional one: if the departing counties differed systematically in cause mix, a withinstate share could still move for the wrong reason. §5.3 bounds that possibility directly, and it is the reason the tail-phase results are trustworthy. 4. Experiment We run four analyses on the reconstructed dataset. E1: cause decomposition. Integrate cell-site-days by cause over the full record, per state and pooled. Tests whether the outage was predominantly physical, and whether the answer depends on terrain. Reported in §5.1. E2: temporal behaviour. Compute the daily transport share and test it for trend, correcting for the serial correlation a monotone recovery induces, and separately for the mid-event change in the reporting panel. Then aggregate into event phases. Reported in §5.2 and §5.3. E3: the 15 October event. An anomaly detected in E2: the North Carolina outage count rose mid-recovery. We test each candidate explanation (rescoping, weather, power, damage, and a region-wide disruption) against an independent source, then examine the county-level geographic signature of the residual. Reported in §5.4. E4: external corroboration. DIRS is self-reported by the providers whose networks failed. We ask whether an independent instrument saw the 15 October event, using IODA’s active-probe signal with Tennessee (in the same activation, measured the same way) as a control. Reported in §5.5. 5. Results 5.1. E1: cause decomposition Table 3 gives the headline result for North Carolina under both denominators. These figures come from Extraction A of the pipeline in Figure 1; the county totals that validate them come from Extraction B. Every figure in this section satisfies the cross-table identity of §3.2. 10

Table 3: Cause decomposition of North Carolina cell-site-days out of service over all 23 reporting days, 27 September – 19 October 2024, under both denominators (§3.4). Transport exceeds power under either.

Cause

Cell-site-days

% attributed

% of all out

Transport (backhaul) Power Damage at site

3325 3012 36

52.2 % 47.3 % 0.56 %

48.6 % 44.0 % 0.53 %

Total attributed Unattributed

6373 472

n/a

93.1 % 6.9 %

Sites out

6845

Two observations follow. Damage at the site is negligible: at 0.56 % of attributed site-days, it is two orders of magnitude below either other cause. Transport exceeds power, which is the finding with engineering consequences, because backup power is the standard resilience investment for cell sites and it addresses only the power share. Battery capacity does not restore a site whose fibre has been cut. What the unattributed residual can and cannot do to this. 6.9 % of North Carolina site-days, and 7.6 % pooled, carry no cause. Reporting that number is not the same as saying what it could do, so we test both conclusions against it adversarially. The damage figure is robust in one direction only. 0.56 % is a floor, not a mean. If ambiguous filings (a site simultaneously flooded, unpowered and cut off) are more likely to be left unattributed when the true cause is damage, then damage is understated. Giving the residual entirely to damage, which is the most hostile assumption available, raises it to 7.4 % of North Carolina out-of-service site-days and 8.6 % pooled. Even that ceiling leaves damage the smallest of the three causes everywhere, so the negligibility claim survives, but it should be stated as “at most 8.6 %, and 1.1 % on the attributions actually filed” rather than as “1.1 %”. The transport-exceeds-power ordering in North Carolina is not robust to the same treatment. Transport leads power by 313 site-days; the residual is 472. Allocating more than 66 % of the residual to power reverses the ordering. Three things bear on whether that allocation is plausible. First, across the 22 North Carolina days with more than 20 sites out, the residual’s size 11

shows no relationship to the power share (r = −0.01, p = 0.96; Spearman ρ = −0.03, p = 0.91), a null result consistent with noise rather than with a causedependent filing convention. Second, the ordering holds under any allocation preserving the observed proportions, since both causes scale together. Third, Tennessee’s residual is 0.8 % and cannot move its 69.9 % transport share, so the transport-dominance finding does not rest on the state with the large residual. None of this excludes a power-skewed residual in North Carolina; it establishes that we see no sign of one. How far this generalises. Table 4 extends the decomposition to all six reporting states and gives the pooled figure, which we report because the North Carolina result alone would misrepresent the event. Table 4: Cause decomposition by state, with the six-state pooled total. Shares are of attributed causes; the last column gives transport as a fraction of all out-of-service sitedays, and the residual column the unattributed remainder. Damage never exceeds 3.8 %. Transport dominance over the whole record is confined to the two mountainous inland states, but the shift toward transport is not (Table 5).

State

Days

Out Transp. Power Dmg Transp. (of out)

TN NC VA GA SC FL

22 23 6 11 11 7

Pooled

23 20,413

1058 6845 616 5707 4905 1282

69.9 % 52.2 % 42.1 % 30.9 % 14.4 % 14.1 %

30.1 % 47.3 % 56.9 % 68.0 % 84.3 % 82.2 %

0.0 % 0.6 % 1.0 % 1.2 % 1.3 % 3.8 %

69.4 % 48.6 % 40.6 % 28.4 % 12.9 % 13.3 %

35.7 % 63.2 % 1.1 %

33.0 %

The pooled figure is the right headline for Helene as a whole: power was the larger cause, at 63.2 % of attributed site-days against 35.7 % for transport. Georgia (5707 site-days) and South Carolina (4905) are comparable in magnitude to North Carolina (6845) and both fall on the power-dominated side. Any claim that transport is the dominant cellular failure mode in hurricanes generally is not supported by these data, and we do not make it. Two things do survive pooling. The first is the negligibility of damage at the site: no state exceeds 3.8 %, and the pooled figure is 1.1 %. Whatever took cell sites out of service during Helene, in every state, it was overwhelmingly not the destruction of the site. The second is the ordering. Tennessee and 12

North Carolina, the mountainous inland states where fibre follows a small number of valley routes, show the highest transport shares, and Florida and South Carolina the lowest. We report the ordering as an observation rather than a result: six states is a small sample, the states differ in exposure duration as well as terrain (Virginia contributes six days, North Carolina 23), and we have no independent measure of route diversity to test the mechanism against. The scoped claim we do make is that in constrained-route terrain the topological failure mode dominates, and that this terrain is where Helene’s outage was largest and longest-lived. 5.2. E2: the temporal inversion The aggregate decomposition understates the point, because the mix is not stationary. Figure 2 plots the daily transport share for North Carolina and Tennessee, the only two states whose records span the whole activation: 23 and 22 reporting days, against 11 for Georgia and South Carolina and 6 for Virginia. A daily trend test on six points is not worth running. The phase-level comparison later in this section (Table 5) covers the shorter records too, and it is where the generality of the pattern is actually tested. Appendix A gives North Carolina’s complete daily record (served, out, and all three cause columns for every reporting day), so that every point in this figure and the next can be checked against the filings. The quantity of interest is the size of the shift. North Carolina’s transport share rises from 7.0 % of attributed outages on 27 September to 85.0 % on 19 October (a twelve-fold change, at an average +2.72 percentage points per reporting day), and Tennessee moves the same way over the same window. That is a change in the composition of the outage large enough that no reasonable reading of the series misses it, and it is the result; the tests below are a check on it, not the basis for it. We report them because a monotone recovery makes consecutive days correlated by construction, so a naive significance test on 23 points would be anticonservative. Residuals about the linear trend carry a lag-1 autocorrelation of 0.22, making the 23 days worth roughly 15 independent observations. Correcting for that, and separately by a moving-block bootstrap that preserves within-block dependence, the weakest of three p-values is 1.5 × 10−6 . Appendix B gives all three with the autocorrelation diagnostics; we do not lean on them, because with a single event there is no replication and a p-value against a no-trend null is not what makes this finding useful. 13

100

panel narrows

transport share of attributed outages (%)

80 60 40 20 0

Tennessee North Carolina

27 Sep

01 Oct

06 Oct

11 Oct

16 Oct

Figure 2: Transport share of attributed cell-site outages, by day. Two states served by different provider mixes show the same inversion.

The tail phase contains the 15 October anomaly (§5.4), so it is fair to ask whether that single day inflates it. It does not: excluding 15 October, the tail is 354 transport site-days against 97 power, a share of 78.1 % rather than 80.9 %. The inversion is not confined to the mountains. The obvious objection to a two-state result is that the two states were chosen after the fact. We therefore ran the same phase decomposition on Georgia and South Carolina, the two power-dominated states with enough site-days to test. Table 5 gives the result. Every state inverts. Georgia climbs from 17.6 % to 54.2 % and crosses 50 % before its record ends; South Carolina, the most power-dominated state in the study, still triples from 5.2 % to 37.5 % in eleven days. The temporal shift from power to transport is therefore a general property of hurricane recovery, not a quirk of mountainous terrain. What terrain changes is the starting level and the endpoint. Tennessee begins its record already transport-dominated and finishes at 93.5 %; South Carolina begins at 5.2 % and, on a record that ends after eleven days, does not reach parity. This is a more useful result than the aggregate comparison 14

Table 5: Transport share of attributed outages, by state and event phase. All four states rise monotonically; what differs is the starting level and how far the record runs. Georgia and South Carolina leave the reporting area after 7 October, so their restoration phase covers 4–7 October and they have no tail; Tennessee enters on 28 September, so it has no onset. Damage never exceeds 4.5 % of attributed outages in any cell.

Onset Acute Restoration Tail State 27 Sep 28 Sep–3 Oct 4–10 Oct 11–19 Oct TN NC GA SC

n/a 7.0 % 17.6 % 5.2 %

59.8 % 47.1 % 30.6 % 12.8 %

84.4 % 58.0 % 54.2 % 37.5 %

93.5 % 80.9 % n/a n/a

in §5.1, because it separates two things the pooled figure conflates: whether the mix inverts, which it does everywhere, and how far it gets, which depends on terrain and on how long the event runs. It also weakens the reading that the coastal states are simply a different kind of event. They are the same event, observed earlier in its trajectory and over a shorter window. The mechanism is unsurprising once stated. Power is restored in days (crews reconnect feeders, generators are refuelled) while severed fibre must be physically located, accessed and respliced, often along roads and bridges that are themselves destroyed. The two causes decay at very different rates, and the slower one comes to dominate. Figure 3 shows the same data as absolute counts: the total collapses as power returns, while the transport band persists. The timing matters for what a mitigation would have to address. Professional response in the first 48 hours has dedicated satellite and radio channels and does not depend on the commercial cellular network. Civilian coordination in weeks two and three does, and that is the window in which the transport share is highest: 80.9 % of attributed North Carolina outages in the tail phase, against 47.1 % in the acute phase. A mitigation aimed at the topological failure mode is therefore aimed at the later, longer part of the outage rather than its peak. 5.3. Does the rescoping produce the trend? The tail phase straddles a change in the reporting panel: North Carolina’s served count falls from 1452 to 901 between the 12 and 13 October reports as 12 counties leave the FCC’s disaster area. If the departing counties were 15

damage power transport

NC cell sites out of service

1000 800 600 400 200 0 27 Sep

01 Oct

06 Oct

11 Oct

16 Oct

Figure 3: North Carolina cell sites out of service by cause. The damage band is not visible at this scale: it peaks at seven sites on 28 September, against 1079 out that day, and totals 36 site-days across the event (Table 3, Appendix A). That invisibility is the finding, not an omission.

disproportionately power-affected, the apparent rise in transport share would be an artefact of who left rather than a fact about recovery. The cause columns are published only at state level, so we cannot compute county cause shares directly. We can, however, bound the confound: if almost all of the outage already sat in the counties that were retained, the panel change can have moved the state share very little. On 12 October, the last day before the rescope, 91 of the 96 North Carolina sites out of service (94.8 %) were in the nine counties the panel retained the next day. The 12 departing counties held 551 of the 1452 sites served (37.9 %) but only 5.2 % of the outage, and only two of them (Burke, 3 sites; Caldwell, 2 sites) had any sites out at all. Both figures are computed on 12 October, where the nine retained counties served exactly the 901 sites that constitute the post-rescope panel. A panel change that removes 5.2 % of the outage cannot manufacture the observed rise from 71.4 % on 12 October to 82.9 % on 14 October, still less the rise from 7.0 % across the event. We regard the confound as bounded and small, and note that the trend within the pre-rescope window alone (27 16

September to 12 October) is ρ = 0.85, n = 16. 5.4. E3: the 15 October event North Carolina’s outage count reached its post-storm minimum of 54 sites on 13 October, sixteen days after landfall. It then rose on two successive days (to 80 on 14 October and 123 on 15 October) before falling back to 58 on the 16th. Both increases are entirely transport: the transport column goes 33 → 68 → 115 while power falls 14 → 14 → 12 and damage stays at zero. The FCC’s own 15 October report notes the increase. We analyse the 14–15 October step in detail, for two reasons. It is the larger of the two, and it is the one for which an independent instrument shows a matching excursion (§5.5). The 13–14 October step has the same signature and we have no reason to think it a different event; treating them separately is conservative, because doing so attributes fewer sites to the mechanism we identify, not more. 5.4.1. Reconciling the two numbers Two figures circulate for this event and they measure different things. Total sites out rose by 43. Sites out due to transport rose by 47. The difference is not an inconsistency: 14 Oct

15 Oct

∆

Transport Power Damage Unattributed

68 14 0 −2

115 12 0 −4

+47 −2 0 −2

Sites out

80

123

+43

The four components sum to the change in the total exactly. (The unattributed residual is negative on these two days because the cause columns slightly exceed the printed total, a reminder that they are not a strict partition; see §6.) Throughout we quote the transport figure, +47, because it is the quantity the analysis is about; the county table below sums to +43 because it counts sites out, for which no county-level cause breakdown is published.

17

Table 6: Elimination of candidate explanations for the 15 October increase. Each rests on a primary source; the entire residual is transport.

Candidate Evidence

Verdict

Rescoping 901 sites served on 13– 16 Oct; identical county list Weather 31 GHCN-Daily stations across all nine counties: max 0.19 in on 14 Oct, 0.06 in on 15 Oct (NOAA National Centers for Environmental Information, 2024); 3rd-driest October in NC since 1895 (Davis, 2024) Power power-caused outages fell 14→12; sites on backup power fell 12→10 Damage zero on both days cable/wireline subRegionscribers out fell wide 43,696→40,963; a event common regional cause would have raised both series (§5.5)

excluded

Transport rose 68→115 (+47)

residual

excluded

excluded

excluded excluded

18

5.4.2. Eliminating alternatives Table 6 tests each candidate explanation against an independent source. The weather exclusion is the one that must carry weight, so we strengthened it. Rather than rely on a summary account, we retrieved daily precipitation for every GHCN-Daily station reporting in the nine counties (NOAA National Centers for Environmental Information, 2024; Menne et al., 2012): 31 stations, all nine counties represented. Every station recorded 0.00 in on 13 October. On 14 October the maximum anywhere was 0.19 in (Hot Springs, Madison County); on 15 October, 0.06 in (Burnsville, Yancey County), with every other station at zero or trace. Asheville Regional Airport recorded 0.03 in for the entire month. The row we previously labelled “broadband” is relabelled here. Cable subscriber outages were never a candidate explanation for cell-site outages; what that series actually rules out is a region-wide disruption (a storm, a regional power event, a metro-scale facility failure) which would have driven both series in the same direction. It fell while cell sites rose. 5.4.3. Geographic signature We state the limit of this evidence before presenting it. DIRS publishes cause only at state level, so nothing that follows is an observation of a countylevel cause. What the county tables give is sites out, by county, on two consecutive days; the attribution of the state-level increase to transport comes from §5.4.1, and the step from a spatial pattern to a single shared facility is an inference from that pattern, not a measurement of it. We cannot name the facility, the provider, or the sites, and no analysis in this paper can. §5.5.1 states what does and does not follow; readers who want the bound before the argument should read it now. Table 7 and Figure 4 give the change per county. Among the affected counties, loss is proportional to site count. The raw rates span a factor of five, from 2.8 % in Haywood to 13.6 % in Yancey, and it would be wrong to call them uniform. Yancey’s figure rests on three sites out of 22 and Madison’s on three out of 43; at those counts the Poisson interval alone spans most of the observed range. Four counties (Buncombe, Henderson, Rutherford and Madison) do sit close together, between 5.1 % and 7.0 %, but we report that as a description and attach no test to it: the four were picked out by their outcome, so testing them for homogeneity would be circular.

19

Mitchell 0

12

sites lost, % of county's sites

Yancey +3 (13.6%)

10

Madison +3 (7.0%)

8

McDowell

Buncombe Haywood +21 0 (6.0%) +3 (2.8%) Rutherford Henderson +5 (5.3%) +8 (5.1%) in panel, no change left panel 14 Oct (no data)

Polk 0

6 4 2 0

Figure 4: Change in cell sites out of service, 14–15 October 2024, across the nine counties in the reporting panel on both days. Six contiguous counties lose sites; three do not. The three unaffected counties are not outside the affected block (McDowell borders Buncombe, Rutherford and Yancey, all affected), which is why we read the pattern as provider- or route-specific rather than geographic. Counties in grey left the reporting panel on 14 October and carry no observation for these days.

20

Table 7: County-level change, 14–15 October 2024, with exact Poisson 95 % intervals on the loss rate and expected losses under a single per-site rate. “Nbrs” counts neighbours within the nine-county panel, not total adjacency. The three counties with no increase are interior neighbours of affected ones, not peripheral to them.

∆ % of county 95 % CI Exp. Nbrs

County

Sites

Buncombe Henderson Rutherford Madison Haywood Yancey

348 +21 158 +8 94 +5 43 +3 106 +3 22 +3

McDowell Polk Mitchell

83 27 20

Total

901 +43

0 0 0

6.0 % 3.7–9.2 16.6 6 5.1 % 2.2–10.0 7.5 4 5.3 % 1.7–12.4 4.5 4 7.0 % 1.4–20.4 2.1 3 2.8 % 0.6–8.3 5.1 3 13.6 % 2.8–39.9 1.1 4 0.0 % 0.0 % 0.0 %

0–4.4 0–13.7 0–18.4

4.0 4 1.3 2 1.0 2 43.0

What the data support is the weaker statement that among the six affected counties the loss is not inconsistent with being proportional to site count. A chi-square test of homogeneity against site-count exposure does not reject a single common per-site rate (χ2 = 4.36, df = 5, p = 0.50), but at 43 events across six counties that test would also fail to reject a genuinely varying rate, so it discriminates between the two hypotheses hardly at all. We report it as a consistency check (the observed spread is what a single rate would produce, so the spread itself is not evidence against a shared cause), and we place no positive weight on it. But the panel as a whole is not proportional, and that is the point. Extending the same test to all nine counties, including the three zeros, gives χ2 = 12.35, df = 8, p = 0.14 (Monte-Carlo p = 0.13 over 106 multinomial draws, since several expected counts are below five). That omnibus test does not reject either, and for the same reason carries no evidential weight: it spends eight degrees of freedom on variation among the affected counties, which is noise, and dilutes the one contrast that matters. The statistics that do bear on the question are the two below, which test a prediction rather than summarise a spread. The contrast is this. McDowell, Polk and Mitchell hold 130 of the panel’s 901 sites (14.4 % of the exposure) and took none of the 43 losses, against 6.2 21

expected. Because those three counties are identified by their outcome, we do not attach a p-value to that comparison directly. Two versions of it that are not conditioned on which counties came up empty. Neither is decisive on its own (43 events across nine counties is not much to work with), but both test a geographic prediction rather than the pattern that suggested it, and both point the same way: • Under a common per-site rate, the counties that happen to record zero losses account for 14.4 % or more of the panel’s sites in 0.7 % of 106 simulated draws, and three or more of the nine counties come up empty in 6.3 %. • McDowell alone. McDowell is singled out by geography, not by its outcome: it is the only county in the panel interior to the affected block, bordering Buncombe, Rutherford and Yancey. It is also the panel’s fourth largest, with 83 sites and 4.0 expected losses. It recorded zero, which a common rate produces with probability 0.016. The unaffected counties are interior, not peripheral. This is the part of the pattern that constrains the explanation most. Computing adjacency directly from Census county boundaries: McDowell borders Buncombe, Rutherford and Yancey, all three of which lost sites; Polk borders Henderson and Rutherford; Mitchell borders Yancey. None of the three sits outside the affected block. A purely geographic cause (a storm cell, a regional power event, a flood) would not skip a county wedged between three affected neighbours while taking the neighbours on every side. A facility that is specific to a provider, or to a route, would: it follows the network graph, not the map. Figure 4 plots the panel so that the adjacency claim can be checked directly. Loss proportional to site count across a contiguous block, with interior neighbours untouched, is the signature of a shared upstream path failing rather than many independent local failures. If a single provider holds roughly a quarter to a third of the region’s sites and lost about a fifth of its own, the arithmetic yields the observed 5 % to 7 % of the total. Recovery to 58 sites out by the 16 October report (one reporting day, exactly 24 hours between snapshots) is consistent with a fibre splice rather than reconstruction. 5.5. E4: external corroboration DIRS is filed by the providers whose networks failed, so the 15 October event rests, so far, on self-report. We therefore asked whether an independent 22

instrument saw it. IODA’s ping-slash24 signal counts /24 address blocks responding to active probes (Georgia Institute of Technology Internet Intelligence Lab, 2026, 2024); the adaptive /24 probing method the signal implements is that of Quan et al. (2013). It is a different operator, a different instrument and a different physical measurement from a provider’s outage filing. Figure 5 plots it for North Carolina and, as a control, Tennessee (in the same DIRS activation, probed by the same platform, and reporting no 15 October increase). Why a cellular backhaul cut is visible to a ping-based signal. Cell sites do not themselves host responsive /24s, so the connection needs stating. In this terrain the backhaul carrying a cell site to the network core and the transport carrying fixed-line subscribers are, very largely, the same fibre in the same conduit along the same valley route. A cut that strands cell sites therefore also strands whatever fixed-line address space rides behind the same facility, and it is the latter that IODA can see. This is not incidental to the corroboration; it is what makes it meaningful. It is also why the corroboration is only ever partial: the two populations overlap on the transport path, not one for one. North Carolina’s responsive-block count is flat at 26,337 through 14 October (the baseline against which we measure), peaks at 26,400 at 02:00 UTC on 15 October, and falls to 26,067 by 05:00 UTC. Relative to that baseline the drawdown is 1.02 %, or 270 blocks; measured from the transient 02:00 peak it would be 333. It remains depressed for twelve hours and recovers from 15:00 UTC. Tennessee over the same window spans 0.29 % end to end and never falls more than 0.18 % below its baseline. The dip is the only one in the four-day window and is not diurnal: the equivalent overnight hours on 14 and 16 October are flat. How unusual is a 1 % dip?. In the raw series, not very; the control is what separates a state-specific excursion from a platform-wide one. Over the whole of October (124 six-hourly samples, Figure 5c) the deepest raw North Carolina drawdown is 2.5 % on 11 October, and Tennessee falls 2.6 % at the same instant. Excursions that move both states by the same relative amount are the measurement platform, not the states, and differencing against the control removes them. In the differenced series the 15 October sample is 4.0 standard deviations below the October median, and only two of the 124 samples in the month reach 0.8 %. 23

responsive /24s (% of 14 Oct mean)

0.5 (a) 0.0 0.5 1.0

DIRS cell sites out

1.5 150 (b) 100

123 80

58

50 14 Oct 00:00

NC TN (% of median)

1.02% ( 270 blocks)

North Carolina Tennessee (control)

0.0

14 Oct 12:00

15 Oct 00:00

15 Oct 12:00

2024, UTC

16 Oct 00:00

16 Oct 12:00

(c)

0.5 15 Oct

1.0 10 Oct

1.5 01

08

15

October 2024, UTC

22

29

Figure 5: (a) IODA active-probe measurement across the event, as a percentage of each state’s own 14 October mean, with Tennessee as a control. (b) The DIRS series, sampled at the instants the reports describe. Dotted lines mark the 09:00 EDT DIRS snapshots; shading marks the interval in which North Carolina sits more than 0.5 % below baseline; the 15 October report was filed while the dip was in progress. (c) The whole of October, North Carolina minus the Tennessee control, against a 7-day rolling median. Only two excursions in the month reach 0.8 %.

24

The other is 10 October. North Carolina shows a state-specific excursion of 1.5 % that day with no corresponding movement in DIRS, whose cellsite count fell monotonically from 9 to 11 October. The two instruments are therefore not in one-to-one correspondence, and we do not claim they are. What we claim is narrower: the single day on which DIRS records an otherwise unexplained increase in cell sites out of service is also one of only two days in the month on which IODA records a North-Carolina-specific connectivity excursion, and the two coincide to within hours. The routing plane says the failure was not upstream. IODA’s BGP signal for North Carolina moves by 3 prefixes out of 79,007 (0.004 %) across the entire twelve-hour window, having last stepped a day earlier. A withdrawn transit path or a failed core facility would appear here; an access-layer or backhaul failure, in which the prefixes remain advertised while the equipment behind them becomes unreachable, would not. That is the shape of what we observe. The timing is the part worth stating carefully. DIRS reports describe 09:00 EDT (13:00 UTC) on their date. At that instant on 14, 15 and 16 October, IODA’s count reads 26,344, 26,131 and 26,314: down 213 blocks at the 15 October snapshot, recovered by the 16th. The two instruments, measuring different things by different means, move together and recover together. IODA also shows recovery beginning about two hours after the 15 October snapshot, which is consistent with the single-reporting-day restoration DIRS records. Reconciling this with the “region-wide event” exclusion. Table 6 excludes a region-wide cause partly because cable and wireline subscribers out fell on 15 October, from 43,696 to 40,963. Here we read a fall in a fixed-line connectivity measure as evidence that something happened. Both readings are correct, but the apparent tension needs resolving explicitly. The two are compatible for three reasons, in increasing order of importance. The series measure different populations: DIRS wireline counts subscribers of the cable and wireline providers that chose to file, while IODA counts responsive address space across every network in the state, including transport providers and networks that file nothing. They have different resolutions: the DIRS figure is one number per day, moving against a restoration trend that was returning thousands of subscribers daily, so a twelve-hour excursion of a few hundred blocks is not separable from it. And the two rows are not making the same claim. The Table 6 row does not assert that no 25

fixed-line disruption occurred on 15 October; it rules out a region-wide cause, which would have driven the subscriber series sharply upward by thousands rather than leaving it falling. A localised transport facility failing, affecting one provider’s backhaul and whatever address space shares its path, is exactly the kind of event that shows up as a 1 % probing excursion while a statewide net subscriber count continues to recover. Both observations are consistent with the same explanation. Limits of the corroboration. IODA publishes no county-level series. Entity codes for all nine counties resolve, but every datasource returns an empty series, and IODA’s own documentation describes country, region and network granularity only. This corroboration is therefore statewide: it confirms that something removed connectivity in North Carolina, in the right direction, on the right day, for about the right duration, and that Tennessee did not experience it. It cannot distinguish the six affected counties from the three unaffected ones. That discrimination remains available only from DIRS, or from NORS. 5.5.1. What the 15 October analysis establishes Eliminating rescoping, weather, power, damage and a region-wide event is well supported: each rests on a primary source, and the timing is now independently corroborated. The inference to a single shared backhaul facility remains indirect: we cannot name it, because DIRS does not identify providers or sites, and IODA cannot localise below state level. What the event does establish does not depend on that mechanism. On 15 October there was no rainfall, no rise in power-caused outages and no reported damage; the 47 sites were, by the filings’ own accounting, intact and powered. A change in connectivity alone removed service across six counties within a single reporting interval, 17 days after landfall: a failure mode distinct from physical destruction, and observable in a public filing. 5.6. Implications Taken together the four analyses say something narrower than “backhaul is the problem”, and more useful. Cell sites overwhelmingly survived Helene physically (§5.1). Whatever mix of power and transport took them out of service, that mix shifted toward transport in every state we can measure, and kept shifting for as long as each record runs (§5.2). And a transport failure alone can remove service from a whole region in the third week of a 26

recovery, with no weather, no power loss and no damage (§5.4, §5.5). Three consequences follow. Resilience investment is mistargeted where terrain constrains routes. Backup power is the standard hardening measure for cell sites and it addresses 47.3 % of this event in North Carolina and 63.2 % of it pooled across all six states. In the mountains, transport diversity addresses the larger and more persistent share, and it is not systematically reported, mandated or measured. The remedy is not hypothetical: a regional broadcaster stayed on the air through the event by substituting satellite connectivity for its terrestrial backhaul (Connelly et al., 2025): the same substitution, one layer over, made by an operator who could choose it. The FCC’s 2026 harmonisation of transport-facility reporting is a necessary precondition for even observing it (Federal Communications Commission, 2026c). Recovery work may cause re-failure. The 15 October event occurred amid intense reconstruction. Debris removal and rebuilding involve heavy earthmoving, which severs fibre. If this generalises, the period of maximum reconstruction activity is also a period of elevated topological failure, a feedback loop that current planning does not represent. Application architecture assumes the wrong failure. Every mainstream communication application requires a path to a distant server, so a region whose local infrastructure survives but whose backhaul is severed loses all of them at once. 6. Limitations and Future Work 6.1. Limitations of the source The cause columns are not a strict partition. The three counts do not sum to the printed total: they fall short by 6.9 % of North Carolina site-days overall and 7.6 % pooled, and on a few days slightly exceed it. We report both denominators and the residual wherever a headline figure appears (§3.4), and test both conclusions against the residual adversarially in §5.1. The outcome is asymmetric and worth restating here: the damage share is a floor (ceiling 8.6 % pooled if the entire residual were damage), while the North Carolina transport-over-power ordering would reverse if more than 66 % of that state’s residual were really power. We find no evidence of such a skew, but we cannot exclude it, and a reader who assumes it should read the North Carolina ordering as unresolved and rely on Tennessee, whose residual is 0.8 %. 27

“Damage” means damage to the cell site. It does not mean the network suffered no physical destruction; Helene destroyed more than 1700 miles of fibre. Much of what DIRS records as a transport outage is downstream of physical damage to a transport facility. Our contribution is to locate that damage on the network graph, not to deny it. The reporting area is rescoped mid-event. Denominators change discontinuously: North Carolina falls from 1452 to 901 sites between the 12 and 13 October reports. A naive reading of the raw subscriber series would show an overnight fall from 325,348 to 84,085 on 8 October; this is Georgia and South Carolina leaving the reporting area, not recovery. We use within-state shares throughout and bound the residual confound in §5.3. Reporting is voluntary and unattributed. DIRS does not identify which provider filed what, nor which specific site failed. We characterise causes in aggregate but cannot attribute an outage to an operator or a facility. Cause is self-reported. “Transport” is the filing provider’s own classification; we have no means to audit it, and a site simultaneously lacking power and backhaul could plausibly be filed under either. The 15 October event is the one place we obtain external corroboration (§5.5), and only at state granularity. 6.2. Limitations of the analysis Cell sites are not people. A site-weighted share is not a populationweighted one: a rural site serves far fewer subscribers than an urban one. Our findings characterise infrastructure, not experienced outage. Populationweighting county outage rates for western North Carolina on 29 September yields roughly 350,000 residents in counties experiencing transport-caused outage, but we regard this as order-of-magnitude only. One event, and a partly terrain-dependent result. These are the causes of one hurricane. Within it, transport dominance over the full record holds in two mountainous states and not in four others, while the temporal shift toward transport holds in all four states long enough to test (§5.2). We scope each claim accordingly, but cannot test the terrain mechanism itself with six states and no route-diversity data. The trend is descriptive. With one event there is no replication. We report autocorrelation-corrected significance (§5.2) because reviewers of measurement work reasonably ask for it, not because a p-value carries inferential weight here. 28

This is an observational study, not a natural experiment. The 15 October event involves no as-if-random assignment. It is an isolating event: an anomaly whose alternative explanations can be eliminated against primary sources. We think that is respectable, and we do not claim more for it. A transport-caused outage is not automatically recoverable. A site without backhaul cannot serve traffic either, since the mobile core is upstream. Our decomposition measures how much of the failure was topological rather than physical, an upper bound on what architectural mitigation could address, not a quantity of service that was available and unused. 6.3. Future work Other DIRS activations. The Commission activates DIRS for every major disaster, and each activation produces a comparable record. Applying this method across events is the direct test of the terrain hypothesis: it would separate whether transport dominance is a property of mountainous terrain, of Helene specifically, or of modern cellular infrastructure generally. Our six-state comparison suggests the first, and cannot establish it. Confirming the 15 October mechanism. The corresponding NORS filings would name the provider, the facility and the root cause. NORS is confidential under 47 C.F.R. §4.2, but the Commission processes requests for confidential records under §0.461, and an aggregated disclosure across four or more providers would suffice to confirm or refute the shared-path hypothesis without exposing competitively sensitive detail. Sub-state external measurement. IODA corroborates the 15 October event at state level (§5.5) but publishes nothing finer. Probing data with county or ASN resolution (or IODA’s per-network signals matched to the providers serving western North Carolina) would test the county-level pattern directly, and more generally allow the cause-labelled DIRS record to be crossvalidated against externally observed outages. Architectural implications. A failure mode in which local infrastructure survives but the path out does not is, in principle, addressable by systems that do not require a path out. Evaluating partition-tolerant designs against this trace, rather than against synthetic outage models, is the direction we intend to pursue.

29

7. Conclusions The largest cell-site outage in the FCC’s public disaster record was not, in the main, caused by destroyed cell sites. Across six states, damage at the site accounts for 1.1 % of attributed cell-site-days and never exceeds 3.8 % anywhere. The sites were standing. What took them out divides by terrain. Pooled across the whole event, commercial power remains the larger cause. But in the mountains of western North Carolina, where Helene’s outage was largest and longest, severed backhaul accounts for 52.2 % of attributed cell-site-days against 47.3 % for power, and its share grows across the event from 7.0 % to 85.0 % until it accounts for essentially all remaining outages, a pattern Tennessee independently reproduces. On 15 October, in the third week of recovery, 47 sites across six contiguous counties lost their backhaul with no weather, no power loss and no damage of any kind, and an independent probing platform recorded the same twelve hours. The towers were, for the most part, standing and powered. What they had lost was the path to the network core. CRediT authorship contribution statement Oluseyi Olukola: Conceptualization, Methodology, Software, Formal analysis, Data curation, Investigation, Validation, Visualization, Writing (original draft), Writing (review and editing). Oare Danielle Addeh: Writing (original draft), Writing (review and editing). Esther Abiodun Konan: Writing (original draft), Writing (review and editing). Nick Rahimi: Conceptualization, Methodology, Supervision, Validation, Writing (review and editing), Project administration. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Data availability The reconstructed dataset (80 state-days and 580 county-days with percause breakdown, the extraction and reconciliation code, the statistical checks 30

reported in §5, and the figure sources) accompanies this submission as supplementary material, and will be deposited in a public repository under a CC-BY-4.0 licence, with a citable DOI, on acceptance. All source documents are public FCC filings, individually cited by document number; the NCEI and IODA queries are given as complete request URLs so that both external checks can be re-run exactly. References Bischof, Z.S., Pitcher, K., Carisimo, E., Meng, A., Nunes, R.B., Padmanabhan, R., Roberts, M.E., Snoeren, A.C., Dainotti, A., 2023. Destination unreachable: Characterizing internet outages and shutdowns, in: Proceedings of the ACM SIGCOMM 2023 Conference, Association for Computing Machinery, New York, NY, USA. pp. 608–621. doi:10.1145/3603269.36 04883. Booker, G., Torres, J., Guikema, S., Sprintson, A., Brumbelow, K., 2010. Estimating cellular network performance during hurricanes. Reliability Engineering & System Safety 95, 337–344. doi:10.1016/j.ress.2009.11 .003. Cho, K., Pelsser, C., Bush, R., Won, Y., 2011. The Japan earthquake: The impact on traffic and routing observed by a local ISP, in: Proceedings of the Special Workshop on Internet and Disasters (SWID ’11), co-located with ACM CoNEXT 2011, Association for Computing Machinery, Tokyo, Japan. pp. 1–8. doi:10.1145/2079360.2079362. claffy, k., Clark, D., 2022. Challenges in measuring the Internet for the public interest. Journal of Information Policy 12, 195–233. doi:10.5325/jinfop oli.12.2022.0003. Connelly, D., Farmer, B., Spasovska, K., 2025. The role of radio and ham radio when everything else fails during crisis: Lessons learned from Hurricane Helene, in: International Crisis and Risk Communication Association Reports (2025 Annual Proceedings), pp. 130–133. doi:10.69931/001c.14 2861. Dainotti, A., Squarcella, C., Aben, E., Claffy, K.C., Chiesa, M., Russo, M., Pescapé, A., 2011. Analysis of country-wide internet outages caused by censorship, in: Proceedings of the 2011 ACM SIGCOMM Conference on 31

Internet Measurement Conference (IMC ’11), Association for Computing Machinery, Berlin, Germany. pp. 1–18. doi:10.1145/2068816.2068818. Dainotti, A., Squarcella, C., Aben, E., Claffy, K.C., Chiesa, M., Russo, M., Pescapé, A., 2014. Analysis of country-wide internet outages caused by censorship. IEEE/ACM Transactions on Networking 22, 1964–1977. doi:10 .1109/TNET.2013.2291244. Davis, C., 2024. October dries out in a monthly rainfall reversal. North Carolina State Climate Office. URL: https://climate.ncsu.edu/b log/2024/11/october-dries-out-in-a-monthly-rainfall-rever sal. published 4 November 2024. Statewide October 2024 precipitation averaged 0.54 in, the third-driest October since 1895. Du, A., 2024. Data-driven telecommunication outage prediction during hurricane events. ASCE-ASME Journal of Risk and Uncertainty in Engineering Systems, Part A: Civil Engineering 10, 04024046. doi:10.1061/AJRUA6.R UENG-1285. uses county-level telecommunication outage data from ten recent US hurricanes. Durairajan, R., Barford, C., Barford, P., 2018. Lights out: Climate change risk to internet infrastructure, in: Proceedings of the Applied Networking Research Workshop (ANRW ’18), Association for Computing Machinery, Montreal, QC, Canada. pp. 9–15. doi:10.1145/3232755.3232775. Federal Communications Commission, 2013. Improving the resiliency of mobile wireless communications networks. Report and Order, PS Docket Nos. 13-239 and 11-60, FCC 13-125. URL: https://docs.fcc.gov/public/ attachments/FCC-13-125A1.pdf. released 27 September 2013. Federal Communications Commission, 2017a. Communications status report for areas impacted by Hurricane Irma. Public Safety and Homeland Security Bureau, report as of 11 September 2017, document DOC-346655A1. URL: https://docs.fcc.gov/public/attachments/DOC-346655A1.pd f. Federal Communications Commission, 2017b. Communications status report for areas impacted by Hurricane Maria. Public Safety and Homeland Security Bureau, report as of 23 September 2017, document DOC-346860A1.

32

URL: https://docs.fcc.gov/public/attachments/DOC-346860A1.pd f. Federal Communications Commission, 2021. Communications status report for areas impacted by Hurricane Ida. Public Safety and Homeland Security Bureau, report as of 30 August 2021, document DOC-375318A1. URL: https://docs.fcc.gov/public/attachments/DOC-375318A1.pdf. Federal Communications Commission, 2024. Communications status reports for areas impacted by hurricane helene. Public Safety and Homeland Security Bureau, daily reports 26 September – 19 October 2024. URL: https://www.fcc.gov/helene. documents DOC-405827A1 through DOC-406771A1. Retrieved from https://docs.fcc.gov/public/at tachments/. The 24 daily filings analysed here span 26 September – 19 October 2024; the 26 September filing predates state-level cause reporting. Federal Communications Commission, 2026a. Disaster information reporting system (DIRS). URL: https://www.fcc.gov/general/disaster-i nformation-reporting-system-dirs-0. voluntary reporting system activated for major disasters; filings are aggregated and published daily. Federal Communications Commission, 2026b. Network outage reporting system (NORS). URL: https://www.fcc.gov/network-outage-reporti ng-system-nors. mandatory outage reporting under 47 C.F.R. Part 4; filings are presumed confidential under 47 C.F.R. §4.2. Federal Communications Commission, 2026c. Resilient networks; amendments to part 4 of the commission’s rules concerning disruptions to communications; new part 4 of the commission’s rules concerning disruptions to communications. Third Report and Order, PS Docket Nos. 21-346 and 15-80, ET Docket No. 04-35, FCC 26-34. URL: https: //docs.fcc.gov/public/attachments/FCC-26-34A1.pdf. adopted 20 May 2026; released 21 May 2026. Federal Communications Commission, 2026d. Resilient networks; concerning disruptions to communications. Final rule, 91 Fed. Reg. 39516 (June 30, 2026) (to be codified at 47 C.F.R. pt. 4). URL: https://www.federalr egister.gov/documents/2026/06/30/2026-13155. fR Doc. 2026-13155; effective 30 June 2026 except 47 C.F.R. § 4.18. 33

of Federal Regulations, C., 2026. Disruptions to communications. 47 C.F.R. pt. 4 (2026). URL: https://www.ecfr.gov/current/title-47/chapte r-I/subchapter-A/part-4. Feeny, N., White, A., Guikema, S., 2026. A quantitative spatial approach to estimate cellular network coverage during natural hazards using publicly available data. Journal of Infrastructure Preservation and Resilience 7, 14. doi:10.1186/s43065-026-00176-0. Gallagher, J.C., 2024. Restoration of Cell Phone Services: Hurricane Helene. CRS In Focus IF12779. Congressional Research Service. URL: https: //www.congress.gov/crs-product/IF12779. Gamboa, S., Chelminski, P.R., Shenvi, C., 2026. Lessons learned from Helene: The role of a rural community hospital in disaster response after a major hurricane. Annals of Emergency Medicine doi:10.1016/j.annemergme d.2026.01.005. disaster Medicine/Concepts section; published online 10 February 2026, volume and pages not yet assigned. Georgia Institute of Technology Internet Intelligence Lab, 2024. IODA v2 API: Internet outage detection and analysis. https://api.ioda.ineti ntel.cc.gatech.edu/v2/. Signal ping-slash24, region entities 4444 (North Carolina) and 4446 (Tennessee). Georgia Institute of Technology Internet Intelligence Lab, 2026. IODA: Internet outage detection and analysis. Online platform, https://ioda.i netintel.cc.gatech.edu/. Griffith, D., Rouil, R., Izquierdo, A., Golmie, N., 2015. Measuring the resiliency of cellular base station deployments, in: 2015 IEEE Wireless Communications and Networking Conference (WCNC), IEEE, New Orleans, LA, USA. pp. 1625–1630. doi:10.1109/WCNC.2015.7127711. Hauser, G., 2025. Bridging communication gaps: Lessons from Hurricane Helene. Domestic Preparedness Journal. URL: https://domesticprepar edness.com/articles/bridging-communication-gaps-lessons-fro m-hurricane-helene. published 30 April 2025. The author is the North Carolina Division of Emergency Management statewide interoperability coordinator and ESF-2 lead for the Helene response.

34

Heidemann, J., Quan, L., Pradkin, Y., 2012. A Preliminary Analysis of Network Outages During Hurricane Sandy. Technical Report ISI-TR-2008685b. USC/Information Sciences Institute. URL: https://ant.isi.edu/ ~johnh/PAPERS/Heidemann12d.html. correction February 2013. Kuhn, D.R., 1997. Sources of failure in the public switched telephone network. Computer 30, 31–36. doi:10.1109/2.585151. Maitland, C., Peha, J.M., 2018. Wireless network recovery following natural disaster: Puerto Rico after Hurricane Maria, in: TPRC 46: The 46th Research Conference on Communication, Information and Internet Policy. URL: https://ssrn.com/abstract=3142393. sSRN Abstract ID 3142393, posted 19 March 2018. Malandrino, F., Chiasserini, C.F., 2017. Quantifying and minimizing the impact of disasters on wireless communications, in: Proceedings of the First CoNEXT Workshop on ICT Tools for Emergency Networks and DisastEr Relief (WICTEND ’17), Association for Computing Machinery, Incheon, Republic of Korea. pp. 26–30. doi:10.1145/3152896.3152902. Menne, M.J., Durre, I., Vose, R.S., Gleason, B.E., Houston, T.G., 2012. An overview of the global historical climatology network-daily database. Journal of Atmospheric and Oceanic Technology 29, 897–910. doi:10.117 5/JTECH-D-11-00103.1. NOAA National Centers for Environmental Information, 2024. Global historical climatology network daily (GHCN-daily), accessed via the NCEI access data service. https://www.ncei.noaa.gov/access/services/data/v1. Dataset daily-summaries, element PRCP; retrieved for 31 stations in nine western North Carolina counties, 13–15 October 2024. Padmanabhan, R., Schulman, A., Levin, D., Spring, N., 2019. Residential links under the weather, in: Proceedings of the ACM Special Interest Group on Data Communication (SIGCOMM ’19), Association for Computing Machinery, Beijing, China. pp. 145–158. doi:10.1145/3341302.33 42084. Quan, L., Heidemann, J., Pradkin, Y., 2013. Trinocular: Understanding internet reliability through adaptive probing. ACM SIGCOMM Computer

35

Communication Review 43, 255–266. doi:10.1145/2534169.2486017. proceedings of ACM SIGCOMM 2013. Reed, D.A., Wang, S., 2018. Numerical modeling of power delivery and telecommunications infrastructure for hurricanes Harvey and Irma, in: Forensic Engineering 2018: Forging Forensic Frontiers, American Society of Civil Engineers, Austin, TX, USA. pp. 1008–1016. doi:10.1061/9780 784482018.097. Service, C.R., 2025. Cellular Network Outage Reporting and Restoration During Disasters. CRS Report R48776. Congressional Research Service. URL: https://www.congress.gov/crs-product/R48776. describes NORS, DIRS and the Mandatory Disaster Response Initiative and their data limitations. Yang, S., He, Y., Ge, Z., Wang, D., Xu, J., 2017. Predictive impact analysis for designing a resilient cellular backhaul network. Proceedings of the ACM on Measurement and Analysis of Computing Systems 1, 1–33. doi:10.114 5/3154488. Appendix A. Full daily record, North Carolina Table A.8 gives the complete per-day North Carolina record underlying Figures 2 and 3. Denominator changes are marked and no values are interpolated. The panel narrows between 12 and 13 October, which is where the served count falls from 1452 to 901. Appendix B. Trend statistics Table B.9 gives the trend statistics summarised in §5.2. They are reported here rather than in the body because the finding is the size of the shift, not the rejection of a no-trend null; these numbers establish only that serial correlation does not account for the shift. Two details are worth stating so they are not mistaken for errors. First, the Durbin–Watson statistic (1.16) is lower than the 2(1 − ρ̂) approximation would suggest from either lag-1 estimate (1.57 and 1.69). The gap is an 2 end effect, P 2 not a contradiction: the exact identity is DW = 2 − 2ρ̂YW − (e1 + 2 en )/ et , and here the two endpoint residuals carry 53 % of the residual sum of squares, almost all of it the 27 September onset day, which sits 26 points 36

Table A.8: North Carolina daily record. “Share” is the transport share, computed against the sum of the three cause columns (§3.4). Note the served-count changes on 29 September and 13 October, and the two-day rise on 14–15 October analysed in §5.4.

Date

Served

Out

Dmg

Trans

Power

Share

27 Sep 28 Sep 29 Sep 30 Sep 1 Oct 2 Oct 3 Oct 4 Oct 5 Oct 6 Oct 7 Oct 8 Oct 9 Oct 10 Oct 11 Oct 12 Oct 13 Oct 14 Oct 15 Oct 16 Oct 17 Oct 18 Oct 19 Oct

1452 1452 1556 1452 1452 1448 1448 1448 1448 1448 1448 1448 1448 1448 1448 1452 901 901 901 901 901 901 901

58 1079 1034 784 707 554 410 321 277 244 209 213 184 154 115 96 54 80 123 58 39 32 20

1 7 5 3 3 3 2 3 3 2 1 1 0 0 0 0 0 0 0 0 1 1 0

4 428 509 385 303 255 178 134 119 116 99 108 116 102 77 60 33 68 115 39 32 28 17

52 622 493 370 338 269 193 149 121 88 69 65 39 35 26 24 14 14 12 8 6 2 3

7.0 % 40.5 % 50.5 % 50.8 % 47.0 % 48.4 % 47.7 % 46.9 % 49.0 % 56.3 % 58.6 % 62.1 % 74.8 % 74.5 % 74.8 % 71.4 % 70.2 % 82.9 % 90.6 % 83.0 % 82.1 % 90.3 % 85.0 %

below the fitted line. At n = 23 that correction term is not negligible, as it is asymptotically. Second, we use the larger of the two lag-1 estimates for the Bartlett correction, which is the conservative choice: the Yule–Walker estimate gives an effective n of 16.9 and a smaller p of 2.2 × 10−7 . The moving-block bootstrap uses 104 resamples of contiguous 4-day blocks, the timescale of the recovery itself.

37

Table B.9: Trend in the North Carolina daily transport share. The naive p-value is anticonservative because a monotone recovery makes consecutive days correlated by construction; the two corrections address that. All three are reported so the correction can be judged.

Statistic Observations n Spearman ρ Pearson r OLS slope

Value 23 0.925 0.919 +2.72 pp/day

Residual lag-1, corr(et , et−1 ) Residual lag-1, Yule–Walker Durbin–Watson Effective n (Bartlett, on 0.215) p, naive p, autocorrelation-adjusted p, moving-block bootstrap

0.215 0.154 1.16 14.9 2.8 × 10−10 1.5 × 10−6 < 10−4

Appendix C. External check queries Both external checks are reproducible from public APIs. Precipitation (§5.4) uses the NCEI Access Data Service, dataset daily-summaries, element PRCP, for 31 stations across the nine counties. Active-probe corroboration (§5.5) uses the IODA v2 API, signal ping-slash24, region entity 4444 (North Carolina) and 4446 (Tennessee), over from=1728864000 to until=1729123200. Full request URLs and the retrieved series are included in the artifact.

38